WO1993009494A1 - Fault-tolerant computer processing using a shadow virtual processor - Google Patents

Fault-tolerant computer processing using a shadow virtual processor Download PDF

Info

Publication number
WO1993009494A1
WO1993009494A1 PCT/US1992/009238 US9209238W WO9309494A1 WO 1993009494 A1 WO1993009494 A1 WO 1993009494A1 US 9209238 W US9209238 W US 9209238W WO 9309494 A1 WO9309494 A1 WO 9309494A1
Authority
WO
WIPO (PCT)
Prior art keywords
backup
primary
computer
virtual machine
traps
Prior art date
Application number
PCT/US1992/009238
Other languages
French (fr)
Inventor
Fred B. Schneider
Butler Lampson
Edward Balkovich
David Thiel
Original Assignee
Digital Equipment Corporation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Digital Equipment Corporation filed Critical Digital Equipment Corporation
Publication of WO1993009494A1 publication Critical patent/WO1993009494A1/en

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/16—Error detection or correction of the data by redundancy in hardware
    • G06F11/1675—Temporal synchronisation or re-synchronisation of redundant processing components
    • G06F11/1691—Temporal synchronisation or re-synchronisation of redundant processing components using a quantum
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/16—Error detection or correction of the data by redundancy in hardware
    • G06F11/20—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements
    • G06F11/202—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements where processing functionality is redundant
    • G06F11/2023—Failover techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/16—Error detection or correction of the data by redundancy in hardware
    • G06F11/20—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements
    • G06F11/202—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements where processing functionality is redundant
    • G06F11/2023—Failover techniques
    • G06F11/2033—Failover techniques switching over of hardware resources
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/16—Error detection or correction of the data by redundancy in hardware
    • G06F11/20—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements
    • G06F11/202—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements where processing functionality is redundant
    • G06F11/2038—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements where processing functionality is redundant with a single idle spare processing component
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/16—Error detection or correction of the data by redundancy in hardware
    • G06F11/20—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements
    • G06F11/202—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements where processing functionality is redundant
    • G06F11/2046—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements where processing functionality is redundant where the redundant components share persistent storage
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/16—Error detection or correction of the data by redundancy in hardware
    • G06F11/20—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements
    • G06F11/2097—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements maintaining the standby controller/processing unit updated
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/16—Error detection or correction of the data by redundancy in hardware
    • G06F11/20—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements
    • G06F11/202—Error detection or correction of the data by redundancy in hardware using active fault-masking, e.g. by switching out faulty elements or by switching in spare elements where processing functionality is redundant
    • G06F11/2023—Failover techniques
    • G06F11/2025—Failover techniques using centralised failover control functionality
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2201/00—Indexing scheme relating to error detection, to error correction, and to monitoring
    • G06F2201/815—Virtual

Definitions

  • the present invention relates generally to fault tolerant computer systems and particularly to methods and systems for providing fault tolerance that is independent of the computer's operating system software.
  • the basic scheme used in most fault tolerant computer systems is to employ a primary and a backup computer. Users interact with the primary computer in order to obtain a service.
  • the primary computer performs the tasks requested by users and informs the backup of its actions so that the backup can take over providing the service if the primary computer fails.
  • hardware failures in the primary computer do not cause interruption of service to the users.
  • the backup processor In a properly implemented instance of this scheme, the backup processor must generate no interactions with its environment before the primary computer has failed. And, after the primary computer has failed, the backup processor must generate interactions with its environment in such a way that the environment is unaware of the primary computer's failure.
  • Fault tolerance in computers is usually implemented either (A) by constructing special purpose computer hardware, or (B) by modifying the computer's operating system.
  • the special purpose hardware approach requires hardware that is intimately related to the computer processor's design. As a result, such computers are usually unnecessarily costly for clients who do not require fault tolerance.
  • the major problem associated with using special purpose operating system code to implement fault tolerance is that the only operating system that can be used on that computer system, and still maintain fault tolerance, is the operating system containing the special purpose code. If a user who needs fault tolerance wants to use another operating system with the computer, extensive (and thus expensive) changes to this second operating system will be required.
  • the goals of the present invention are (1) to provide a fault tolerant computer system with little or no added cost for clients and processes that do not require fault tolerance, and (2) to provide a fault tolerance mechanism that works regardless of the operating system software used by the system's clients. While the "special-purpose hardware" and “modified operating system” approaches are both capable of meeting the basic requirements for a fault-tolerant computer system, the present invention overcomes cost problems associated with the "special-purpose hardware” approach and provides more flexibility in terms of operating system selection than the "modified operating system” approach.
  • the present invention is a fault-tolerant computer system having a primary computer and backup computer, both of which are capable of delaying delivery of interrupts to the computer's operating system due to aysnchronous activity by associated components (e.g., I/O devices) and the use of instruction pipelining.
  • Primary and backup virtual processors running on the computers are controlled by corresponding virtual machine monitors. Only programs that need to be fault tolerant are run on both the primary and backup virtual processors. Programs that do not need to be fault tolerant are run on other virtual processors on the primary computer.
  • the computer hardware directly executes only user-mode instructions of the virtual processors, but all kernel-mode instructions (e.g., I/O instructions) are trapped and handled by the virtual machine monitors.
  • Each computer also has a recovery register that can be loaded with a value and that generates a hardware interrupt after a corresponding number of instructions are executed.
  • the hardware interrupts generated by the recovery register mark the boundary between "epochs" in the instruction stream.
  • the backup virtual processor Prior to failure of the primary computer, the backup virtual processor generates no interactions with its environment. To accomplish this, the backup computer's virtual machine monitor converts all I/O instructions into no-ops (i.e., null operations). Furthermore, the primary computer sends copies of all I/O interrupts destined for delivery to the primary virtual processor also to the backup computer so that the virtual processor in the backup computer receives the same responses as the virtual processor running in the primary computer.
  • Context switches and user-mode drain instructions are also handled by the virtual machine monitors so as to ensure that all resulting calls to interrupt and trap handling routines are executed at identical points in the instruction streams of both the primary and backup virtual processors.
  • a special I/O protocol is defined. Whenever such a failure occurs, all I/O operation interrupts signifying operation completion that are received during the epoch in which a failure occurrs are deleted, and interrupts signifying an unknown outcome are generated for all I/O devices that are in use.
  • the "unknown outcome" interrupts are processed by the backup virtual processor at the beginning of the next epoch, the software in the backup virtual processor re-initiates all outstanding I/O operations for which a "completed" interrupt was not received.
  • processor failures look like ordinary I/O device failures to the software running in the backup virtual processor.
  • programs that correctly cope with ordinary I/O failures will work correctly even when I/O operations are outstanding during a fail-over.
  • Figure 1 is a block diagram of a fault tolerant computer system in accordance with the present invention.
  • Figure 2 is a block diagram of software routines and data structures associated with a virtual machine monitor that implements fault tolerance.
  • Figure 3 is a flow chart of the epoch handling routine executed by the virtual machine monitors of the preferred embodiment.
  • Figure 4 represents an interrupt information message sent from a primary computer to a backup computer.
  • Figure 5 represents the data structures that store buffered trap information.
  • Figure 6 is a flow chart of the process for executing a modified user-mode drain instruction.
  • Figure 7 is a flow chart of the process for handling I/O operations during a fail-over from the primary to the backup computer.
  • Figure 1 shows a fault tolerant computer system 100 having a primary computer 102, a backup computer 104 and a FIFO communication channel 106 used primarily for transferring information from the primary to the backup computer.
  • Each computer has its own primary memory 112, 114, but the two share use of other input/output (I/O) devices 120, including disk storage devices and any other secondary memory devices that are used by software in both computers.
  • I/O input/output
  • Virtual Machine Monitor Before discussing how the two computers 102 and 104 work together in the preferred embodiment, it is important to first explain the "virtual machine" configuration of the software in just one of the machines 102 because the fault tolerance methodology of the present invention makes use of "virtual machines” 130, 132 and a virtual machine monitor 134.
  • VMF370 virtual machine monitor software
  • the VMF370 operating system software was used to simulate multiprocessing by using only a single real processor, allowing each virtual machine to act as though it were running on a processor unto itself.
  • One of the known properties of the virtual machine monitor technology is that it allows clients using the same computer processor at the same time to be running applications under different operating systems. This is quite different from a typical computer system, in which only one operating system can be running at a time and that operating system manages all of the system's resources on behalf of all users.
  • VMM virtual machine monitor's
  • the VMM allows all user-mode instructions to be executed by the CPU without interference, while kernel-mode instructions are trapped by the VMM and then processed so as to keep track of and isolate the machine state associated with each virtual machine running on the computer (see Figure 2).
  • the programs running in each virtual machine are executed in user mode, even though the virtual machine includes its own operating system.
  • the trap vectors for user-mode instruction violations are set so that control of all such traps is passed to the VMM.
  • all kernel-mode instructions are automatically trapped and passed to the virtual machine monitor for handling.
  • interrupt vectors for the (real) machine are set so that such interrupts are passed to the VMM, which then dispatches them to the correct virtual machine.
  • virtual processor is used interchangeably with the term “virtual machine”.
  • Recovery Register is defined herein to mean a register that is automatically decremented (or incremented) each time that an instruction is executed and that causes a hardware interrupt of type "recovery register” when its value reaches zero (or any other fixed, predefined value).
  • recovery register it is possible to ensure that control is passed to a virtual machine monitor at a predetermined point in the instruction stream of any virtual processor. The use of the recovery register in the context of the present invention will be described below.
  • the present invention also requires that all user-mode instructions executed b both computers produce identical results in all cases, even in situations in whic the result is normally "undefined".
  • the processors 102, 104 mus be designed so that when an error condition such as an overflow occurs, the resulting data values and virtual machine states in the two processors are identical.
  • VMM Virtual machine monitor
  • Primary virtual processor 130 and backup virtual processor 140 contain identical software, including both operating system software and application software. For this reason, the backup virtual processor 140 is also called a "shadow" virtual processor.
  • a fault monitor 150 which sends a signal to the backup computer 104.
  • the VMM 144 in the backup computer will receive this signal, causing that computer to change the Processor_Type of the virtual processor 140 to "Primary".
  • the changeover from backup mode to primary mode by the backup VMM 144, called a fail-over is discussed in more detail below with regard to the handling of input output (I/O) instructions.
  • I/O input output
  • interrupt is defined to mean an asynchronous signal generated by hardware events, such as input/output interrupts, and clock interrupts.
  • traps are defined to mean synchronous signals generated in response to the execution of software instructions, such as traps caused by an overflow or underflow.
  • interrupts are, by definition, asynchronous, it is difficult to control when an interrupt will be delivered to the virtual processor.
  • the present invention overcomes this problem by delaying the delivery of interrupts until a well defined point in the instruction streams of both the primary and virtual processors.
  • delaying or “buffering” interrupts and traps means storing representations of these events and preventing the virtual processors involved from immediately processing those events.
  • delivering buffered interrupts and traps to a virtual processor (or to a hardware processor) means initiating processing of those interrupts and traps by the interrupt handling and trap handling routines in the virtual processor's operating system.
  • interrupts are handled by the VMM 134. More particularly, interrupts cause the computer processor to begin execution of a kernel-mode interrupt handling routine that is part of the VMM. Since interrupts are asynchronous, the VMM 134 has some freedom to delay the delivery of interrupts to its virtual processors. In accordance with the present invention, the VMM 134 exploits this freedom by delaying delivery of an interrupt so that delivery of interrupts always occurs at a planned, fixed point in a virtual processor's instruction stream. The way this is done is as follows.
  • Each virtual processor 130 is allowed to execute for an "epoch" comprising a known number of instructions (e.g., 100 instructions) before transferring control to the VMM 134.
  • the two VMMs 134 and 144 running on the primary computer and backup computer must use the same scheme for computing epoch lengths to ensure that a backup virtual processor 140 transfers control to its VMM 144 at the same point in its instruction stream as the primary virtual processor 130 transfers control to its VMM 34.
  • an epoch's length is determined by counting instructions executed.
  • the "best" epoch length is determined by balancing the following countervailing factors: (A) the overhead associated with the epoch mechanism is inversely proportional to the epoch's length, and (B) excessive delay in delivery of interrupts and traps can hurt system performance.
  • each epoch is terminated either when (A) the recovery register causes an interrupt because the specified number of instructions has been executed, or (B) a drain instruction is executed, whichever occurs first. Drain instructions are discussed more below.
  • This modified epoch ending mechanism satisfies the criteria that all instructions for handling interrupts and traps must be executed at exactly the same point in the two virtual processors' instruction streams. Note that both virtual processors will encounter drain instructions at exactly the same points in their instruction streams.
  • Steps 200, 204, and 206 are handled by the Virtual Machine Monitors in the two computers, while step 202 represents the user-mode process being run by the primary and backup virtual processors.
  • the recovery register 122, 124 (see Figure 1) for each virtual processor 130 and 140 is set to a known value (step 200) and is decremented once for each command executed by the corresponding virtual processor (step 202). Execution of the software running in each virtual processor is allowed to continue for one "epoch" (step 202) until the recovery register reaches a value of zero. During this time, all interrupts are buffered by the VMM (step 204) and messages containing copies of those interrupts and any associated data (e.g., the data from a disk read operation) are sent to the backup computer over a first in first out (FIFO) communication channel 106 (see Figure 1).
  • FIFO first in first out
  • the copies of the buffered interrupts are delivered in the form of messages to the backup computer, and the virtual machine monitor in the backup computer is programmed to receive those messages and to treat them as though they were ordinary interrupts.
  • the backup computer's virtual machine monitor buffers the received interrupts in the same way as the primary computer's virtual machine monitor buffers those interrupts.
  • the format of messages sent in step 204 that inform the backup VMM of th interrupts that occurred during each epoch on the primary computer is show in Figure 4. By tagging each message with the epoch number during which i was received by the primary virtual processor, the backup VMM can determine which interrupts to deliver to the backup VMM 144 at each new epoch, even if the backup computer is running more than one epoch behind the vast computer.
  • the recovery register could be used to count any quantity (e.g., memory references) that allows a virtual processor's instruction stream to be broken at a predictable point.
  • Traps are like interrupts, but result from execution of an instruction. Traps that are delivered immediately following execution of the causative instruction offer no problem for ensuring that a primary and a backup virtual processor execute the same sequence of instructions. This is because if an instruction causes a trap when executed by the primary virtual processor, it will cause the same trap when it is executed by the backup. Thus, the next instruction executed by the primary and the backup will be the same: the next instruction executed will be the first instruction of the trap handling routine.
  • the user-mode drain instruction does not cause a problem. Since most RISC processors cannot guarantee immediate delivery of all traps, operation of the user-mode drain instruction as described above is inconsistent with the need for the primary and backup virtual processors to have identical instruction streams.
  • the data structure used for the delivery of a set of buffered traps in the preferred embodiment is shown in Figure 5.
  • the computer has an array 210 of registers, where each register 212 in the array is capable of storing a data value that results from executing an instruction.
  • any register 212 can be the destination associated with an instruction whose execution causes a trap.
  • Two vectors V1 and V2 (214, 216) are set by the computer hardware to indicate what traps have occurred, as follows:
  • VI (i) 1 if Register i was the destination for an instruction that caused a trap, e.g., if Register i contains a potentially erroneous value such as an overflow value;
  • V2(i) 1 if a Class i trap has occurred.
  • bits associated with each delivered trap are added to copies of vectors V1 and V2 being maintained by the VMM.
  • vectors V1 and V2 of the VMM are delivered to the virtual processor for handling of the buffered traps.
  • VMM 144 on the backup computer ignores interrupts delivered by its own hardware and uses instead messages received from the primary VMM 134, the backup VMM 144 processes its own traps. This is possible because the traps that occur in epoch X in the primary virtual processor will also be delivered by the end of epoch X in the backup virtual processor, since the epoch end coincides with a context switch.
  • the present invention modifies operation of the user-mode drain instruction as follows.
  • a new status bit called Drain_Priv
  • the user-mode drain instruction operates as described above (see steps 250, 252, 254).
  • Drain_Priv is equal to 1
  • execution of the drain instruction in user-mode causes an immediate context switch (step 256) to an appropriate handler.
  • the context switch causes all traps buffered by the hardware, if any, to be flushed and delivered to the VMM 134 (step 258).
  • the VMM 134 adds the flushed traps to those already represented by the VMM versions of vectors V1 and V2, and then immediately delivers all of the buffered traps (including both those previously buffered and those just now flushed by the drain instruction) in the form of vectors V1 and V2 to the virtual processor 130. Then, control is returned by the VMM to the virtual processor (step 260) for executing the next instruction.
  • the backup virtual processor reaches the same drain instruction in its instruction stream, it will perform exactly the same steps as the primary virtual processor in the same sequence.
  • the present invention when using a hardware processor that can delay the delivery of traps, the present invention (A) buffers all delivered traps in the VMM and delivers them at the start of each epoch, and (B) modifies operation of the user-mode drain instruction so as to cause a context switch, which causes hardware-buffered traps to be delivered, followed by delivery of all traps buffered by both the hardware and the VMM to the virtual processor.
  • One alternate embodiment of the above described process for handling drain instructions is as follows. In some situations it may not be advantageous to deliver delayed traps (in response to a drain instruction) to the virtual processor before delivery of interrupts that occurred prior to the delayed traps. To avoid this potential problem, the epoch control software in the VMMs is modified to end each epoch when (A) the recovery register reaches a value of zero due to execution of the specified number of instructions, or (B) a drain instruction is executed, whichever occurs first. By ending the epoch, control is passed to step 206 in Figure 3, causing all delayed interrupts as well as all delayed traps to be delivered to the virtual processor.
  • the backup computer When the primary computer fails (step 304 of Figure 7), the backup computer must take over as the new primary computer. Thus, the backup's VMM 144 must go from a mode in which it is suppressing all of the backup virtual processor's I/O operations to a mode where the backup virtual processor's operations are actually performed. This "fail-over" must be done in such a way that the environment is unaware of the primary computer's failure and is accomplished as described below.
  • the present invention avoids this "exactly once" problem with a new I/O interface specification.
  • This specification is the one seen by virtual processors, and so it is what operating system programmers and applications programmers must comply with when using the present invention.
  • I/O operations are governed by the following rules (the rules are formulated for an interrupt-based I/O interface but have analogues for other I/O models):
  • connection (which is a type of I/O operation) to the device that will perform that operation.
  • the (virtual) processor can then issue I/O operations to that device.
  • Rule 1 Delivery of a "completed" interrupt for an issued I/O operation implies that the operation was performed by the device.
  • Rule 2 Concurrent I/O operations can be performed by the I/O device in any order.
  • Rule 3 Delivery of a "disconnected" interrupt from an I/O device to a processor implies that outstanding I/O operations (i.e., those for which a "completed” interrupt has not been delivered) may or may not have been performed.
  • the VMM on the backup computer changes the setting of an internal "processor type" variable to indicate that it is now the primary processor (step 308), and it may also execute a number of other fail-over transition routines that are not relevant to the present invention (e.g., sending a hardware failure message to the system's operator console) (step 310). Then the backup VMM proceeds to take care of I/O operations in process during the fail-over.
  • the backup VMM deletes all I/O "completed" interrupts received during the epoch in which the failure occurred (step 312).
  • the backup VMM knows the epoch in which the failure occurred because it receives an epoch message for every epoch that the primary VMM has completed. Thus if the last epoch message received from the primary VMM prior to failure was for epoch X, then the failure occurred in epoch X+1.
  • the backup VMM adds to the set of buffered interrupts that will be delivered at the start of epoch X+1 , a "disconnected" interrupt for every I/O connection open during epoch X (step 314).
  • the backup virtual processor After steps 312 and 314, the backup virtual processor completes execution of epoch X (step 316), then the backup VMM delivers all buffered interrupts and traps to the backup virtual processor (step 318). Since the buffered interrupts include disconnects for all the I/O devices in use, the operating system or application software (whichever initiated connection to the I/O devices) running in the backup virtual processor will send new "connect" signals to all the I/O devices, and it will also reissue all outstanding I/O operations that were issued before epoch X+1 and for which a "completed" interrupt was not received (step 320).
  • the techniques of the present invention could be implemented without using virtual processors and virtual machine monitors.
  • the delivery points would be defined by a recovery register or similar mechanism with additional delivery points occurring as defined by drain instructions.
  • the handling of I/O operations disrupted by a fail-over would be the same as described above.
  • This alternate embodiment of the present invention is suitable for use with RISC processors and satisfies the requirement of needing identical instruction streams in the primary and backup processors even though interrupts and traps may be delayed in each processor in an unpredictable fashion.

Abstract

A fault-tolerant computer system has primary and backup computers. Primary and backup virtual machines running on the computers are controlled by corresponding virtual machine monitors. The virtual machines execute only user-mode instructions, while all kernel-mode instructions are trapped and handled by the virtual machine monitors. Each computer has a recovery register that generates a hardware interrupt each time that a specified number of instructions, called an epoch, are executed. Prior to failure of the primary computer, the backup computer's virtual machine monitor converts all I/O instructions into no-ops and the primary computer sends copies of all I/O interrupts to the backup computer. To ensure that the instruction streams in the primary and backup virtual machines are identical and that all instructions for handling interrupts and traps are executed at exactly the same point in the two virtual machines' instruction streams, all interrupts and traps that occur on the primary computer during an epoch are buffered by the virtual machine monitor. At the end of each epoch, the buffered interrupts and traps are delivered to the primary computer's virtual machine and a message is sent to the backup computer allowing the just completed epoch to be executed by the backup virtual machine. Whenever a fail-over occurs, all I/O operation completed interrupts for the epoch in which the failure occurred are deleted, and 'disconnected' interrupts are generated for all I/O devices in use. The backup virtual processor re-connects to the I/O devices and then reissues outstanding I/O operations for which an operation completed interrupt was not received. As a result, processor failures look like ordinary I/O device failures to the software running in the backup virtual machine.

Description

FAULT TOLERANT COMPUTER PROCESSING USING A SHADOW VIRTUAL PROCESSOR
The present invention relates generally to fault tolerant computer systems and particularly to methods and systems for providing fault tolerance that is independent of the computer's operating system software.
BACKGROUND OF THE INVENTION
The basic scheme used in most fault tolerant computer systems is to employ a primary and a backup computer. Users interact with the primary computer in order to obtain a service. The primary computer performs the tasks requested by users and informs the backup of its actions so that the backup can take over providing the service if the primary computer fails. Thus, hardware failures in the primary computer do not cause interruption of service to the users.
In a properly implemented instance of this scheme, the backup processor must generate no interactions with its environment before the primary computer has failed. And, after the primary computer has failed, the backup processor must generate interactions with its environment in such a way that the environment is unaware of the primary computer's failure. Fault tolerance in computers is usually implemented either (A) by constructing special purpose computer hardware, or (B) by modifying the computer's operating system. The special purpose hardware approach requires hardware that is intimately related to the computer processor's design. As a result, such computers are usually unnecessarily costly for clients who do not require fault tolerance.
The major problem associated with using special purpose operating system code to implement fault tolerance is that the only operating system that can be used on that computer system, and still maintain fault tolerance, is the operating system containing the special purpose code. If a user who needs fault tolerance wants to use another operating system with the computer, extensive (and thus expensive) changes to this second operating system will be required.
The goals of the present invention are (1) to provide a fault tolerant computer system with little or no added cost for clients and processes that do not require fault tolerance, and (2) to provide a fault tolerance mechanism that works regardless of the operating system software used by the system's clients. While the "special-purpose hardware" and "modified operating system" approaches are both capable of meeting the basic requirements for a fault-tolerant computer system, the present invention overcomes cost problems associated with the "special-purpose hardware" approach and provides more flexibility in terms of operating system selection than the "modified operating system" approach.
SUMMARY OF THE INVENTION
In summary, the present invention is a fault-tolerant computer system having a primary computer and backup computer, both of which are capable of delaying delivery of interrupts to the computer's operating system due to aysnchronous activity by associated components (e.g., I/O devices) and the use of instruction pipelining. Primary and backup virtual processors running on the computers are controlled by corresponding virtual machine monitors. Only programs that need to be fault tolerant are run on both the primary and backup virtual processors. Programs that do not need to be fault tolerant are run on other virtual processors on the primary computer. The computer hardware directly executes only user-mode instructions of the virtual processors, but all kernel-mode instructions (e.g., I/O instructions) are trapped and handled by the virtual machine monitors.
Each computer also has a recovery register that can be loaded with a value and that generates a hardware interrupt after a corresponding number of instructions are executed. The hardware interrupts generated by the recovery register mark the boundary between "epochs" in the instruction stream.
Prior to failure of the primary computer, the backup virtual processor generates no interactions with its environment. To accomplish this, the backup computer's virtual machine monitor converts all I/O instructions into no-ops (i.e., null operations). Furthermore, the primary computer sends copies of all I/O interrupts destined for delivery to the primary virtual processor also to the backup computer so that the virtual processor in the backup computer receives the same responses as the virtual processor running in the primary computer.
To ensure that the instruction streams in the primary and backup virtual processors are identical and that all instructions for handling interrupts and traps are executed at exactly the same point in the two virtual processors' instruction streams, all interrupts and traps that occur on the primary computer (and are destined for the primary virtual processor) during an epoch are buffered by the virtual machine monitor, with copies of the interrupts being sent to the backup computer. Then, at the end of the epoch, all the buffered interrupts and traps are delivered to the primary computer's virtual processor by the virtual machine monitor on the primary computer. In addition, a message is sent to the backup computer allowing the just completed epoch to be executed by the backup virtual processor. As a result, all interrupt and trap handling routines are executed at identical points in the instruction streams of both the primary and backup virtual processors.
Context switches and user-mode drain instructions are also handled by the virtual machine monitors so as to ensure that all resulting calls to interrupt and trap handling routines are executed at identical points in the instruction streams of both the primary and backup virtual processors.
Finally, to provide seamless transfer of I/O operations after failure of the primary computer, a special I/O protocol is defined. Whenever such a failure occurs, all I/O operation interrupts signifying operation completion that are received during the epoch in which a failure occurrs are deleted, and interrupts signifying an unknown outcome are generated for all I/O devices that are in use. When the "unknown outcome" interrupts are processed by the backup virtual processor at the beginning of the next epoch, the software in the backup virtual processor re-initiates all outstanding I/O operations for which a "completed" interrupt was not received. As a result, processor failures look like ordinary I/O device failures to the software running in the backup virtual processor. Furthermore, programs that correctly cope with ordinary I/O failures will work correctly even when I/O operations are outstanding during a fail-over.
BRIEF DESCRIPTION OF THE DRAWINGS
Additional objects and features of the invention will be more readily apparent from the following detailed description and appended claims when taken in conjunction with the drawings, in which: Figure 1 is a block diagram of a fault tolerant computer system in accordance with the present invention.
Figure 2 is a block diagram of software routines and data structures associated with a virtual machine monitor that implements fault tolerance.
Figure 3 is a flow chart of the epoch handling routine executed by the virtual machine monitors of the preferred embodiment.
Figure 4 represents an interrupt information message sent from a primary computer to a backup computer.
Figure 5 represents the data structures that store buffered trap information.
Figure 6 is a flow chart of the process for executing a modified user-mode drain instruction.
Figure 7 is a flow chart of the process for handling I/O operations during a fail-over from the primary to the backup computer.
DESCRIPTION OF THE PREFERRED EMBODIMENT
Figure 1 shows a fault tolerant computer system 100 having a primary computer 102, a backup computer 104 and a FIFO communication channel 106 used primarily for transferring information from the primary to the backup computer. Each computer has its own primary memory 112, 114, but the two share use of other input/output (I/O) devices 120, including disk storage devices and any other secondary memory devices that are used by software in both computers.
Virtual Machine Monitor Before discussing how the two computers 102 and 104 work together in the preferred embodiment, it is important to first explain the "virtual machine" configuration of the software in just one of the machines 102 because the fault tolerance methodology of the present invention makes use of "virtual machines" 130, 132 and a virtual machine monitor 134.
The use of virtual machines and virtual machine monitors by International Business Machines in conjunction with the IBM360/67, IBM370, and its successor processors is well known and well documented. The VMF370 operating system software was used to simulate multiprocessing by using only a single real processor, allowing each virtual machine to act as though it were running on a processor unto itself. One of the known properties of the virtual machine monitor technology is that it allows clients using the same computer processor at the same time to be running applications under different operating systems. This is quite different from a typical computer system, in which only one operating system can be running at a time and that operating system manages all of the system's resources on behalf of all users.
The basic idea behind the virtual machine monitor's (VMM's) operation is as follows. All instructions are divided into two groups: user-mode instructions and kernel-mode instructions. The basis of this partition is well known and documented, with respect to virtual machines, for example in Popeck and Goldberg, "Formal Requirements for Virtualizable Third Generation Architectures," Communications of the ACM, Vol. 17, No. 7, pp. 412-421, July 1974.
The VMM allows all user-mode instructions to be executed by the CPU without interference, while kernel-mode instructions are trapped by the VMM and then processed so as to keep track of and isolate the machine state associated with each virtual machine running on the computer (see Figure 2). To implement this, the programs running in each virtual machine are executed in user mode, even though the virtual machine includes its own operating system. Also, the trap vectors for user-mode instruction violations are set so that control of all such traps is passed to the VMM. As a result, since the virtual machine is running in user mode, all kernel-mode instructions are automatically trapped and passed to the virtual machine monitor for handling. Finally, interrupt vectors for the (real) machine are set so that such interrupts are passed to the VMM, which then dispatches them to the correct virtual machine.
While there are, of course, many more details and complexities involved in the actual design and implementation of a virtual machine monitor, the design and implementation of virtual machine monitors is known to those skilled in the art. Therefore, this document shall discuss only those aspects of virtual machine monitors that are not already known to those skilled in the art.
For the purposes of this document, the term "virtual processor" is used interchangeably with the term "virtual machine".
Hardware Requirements for Fault Tolerant System: Recovery Register and Functional User-mode Instructions The present invention requires that the primary computer 102 and backup computer 104 both have a recovery register 122, 124. The term "recovery register" is defined herein to mean a register that is automatically decremented (or incremented) each time that an instruction is executed and that causes a hardware interrupt of type "recovery register" when its value reaches zero (or any other fixed, predefined value). Thus, on a processor that has a recovery register, it is possible to ensure that control is passed to a virtual machine monitor at a predetermined point in the instruction stream of any virtual processor. The use of the recovery register in the context of the present invention will be described below. The present invention also requires that all user-mode instructions executed b both computers produce identical results in all cases, even in situations in whic the result is normally "undefined". For instance, the processors 102, 104 mus be designed so that when an error condition such as an overflow occurs, the resulting data values and virtual machine states in the two processors are identical.
Implementing Fault Tolerance With Virtual Machine Monitor
In the preferred embodiment, all programs on the primary computer 102 that need to be made fault tolerant are executed by a first virtual processor 130, while programs that do not need to be fault tolerant are executed on one or more other virtual processors 132 in the primary and backup computers. Virtual machine monitor (VMM) 134 maintains a separate machine state for each virtual processor, and generally performs the functions of a VMM as described above.
Primary virtual processor 130 and backup virtual processor 140 contain identical software, including both operating system software and application software. For this reason, the backup virtual processor 140 is also called a "shadow" virtual processor. The VMMs 134 and 144 for the primary and backup computers are not completely identical, however, because the backup computer's VMM 144 contains a state variable (herein called Processor_Type 146, see Figure 2) indicating that virtual processor 140 is the backup virtual processor (Processor_Type=Backup). Since the backup virtual processor 140 is not supposed to affect its environment in any way (unless the primary computer has failed), all I/O instructions executed by the backup virtual processor 140 are converted into no-ops (i.eM null operations) by the backup VMM 144.
If the primary computer 102 fails, that condition is detected by a fault monitor 150, which sends a signal to the backup computer 104. The VMM 144 in the backup computer will receive this signal, causing that computer to change the Processor_Type of the virtual processor 140 to "Primary". The changeover from backup mode to primary mode by the backup VMM 144, called a fail-over, is discussed in more detail below with regard to the handling of input output (I/O) instructions. There are many different ways of implementing a fault monitor 150 known to those skilled in the art. The type of fault monitor does not matter so long as the backup computer 104 is reliably informed whenever a hardware fault in the primary computer 102 causes that computer to stop working and the last message sent along communications channel 106 is received before the failure is signalled.
Delivery of Interrupts and Use of Epochs
For the purposes of this document, the term "interrupt" is defined to mean an asynchronous signal generated by hardware events, such as input/output interrupts, and clock interrupts. By way of contrast, "traps" are defined to mean synchronous signals generated in response to the execution of software instructions, such as traps caused by an overflow or underflow.
The major problem associated with interrupts is that because they are, by definition, asynchronous, it is difficult to control when an interrupt will be delivered to the virtual processor. In the context of the present invention, it is essential that the primary and backup virtual processors execute exactly the same instructions in exactly the same order. If asynchronous interrupts were delivered to the primary and backup virtual processors at different points in their instruction streams, this requirement would be violated. The present invention overcomes this problem by delaying the delivery of interrupts until a well defined point in the instruction streams of both the primary and virtual processors.
It should be noted that "delaying" or "buffering" interrupts and traps means storing representations of these events and preventing the virtual processors involved from immediately processing those events. Conversely, "delivering" buffered interrupts and traps to a virtual processor (or to a hardware processor) means initiating processing of those interrupts and traps by the interrupt handling and trap handling routines in the virtual processor's operating system.
All interrupts are handled by the VMM 134. More particularly, interrupts cause the computer processor to begin execution of a kernel-mode interrupt handling routine that is part of the VMM. Since interrupts are asynchronous, the VMM 134 has some freedom to delay the delivery of interrupts to its virtual processors. In accordance with the present invention, the VMM 134 exploits this freedom by delaying delivery of an interrupt so that delivery of interrupts always occurs at a planned, fixed point in a virtual processor's instruction stream. The way this is done is as follows.
Each virtual processor 130 is allowed to execute for an "epoch" comprising a known number of instructions (e.g., 100 instructions) before transferring control to the VMM 134. The two VMMs 134 and 144 running on the primary computer and backup computer must use the same scheme for computing epoch lengths to ensure that a backup virtual processor 140 transfers control to its VMM 144 at the same point in its instruction stream as the primary virtual processor 130 transfers control to its VMM 34. In the preferred embodiment, an epoch's length is determined by counting instructions executed. For any particular implementation of the present invention, the "best" epoch length is determined by balancing the following countervailing factors: (A) the overhead associated with the epoch mechanism is inversely proportional to the epoch's length, and (B) excessive delay in delivery of interrupts and traps can hurt system performance.
In one preferred embodiment, each epoch is terminated either when (A) the recovery register causes an interrupt because the specified number of instructions has been executed, or (B) a drain instruction is executed, whichever occurs first. Drain instructions are discussed more below. This modified epoch ending mechanism satisfies the criteria that all instructions for handling interrupts and traps must be executed at exactly the same point in the two virtual processors' instruction streams. Note that both virtual processors will encounter drain instructions at exactly the same points in their instruction streams.
The flow chart in Figure 3 represents processing in both the primary and backup computers. Steps 200, 204, and 206 are handled by the Virtual Machine Monitors in the two computers, while step 202 represents the user-mode process being run by the primary and backup virtual processors.
Referring to Figures 1 and 3, at the beginning of each epoch, the recovery register 122, 124 (see Figure 1) for each virtual processor 130 and 140 is set to a known value (step 200) and is decremented once for each command executed by the corresponding virtual processor (step 202). Execution of the software running in each virtual processor is allowed to continue for one "epoch" (step 202) until the recovery register reaches a value of zero. During this time, all interrupts are buffered by the VMM (step 204) and messages containing copies of those interrupts and any associated data (e.g., the data from a disk read operation) are sent to the backup computer over a first in first out (FIFO) communication channel 106 (see Figure 1).
The copies of the buffered interrupts are delivered in the form of messages to the backup computer, and the virtual machine monitor in the backup computer is programmed to receive those messages and to treat them as though they were ordinary interrupts. Thus, the backup computer's virtual machine monitor buffers the received interrupts in the same way as the primary computer's virtual machine monitor buffers those interrupts. The format of messages sent in step 204 that inform the backup VMM of th interrupts that occurred during each epoch on the primary computer is show in Figure 4. By tagging each message with the epoch number during which i was received by the primary virtual processor, the backup VMM can determine which interrupts to deliver to the backup VMM 144 at each new epoch, even if the backup computer is running more than one epoch behind the primar computer.
At the end of the epoch, when the recovery register in the primary computer 122 reaches a value of zero, all the interrupts buffered by the primary computer's VMM 134 (as well as all traps buffered by the VMM, as discussed below) are delivered to the primary virtual processor 130 (step 206), and then a message containing the number of the completed epoch is sent to the backup computer. At the backup computer, receipt of this epoch number message initiates execution of the instructions for the next epoch. This execution starts with the delivery of all buffered interrupts that were deliverd to the primary virtual processor at this epoch boundary. Thus, the backup virtual processor always runs at least one epoch behind the primary virtual processor. As will be described below, this one epoch delay is required for successfully handling failures of the primary virtual processor that occur mid-epoch during the processing of I/O operations.
In alternate embodiments, the recovery register could be used to count any quantity (e.g., memory references) that allows a virtual processor's instruction stream to be broken at a predictable point.
Delivery of Delayed Traps
Traps are like interrupts, but result from execution of an instruction. Traps that are delivered immediately following execution of the causative instruction offer no problem for ensuring that a primary and a backup virtual processor execute the same sequence of instructions. This is because if an instruction causes a trap when executed by the primary virtual processor, it will cause the same trap when it is executed by the backup. Thus, the next instruction executed by the primary and the backup will be the same: the next instruction executed will be the first instruction of the trap handling routine.
When the computer hardware being used is such that trap delivery can be delayed (i.e., by the computer hardware), the exact amount of delay is typically undefined, meaning that two computers running the exact same program might deliver delayed traps at different points in the program's instruction stream. Whenever this is the case, and it often is for RISC processors (reduced instruction set computers) or when instruction pipelining is employed, other actions need to be taken to ensure that traps are delivered to the backup virtual processor at the same point in its instruction stream as where it was delivered to the primary virtual processor.
Usually, on RISC processors and other computers that use instruction pipelining techniques, the programmer can force delivery of a delayed trap. For the purposes of this discussion, we will assume that traps may be delivered at any time after they have occurred, but that all undelivered traps will be delivered by the processor upon:
(A) execution of a special user-mode "drain" instruction;
(B) a processor context switch in response to a (hardware) interrupt; or
(C) execution of an instruction that causes the processor to leave user-mode.
For example, whenever a user-mode drain instruction is executed, all traps that have been buffered by the computer hardware are flushed and delivered immediately (in, for example, the form of vectors V1 and V2) to the virtual processor that executed the drain instruction. Conventionally, the drain instruction forces immediate delivery of delayed traps, but does not force delivery of pending interrupts. To understand the nature of the problem presented by delayed traps, consider the following situation. Suppose the primary virtual processor executes an instruction that leads to delivery of a trap. The VMM's trap handler software would actually receive control for the purpose of processing the trap. However, the VMM cannot buffer the trap, because it if does, subsequent execution of the user-mode drain instruction by the virtual processor would cause that virtual processor to receive information regarding only a subset of its delayed traps. In particular, the virtual processor would not learn of the trap already buffered by the VMM.
On the other hand, if the VMM can and does immediately deliver all traps to the primary virtual processor, then the user-mode drain instruction does not cause a problem. Since most RISC processors cannot guarantee immediate delivery of all traps, operation of the user-mode drain instruction as described above is inconsistent with the need for the primary and backup virtual processors to have identical instruction streams.
We must ensure that the same trap will be delivered to the backup virtual processor at the same point in its instruction stream. To do this, two mechanisms are used: (1) delivery at the end of each epoch of hardware buffered and VMM buffered traps, and (2) use of a modified drain instruction so as to ensure delivery of all hardware buffered and VMM buffered traps upon execution of user-mode drain instructions.
The data structure used for the delivery of a set of buffered traps in the preferred embodiment is shown in Figure 5. In particular, the computer has an array 210 of registers, where each register 212 in the array is capable of storing a data value that results from executing an instruction. In the preferred embodiment, any register 212 can be the destination associated with an instruction whose execution causes a trap. Two vectors V1 and V2 (214, 216) are set by the computer hardware to indicate what traps have occurred, as follows:
VI (i) = 1 if Register i was the destination for an instruction that caused a trap, e.g., if Register i contains a potentially erroneous value such as an overflow value;
V2(i) = 1 if a Class i trap has occurred.
Using this scheme, a program will need to determine which register is associated with each type of trap, as will be understood by those skilled in the art. In alternate embodiments, data describing traps that have occurred could be stored in other ways, so long as all the traps are buffered for later processing.
During the middle of an epoch, bits associated with each delivered trap are added to copies of vectors V1 and V2 being maintained by the VMM. At the end of each epoch, vectors V1 and V2 of the VMM are delivered to the virtual processor for handling of the buffered traps.
In accordance with the present invention, while VMM 144 on the backup computer ignores interrupts delivered by its own hardware and uses instead messages received from the primary VMM 134, the backup VMM 144 processes its own traps. This is possible because the traps that occur in epoch X in the primary virtual processor will also be delivered by the end of epoch X in the backup virtual processor, since the epoch end coincides with a context switch.
To overcome the problem associated with the fact that both the computer hardware and the VMM can buffer traps, the present invention modifies operation of the user-mode drain instruction as follows. As shown in Figure 2, a new status bit, called Drain_Priv, is added to the machine state of the hardware processor. Referring to Figure 6, when the Drain_Priv flag is reset (i.e., equal to 0), the user-mode drain instruction operates as described above (see steps 250, 252, 254). However, when Drain_Priv is equal to 1 (see step 252), execution of the drain instruction in user-mode causes an immediate context switch (step 256) to an appropriate handler. The context switch causes all traps buffered by the hardware, if any, to be flushed and delivered to the VMM 134 (step 258). The VMM 134 adds the flushed traps to those already represented by the VMM versions of vectors V1 and V2, and then immediately delivers all of the buffered traps (including both those previously buffered and those just now flushed by the drain instruction) in the form of vectors V1 and V2 to the virtual processor 130. Then, control is returned by the VMM to the virtual processor (step 260) for executing the next instruction. When the backup virtual processor reaches the same drain instruction in its instruction stream, it will perform exactly the same steps as the primary virtual processor in the same sequence.
Operation of the drain instruction is not modified for virtual processors with Drain_Priv=0, so as to minimize impact of the invention on processes not requiring fault tolerance.
In summary, when using a hardware processor that can delay the delivery of traps, the present invention (A) buffers all delivered traps in the VMM and delivers them at the start of each epoch, and (B) modifies operation of the user-mode drain instruction so as to cause a context switch, which causes hardware-buffered traps to be delivered, followed by delivery of all traps buffered by both the hardware and the VMM to the virtual processor.
One alternate embodiment of the above described process for handling drain instructions is as follows. In some situations it may not be advantageous to deliver delayed traps (in response to a drain instruction) to the virtual processor before delivery of interrupts that occurred prior to the delayed traps. To avoid this potential problem, the epoch control software in the VMMs is modified to end each epoch when (A) the recovery register reaches a value of zero due to execution of the specified number of instructions, or (B) a drain instruction is executed, whichever occurs first. By ending the epoch, control is passed to step 206 in Figure 3, causing all delayed interrupts as well as all delayed traps to be delivered to the virtual processor.
Transparent Fail-Over For Handling I/O Operations
As discussed above with reference to Figure 1 , since the backup virtual processor 140 is not supposed to affect its environment in any way (unless the primary computer has failed), all I/O instructions executed by the backup virtual processor 140 are converted into no-ops by the backup VMM 144. Since all instructions that start I/O operations are kernel-mode instructions, they are all intercepted by the VMM 144, which then converts them into no-ops as long as the primary computer has not failed (see steps 300 and 302 in Figure 7). Thus, only the I/O operations from the primary computer 102 affect the environment (disks, etc.)
Even though I/O operations issued by the backup virtual processor 140 are translated into no-ops, a program being executed by the backup virtual processor has no way of knowing this fact. Each I/O operation issued by the backup is necessarily also issued by the primary, and each I/O interrupt that is delivered by the primary VMM 134 to the primary virtual processor 130 is also delivered to the backup VMM 144 for delivery to the backup virtual processor 140 (see Figure 3).
When the primary computer fails (step 304 of Figure 7), the backup computer must take over as the new primary computer. Thus, the backup's VMM 144 must go from a mode in which it is suppressing all of the backup virtual processor's I/O operations to a mode where the backup virtual processor's operations are actually performed. This "fail-over" must be done in such a way that the environment is unaware of the primary computer's failure and is accomplished as described below.
Implementing the fail-over is made difficult by the fact that it is usually not possible for the primary computer to (1) issue an I/O operation and (2) inform the backup of this fact, all in a single atomic (indivisible) action. As a result, it is possible for the primary computer to fail between these two operations. Also, no protocol exists that ensures an I/O operation will be performed exactly once (by either the primary or else the backup) if it is possible for there to be a failure between actions (1) and (2) above. Notice that the requirement that "the environment be unaware of the primary computer's failure" is violated if an I/O operation is skipped or performed more than once just because a failure occurred at an inopportune moment.
The present invention avoids this "exactly once" problem with a new I/O interface specification. This specification is the one seen by virtual processors, and so it is what operating system programmers and applications programmers must comply with when using the present invention.
According to the present invention, I/O operations are governed by the following rules (the rules are formulated for an interrupt-based I/O interface but have analogues for other I/O models):
Rule 0: To perform an I/O operation, a processor must first issue a
"connect" (which is a type of I/O operation) to the device that will perform that operation. The (virtual) processor can then issue I/O operations to that device. Rule 1: Delivery of a "completed" interrupt for an issued I/O operation implies that the operation was performed by the device. Rule 2: Concurrent I/O operations can be performed by the I/O device in any order. Rule 3: Delivery of a "disconnected" interrupt from an I/O device to a processor implies that outstanding I/O operations (i.e., those for which a "completed" interrupt has not been delivered) may or may not have been performed.
An important corollary to the above rules is that when a "disconnected" interrupt is received from an I/O device and one or more outstanding I/O operations were pending for that device (i.e., for which no "completed" interrupt has been received), then the program that receives the "disconnected" interrupt (A) must issue a new "connect" and (B) must reissue the outstanding I/O operations. In other words, use of the present invention requires that systems and applications programmers write their programs so as to respond to "disconnected" interrupts in this way.
According the above described scheme for delivering interrupts, "completed" and "disconnected" interrupts are delivered to a virtual processor only at the start of epochs. Also, these interrupts are delivered to both the primary and the backup. Consequently, when the backup VMM detects that the primary computer has failed (step 306), it proceeds as follows:
The VMM on the backup computer changes the setting of an internal "processor type" variable to indicate that it is now the primary processor (step 308), and it may also execute a number of other fail-over transition routines that are not relevant to the present invention (e.g., sending a hardware failure message to the system's operator console) (step 310). Then the backup VMM proceeds to take care of I/O operations in process during the fail-over.
First, the backup VMM deletes all I/O "completed" interrupts received during the epoch in which the failure occurred (step 312). The backup VMM knows the epoch in which the failure occurred because it receives an epoch message for every epoch that the primary VMM has completed. Thus if the last epoch message received from the primary VMM prior to failure was for epoch X, then the failure occurred in epoch X+1.
Second, the backup VMM adds to the set of buffered interrupts that will be delivered at the start of epoch X+1 , a "disconnected" interrupt for every I/O connection open during epoch X (step 314).
After steps 312 and 314, the backup virtual processor completes execution of epoch X (step 316), then the backup VMM delivers all buffered interrupts and traps to the backup virtual processor (step 318). Since the buffered interrupts include disconnects for all the I/O devices in use, the operating system or application software (whichever initiated connection to the I/O devices) running in the backup virtual processor will send new "connect" signals to all the I/O devices, and it will also reissue all outstanding I/O operations that were issued before epoch X+1 and for which a "completed" interrupt was not received (step 320).
The combination of these steps 312 and 314 causes fail-over from the primary computer to the backup computer to be made to look like an I/O failure insofar as the programs executing in the virtual processors 130 and 140 are concerned. Thus, programs that work correctly (i.e., in accordance with Rules 0-3) when messages to or from the I/O system are lost will also work correctly in a fault-tolerant computer system having primary and backup computers when a failure occurs during an I/O operation.
In summary, the potential problems associated with failures concurrent with outstanding I/O operations are solved by carefully not shielding the virtual processor from all failures. This allows processor failures to look like ordinary I/O device failures (such as lost I/O request messages and lost completed messages). A program that correctly copes with ordinary I/O failures and that complies with the above Rules 0-3, will work correctly even when I/O operations are outstanding during a fail-over.
Alternate Embodiments
While the motivation for the present invention was to make fault tolerance independent of the operating system being used, resulting in a preferred embodiment using virtual processors and virtual machine monitors, the techniques of the present invention could be implemented without using virtual processors and virtual machine monitors. In particular, one could modify the interrupt and trap handling routines of an operating system to (1) buffer interrupts and traps in software maintained structures and (2) deliver (i.e., initiating processing of) the delayed interrupts and traps at predefined points in the instruction stream of a computer. The delivery points would be defined by a recovery register or similar mechanism with additional delivery points occurring as defined by drain instructions. The handling of I/O operations disrupted by a fail-over would be the same as described above. This alternate embodiment of the present invention is suitable for use with RISC processors and satisfies the requirement of needing identical instruction streams in the primary and backup processors even though interrupts and traps may be delayed in each processor in an unpredictable fashion.
While the present invention has been described with reference to a few specific embodiments, the description is illustrative of the invention and is not to be construed as limiting the invention. Various modifications may occur to those skilled in the art without departing from the true spirit and scope of the invention as defined by the appended claims.

Claims

WHAT IS CLAIMED IS:
1. A fault tolerant computer system, comprising: a primary computer on which is running a primary virtual machine monitored by a primary virtual machine monitor; a backup computer on which is running a backup virtual machine monitored by a backup virtual machine monitor; said backup computer coupled to said primary computer for communication of messages therebetween; said primary and backup virtual machines executing substantially identical streams of instructions; a fault detector coupled to said primary computer and said backup computer for sending a fault message to said backup virtual machine monitor when said primary computer fails; said primary virtual machine monitor including delay means for trapping and buffering all interrupts and traps associated with operation of said primary virtual machine, for sending copies of said interrupts to said backup virtual machine monitor, and for delivering said buffered interrupts and traps to said primary virtual machine at predefined points in said stream of instructions executed by said primary virtual machine; and said backup virtual machine monitor including delay means for trapping and buffering all traps associated with operation of said backup virtual machine as well as said interrupts sent by said primary virtual machine monitor, and for delivering said buffered interrupts and traps to said backup virtual machine at predefined points in said stream of instructions executed by said backup virtual machine.
2. The fault tolerant computer system of claim 1 , wherein said primary computer and said backup computer both include a recovery register that stores a counter value that is automatically decremented during execution of a stream of instructions and that generates a recovery register interrupt signal when said counter value reaches a predefined terminal value; said primary and backup virtual machine monitors both including epoch means for stopping execution of said primary and backup virtual machines and for initializing said recovery register in said primary and backup computers to a preselected starting value whenever a recovery register interrupt signal is generated; said epoch means in said primary machine monitor including means for delivering said buffered interrupts and traps to said primary virtual machine whenever a recovery register interrupt signal is generated, for sending an epoch end notification message to said backup virtual machine monitor; and said epoch means in said backup virtual machine monitor including means for delivering said buffered interrupts and traps to said backup virtual machine after a recovery register interrupt signal is generated and an epoch end notification message is received from said primary virtual machine monitor.
3. The fault tolerant computer system of claim 2, wherein said primary and backup computers share access to at least one input/output device; said primary and backup virtual machines' execution of said streams of instructions is divided into a sequence of epochs, each epoch starting when said recovery register is initialized and ending when said recovery register interrupt signal is generated; said backup virtual machine including: backup input/output operation means for converting input/output commands to said at least one input/output device into null operation commands so long as said primary computer has not failed; software means for keeping track of all outstanding input/output operations not yet completed; and fail-over means, responsive to a fault message from said faul detector, for identifying which epoch in said sequence of epochs said primary computer failed during, deleting all buffered interrupts associated with said identified epoch, establishing a connection to each input/output device for which an input/output operation is outstanding, and reissuing all of said outstanding input/output operations.
4. The fault tolerant computer system of claim 1 , wherein said primary and backup computers have pipelined instruction decoders and hardware means for temporarily buffering traps caused by interrupts; said primary virtual machine monitor includes means for trapping user-mode drain instructions executed by said primary virtual machine monitor, performing a context switch, flushing said traps buffered by said hardware means of said primary computer, and then delivering both the traps that were buffered by said hardware means of said primary computer and any traps buffered by said primary virtual machine monitor to said primary virtual machine; and said backup virtual machine monitor includes means for trapping user-mode drain instructions executed by said backup virtual machine monitor, performing a context switch, flushing said traps buffered by said hardware means of said backup computer, and then delivering both the traps that were buffered by said hardware means of said backup computer and any traps buffered by said backup virtual machine monitor to said backup virtual machine.
5. A fault tolerant computer system, comprising: a primary computer; a backup computer coupled to said primary computer for communication of messages therebetween; said primary and backup computers executing substantially identical streams of instructions; a fault detector coupled to said primary computer and said backup computer for sending a fault message to said backup computer when said primary computer fails; said primary computer including delay means for trapping and buffering all interrupts and traps associated with operation of said primary computer, for sending copies of said interrupts to said backup computer, and for delivering said buffered interrupts and traps to said primary computer at predefined points in said stream of instructions executed by said primary computer; and said backup computer including delay means for trapping and buffering all traps associated with operation of said backup computer as well as said interrupts sent by said primary computer, and for delivering said buffered interrupts and traps to said backup computer at predefined points in said stream of instructions executed by said backup computer.
6. The fault tolerant computer system of claim 5, wherein said primary computer and said backup computer both include a recovery register that stores a counter value that is automatically decremented during execution of said streams of instructions and which generates a recovery register interrupt signal when said counter value reaches a predefined terminal value; said primary and backup computer both including epoch means for initializing said recovery register in said primary and backup computers to a preselected starting value whenever a recovery register interrupt signal is generated, said epoch means in said primary computer including means for said initiating processing of said buffered interrupts and traps by said primary computer whenever a recovery register interrupt signal is generated, for sending an epoch end notification message to said backup computer; and said epoch means in said backup computer including means for initiating processing of said buffered interrupts and traps by said backup computer after a recovery register interrupt signal is generated and an epoch end notification message is received from said primary computer.
7. The fault tolerant computer system of claim 6, wherein said primary and backup computers share access to at least one input/output device; said primary and backup computers' execution of said streams of instructions is divided into a sequence of epochs, each epoch starting when said recovery register is initialized and ending when said recovery register interrupt signal is generated; said backup computer including: backup input/output operation means for converting input/output commands to said at least one input/output device into null operation commands so long as said primary computer has not failed; software means for keeping track of all outstanding input/output operations not yet completed; and fail-over means, responsive to a fault message from said fault detector, for identifying which epoch in said sequence of epochs said primary computer failed during, deleting all buffered interrupts associated with said identified epoch, establishing a connection to each input/output device for which an input/output operation is outstanding, and reissuing all of said outstanding input/output operations.
8. The fault tolerant computer system of claim 5, wherein said primary and backup computers have pipelined instruction decoders and hardware means for temporarily buffering traps caused by synchronous interrupts; said primary computer includes means for trapping user-mode drain instructions, performing a context switch, flushing said traps buffered by said hardware means of said primary computer, and then delivering to said primary computer for processing both the traps that were buffered by said hardware means of said primary computer and any traps buffered by said primary computer; and said backup virtual machine monitor includes means for trapping user-mode drain instructions executed by said backup virtual machine monitor, performing a context switch, flushing said traps buffered by said hardware means of said backup computer, and then delivering both the traps that were buffered by said hardware means of said backup computer and any traps buffered by said backup virtual machine monitor to said backup virtual machine.
9. A fault tolerant data processing method, comprising the steps of: running a primary virtual machine on a primary computer and monitoring said primary virtual machine's operation with a primary virtual machine monitor; running a backup virtual machine on a backup computer and monitoring said back virtual machine's operation with a backup virtual machine monitor; executing substantially identical streams of instructions on said primary and backup virtual machines; sending a fault message to said backup virtual machine monitor when said primary computer fails; at said primary computer, trapping and buffering all interrupts and traps associated with operation of said primary virtual machine, sending copies of said interrupts to said backup virtual machine monitor, and delivering said buffered interrupts and traps to said primary virtual machine at predefined points in said stream of instructions executed by said primary virtual machine; and at said backup computer, trapping and buffering all traps associated with operation of said backup virtual machine as well as said interrupts sent by said primary virtual machine monitor, and delivering said buffered interrupts and traps to said backup virtual machine at predefined points in said stream of instructions executed by said backup virtual machine.
10. The fault tolerant data processing method of claim 9, wherein said primary computer and said backup computer both include a recovery register that stores a counter value that is automatically decremented during execution of a stream of instructions and which generates a recovery register interrupt signal when said counter value reaches a predefined terminal value; stopping execution of said primary and backup virtual machines and initializing said recovery register in said primary and backup computers to a preselected starting value whenever a recovery register interrupt signal is generated; whenever a recovery register interrupt signal is generated in said primary computer, delivering said buffered interrupts and traps to said primary virtual machine and sending an epoch end notification message to said backup virtual ■ machine monitor; and whenever a recovery register interrupt signal is generated in said backup computer and an epoch end notification message is received, delivering said buffered interrupts and traps to said backup virtual machine.
11. The fault tolerant data processing method of claim 10, wherein said primary and backup computers share access to at least one input/output device; dividing said primary and backup virtual machines' execution of said streams of instructions into a sequence of epochs, each epoch starting when said recovery register is initialized and ending when said recovery register interrupt signal is generated; said backup virtual machine: converting input/output commands to said at least one input/output device into null operation commands so long as said primary computer has not failed; keeping track of all outstanding input/output operations not yet completed; and responsive to said fault message, identifying which epoch in said sequence of epochs said primary computer failed during, deleting all buffered interrupts associated with said identified epoch, establishing a connection to each input/output device for which an input/output operation is outstanding, and reissuing all of said outstanding input/output operations.
12. The fault tolerant processing method of claim 11, wherein said primary and backup computers have pipelined instruction decoders and hardware means for temporarily buffering traps caused by synchronous interrupts; atsaid primary computer, trapping user-mode drain instructions executed by said primary virtual machine monitor, performing a context switch, flushing said traps buffered by said hardware means of said primary computer, and then delivering both the traps that were buffered by said hardware means of said primary computer and any traps buffered by said primary virtual machine monitor to said primary virtual machine; and at said backup computer, trapping user-mode drain instructions executed by said backup virtual machine monitor, performing a context switch, flushing said traps buffered by said hardware means of said backup computer, and then delivering both the traps that were buffered by said hardware means of said backup computer and any traps buffered by said backup virtual machine monitor o said backup virtual machine.
13. A fault tolerant data processing method, comprising the steps of: executing substantially identical streams of instructions in a primary computer and a backup computer; sending a fault message to said backup computer when said primary 5 computer fails; at said primary computer, trapping and buffering all interrupts and traps associated with operation of said primary computer, sending copies of said interrupts to said backup computer, and initiating processing of said buffered interrupts and traps by said primary computer at predefined points in said stream
10 of instructions executed by said primary computer; and at said backup computer, trapping and buffering all traps associated with operation of said backup computer as well as said interrupts sent by said primary computer, and initiating processing of said buffered interrupts and traps by said backup computer at predefined points in said stream of instructions executed
15 by said backup computer.
14. The fault tolerant data processing method of claim 13, wherein said primary computer and said backup computer both include a recovery register that stores a counter value that is automatically decremented during execution of said
20 streams of instructions and which generates a recovery register interrupt signal when said counter value reaches a predefined terminal value; at said primary computer, whenever a recovery register interrupt signal is generated, initializing said recovery register in said primary and backup computers to a preselected starting value, initiating processing of said buffered
25 interrupts and traps by said primary computer whenever a recovery register interrupt signal is generated, and sending an epoch end notification message to said backup computer; and κ> at said backup computer, after a recovery register interrupt signal is generated and an epoch end notification message is received from said primary computer, initiating processing of said buffered interrupts and traps by said backup computer.
15. The fault tolerant data processing method of claim 14, wherein said primary and backup computers share access to at least one input/output device; dividing said primary and backup computers' execution of said streams of instructions into a sequence of epochs, each epoch starting when said recovery register is initialized and ending when said recovery register interrupt signal is generated; said backup computer: converting input/output commands to said at least one input/output device into null operation commands so long as said primary computer has not failed; keeping track of all outstanding input/output operations not yet completed; and responsive to said fault message, identifying which epoch in said sequence of epochs said primary computer failed during, deleting all buffered interrupts associated with said identified epoch, establishing a connection to each input/output device for which an input/output operation is outstanding, and reissuing all of said outstanding input output operations.
16. The fault tolerant processing method of claim 15, wherein said primary and backup computers have pipelined instruction decoders and hardware means for temporarily buffering traps caused by synchronous interrupts; at said primary computer, trapping user-mode drain instructions executed by said primary computer, performing a context switch, flushing said traps buffered by said hardware means of said primary computer, and then initiating processing by said primary computer of both the traps that were buffered by said hardware means of said primary computer and any traps buffered by said primary computer; and at said backup computer, trapping user-mode drain instructions executed by said backup computer, performing a context switch, flushing said traps buffered by said hardware means of said backup computer, and then initiating processing by said backup computer of both the traps that were buffered by said hardware means of said backup computer and any traps buffered by said backup computer.
PCT/US1992/009238 1991-10-28 1992-10-27 Fault-tolerant computer processing using a shadow virtual processor WO1993009494A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US78351991A 1991-10-28 1991-10-28
US783,519 1991-10-28

Publications (1)

Publication Number Publication Date
WO1993009494A1 true WO1993009494A1 (en) 1993-05-13

Family

ID=25129530

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US1992/009238 WO1993009494A1 (en) 1991-10-28 1992-10-27 Fault-tolerant computer processing using a shadow virtual processor

Country Status (2)

Country Link
US (1) US5488716A (en)
WO (1) WO1993009494A1 (en)

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1998055922A1 (en) * 1997-06-04 1998-12-10 Marathon Technologies Corporation Loosely-coupled, synchronized execution
WO1999026133A2 (en) * 1997-11-14 1999-05-27 Marathon Technologies Corporation Method for maintaining the synchronized execution in fault resilient/fault tolerant computer systems
US6438707B1 (en) 1998-08-11 2002-08-20 Telefonaktiebolaget Lm Ericsson (Publ) Fault tolerant computer system
WO2004034172A2 (en) * 2002-09-12 2004-04-22 Siemens Aktiengesellschaft Method for synchronizing events, particularly for processors of fault-tolerant systems
EP1674992A1 (en) * 2004-12-22 2006-06-28 Microsoft Corporation Replicated virtual machine
US7155704B2 (en) 1998-06-30 2006-12-26 Sun Microsystems, Inc. Determinism in a multiprocessor computer system and monitor and processor therefor
EP1907935A2 (en) * 2005-06-24 2008-04-09 Syncsort Incorporated System and method for virtualizing backup images
FR3023398A1 (en) * 2014-07-01 2016-01-08 Bull Sas METHOD AND APPARATUS FOR SYNCHRONIZED EXECUTION OF AN APPLICATION IN A HIGH-AVAILABILITY ENVIRONMENT
US9378099B2 (en) 2005-06-24 2016-06-28 Catalogic Software, Inc. Instant data center recovery
US9753754B2 (en) 2004-12-22 2017-09-05 Microsoft Technology Licensing, Llc Enforcing deterministic execution of threads of guest operating systems running in a virtual machine hosted on a multiprocessor machine

Families Citing this family (94)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5978565A (en) 1993-07-20 1999-11-02 Vinca Corporation Method for rapid recovery from a network file server failure including method for operating co-standby servers
AU2663095A (en) * 1994-06-10 1996-01-05 Sequoia Systems, Inc. Main memory system and checkpointing protocol for fault-tolerant computer system
US5751992A (en) * 1994-09-23 1998-05-12 International Business Machines Corporation Computer program product for continuous destaging of changed data from a shared cache in a multisystem shared disk environment wherein castout interest is established in a hierarchical fashion
US5794242A (en) * 1995-02-07 1998-08-11 Digital Equipment Corporation Temporally and spatially organized database
JP3657665B2 (en) * 1995-02-14 2005-06-08 富士通株式会社 Multiple computer systems coupled to shared memory and control method of multiple computer systems coupled to shared memory
US5696895A (en) * 1995-05-19 1997-12-09 Compaq Computer Corporation Fault tolerant multiple network servers
US5675723A (en) * 1995-05-19 1997-10-07 Compaq Computer Corporation Multi-server fault tolerance using in-band signalling
JP3086779B2 (en) * 1995-06-19 2000-09-11 株式会社東芝 Memory state restoration device
US5737514A (en) * 1995-11-29 1998-04-07 Texas Micro, Inc. Remote checkpoint memory system and protocol for fault-tolerant computer system
US5745672A (en) * 1995-11-29 1998-04-28 Texas Micro, Inc. Main memory system and checkpointing protocol for a fault-tolerant computer system using a read buffer
US5864657A (en) * 1995-11-29 1999-01-26 Texas Micro, Inc. Main memory system and checkpointing protocol for fault-tolerant computer system
US6173332B1 (en) * 1996-03-06 2001-01-09 Paul L. Hickman Method and apparatus for computing over a wide area network
US5822534A (en) * 1996-06-04 1998-10-13 Sun Microsystems, Inc. Method and apparatus for selectively unmanaging elements on a network
US6032271A (en) * 1996-06-05 2000-02-29 Compaq Computer Corporation Method and apparatus for identifying faulty devices in a computer system
TW379298B (en) * 1996-09-30 2000-01-11 Toshiba Corp Memory updating history saving device and memory updating history saving method
US20050021477A1 (en) * 1997-01-29 2005-01-27 Ganapathy Krishnan Method and system for securely incorporating electronic information into an online purchasing application
US6038677A (en) * 1997-03-31 2000-03-14 International Business Machines Corporation Automatic resource group formation and maintenance in a high availability cluster configuration
US7389312B2 (en) 1997-04-28 2008-06-17 Emc Corporation Mirroring network data to establish virtual storage area network
US6799224B1 (en) 1998-03-10 2004-09-28 Quad Research High speed fault tolerant mass storage network information server
US6260155B1 (en) 1998-05-01 2001-07-10 Quad Research Network information server
US6496847B1 (en) * 1998-05-15 2002-12-17 Vmware, Inc. System and method for virtualizing computer systems
US6397242B1 (en) 1998-05-15 2002-05-28 Vmware, Inc. Virtualization system including a virtual machine monitor for a computer with a segmented architecture
US8631066B2 (en) * 1998-09-10 2014-01-14 Vmware, Inc. Mechanism for providing virtual machines for use by multiple users
US6247141B1 (en) 1998-09-24 2001-06-12 Telefonaktiebolaget Lm Ericsson (Publ) Protocol for providing replicated servers in a client-server system
US7516453B1 (en) * 1998-10-26 2009-04-07 Vmware, Inc. Binary translator with precise exception synchronization mechanism
JP3693226B2 (en) * 1999-06-30 2005-09-07 矢崎総業株式会社 Microcomputer backup device and automotive power window control device
GB2353113B (en) * 1999-08-11 2001-10-10 Sun Microsystems Inc Software fault tolerant computer system
US7703107B2 (en) * 2000-04-06 2010-04-20 Infineon Technologies Ag Virtual machine interface for hardware reconfigurable and software programmable processors
US8020176B2 (en) 2000-04-06 2011-09-13 Infineon Technologies Ag Virtual machine interface for hardware reconfigurable and software programmable processors
US7162615B1 (en) 2000-06-12 2007-01-09 Mips Technologies, Inc. Data transfer bus communication using single request to perform command and return data to destination indicated in context to allow thread context switch
US7263476B1 (en) * 2000-06-12 2007-08-28 Quad Research High speed information processing and mass storage system and method, particularly for information and application servers
US7020709B1 (en) * 2000-06-30 2006-03-28 Intel Corporation System and method for fault tolerant stream splitting
GB0112781D0 (en) 2001-05-25 2001-07-18 Global Continuity Plc Method for rapid recovery from a network file server failure
US7620955B1 (en) 2001-06-08 2009-11-17 Vmware, Inc. High-performance virtual machine networking
US7730153B1 (en) * 2001-12-04 2010-06-01 Netapp, Inc. Efficient use of NVRAM during takeover in a node cluster
US7360034B1 (en) * 2001-12-28 2008-04-15 Network Appliance, Inc. Architecture for creating and maintaining virtual filers on a filer
KR100441712B1 (en) * 2001-12-29 2004-07-27 엘지전자 주식회사 Extensible Multi-processing System and Method of Replicating Memory thereof
US7308576B2 (en) * 2001-12-31 2007-12-11 Intel Corporation Authenticated code module
US7631196B2 (en) * 2002-02-25 2009-12-08 Intel Corporation Method and apparatus for loading a trustable operating system
US6961870B2 (en) * 2002-03-13 2005-11-01 Inventec Corporation Data exchange update and back-up system and method between dual operating systems of a computer
US20030229794A1 (en) * 2002-06-07 2003-12-11 Sutton James A. System and method for protection against untrusted system management code by redirecting a system management interrupt and creating a virtual machine container
JP2004102324A (en) * 2002-09-04 2004-04-02 Oki Electric Ind Co Ltd Interrupt program module, recording medium storing the module and computer capable of interrupt processing
WO2004034199A2 (en) 2002-10-04 2004-04-22 Starent Networks Corporation Managing resources for ip networking
US7028218B2 (en) * 2002-12-02 2006-04-11 Emc Corporation Redundant multi-processor and logical processor configuration for a file server
DE10313318A1 (en) * 2003-03-25 2004-10-21 Giesecke & Devrient Gmbh Controlled execution of a program intended for a virtual machine on a portable data carrier
US7415708B2 (en) * 2003-06-26 2008-08-19 Intel Corporation Virtual machine management using processor state information
US20050039074A1 (en) * 2003-07-09 2005-02-17 Tremblay Glenn A. Fault resilient/fault tolerant computing
US7146530B2 (en) * 2003-07-18 2006-12-05 Hewlett-Packard Development Company, L.P. Targeted fault tolerance by special CPU instructions
US9020801B2 (en) * 2003-08-11 2015-04-28 Scalemp Inc. Cluster-based operating system-agnostic virtual computing system
US7424709B2 (en) * 2003-09-15 2008-09-09 Intel Corporation Use of multiple virtual machine monitors to handle privileged events
US7366305B2 (en) 2003-09-30 2008-04-29 Intel Corporation Platform and method for establishing trust without revealing identity
US8037314B2 (en) 2003-12-22 2011-10-11 Intel Corporation Replacing blinded authentication authority
US20050204186A1 (en) * 2004-03-09 2005-09-15 Rothman Michael A. System and method to implement a rollback mechanism for a data storage unit
US8230085B2 (en) * 2004-04-12 2012-07-24 Netapp, Inc. System and method for supporting block-based protocols on a virtual storage appliance executing within a physical storage appliance
US7818615B2 (en) * 2004-09-16 2010-10-19 Invensys Systems, Inc. Runtime failure management of redundantly deployed hosts of a supervisory process control data acquisition facility
US7676558B2 (en) * 2004-11-12 2010-03-09 International Business Machines Corporation Configuring shared devices over a fabric
US7487403B2 (en) * 2004-11-12 2009-02-03 International Business Machines Corporation Method for handling a device failure
US7480822B1 (en) * 2005-07-13 2009-01-20 Symantec Corporation Recovery and operation of captured running states from multiple computing systems on a single computing system
US20070074067A1 (en) * 2005-09-29 2007-03-29 Rothman Michael A Maintaining memory reliability
US8010829B1 (en) 2005-10-20 2011-08-30 American Megatrends, Inc. Distributed hot-spare storage in a storage cluster
US8010485B1 (en) 2005-10-20 2011-08-30 American Megatrends, Inc. Background movement of data between nodes in a storage cluster
US7996608B1 (en) 2005-10-20 2011-08-09 American Megatrends, Inc. Providing redundancy in a storage system
US20090106754A1 (en) * 2005-12-10 2009-04-23 Benjamin Liu Handling a device related operation in a virtualization enviroment
US20070162594A1 (en) * 2006-01-12 2007-07-12 Microsoft Corporation Controlled disconnection of a network device
US7814495B1 (en) * 2006-03-31 2010-10-12 V Mware, Inc. On-line replacement and changing of virtualization software
US7533207B2 (en) * 2006-12-06 2009-05-12 Microsoft Corporation Optimized interrupt delivery in a virtualized environment
US8498967B1 (en) * 2007-01-30 2013-07-30 American Megatrends, Inc. Two-node high availability cluster storage solution using an intelligent initiator to avoid split brain syndrome
US7908448B1 (en) 2007-01-30 2011-03-15 American Megatrends, Inc. Maintaining data consistency in mirrored cluster storage systems with write-back cache
US8046548B1 (en) 2007-01-30 2011-10-25 American Megatrends, Inc. Maintaining data consistency in mirrored cluster storage systems using bitmap write-intent logging
US20080228971A1 (en) * 2007-03-13 2008-09-18 Rothman Michael A Device modeling in a multi-core environment
US8099274B2 (en) * 2007-03-30 2012-01-17 International Business Machines Corporation Facilitating input/output processing of one or more guest processing systems
US7805631B2 (en) * 2007-05-03 2010-09-28 Microsoft Corporation Bare metal recovery from backup media to virtual machine
US7877347B2 (en) * 2007-05-03 2011-01-25 Payton David W Method and system for independently observing and modifying the activity of an actor processor
US7840839B2 (en) * 2007-11-06 2010-11-23 Vmware, Inc. Storage handling for fault tolerance in virtual machines
US7949850B2 (en) * 2007-12-28 2011-05-24 Intel Corporation Methods and appratus for demand-based memory mirroring
US7925923B1 (en) * 2008-01-31 2011-04-12 Hewlett-Packard Development Company, L.P. Migrating a virtual machine in response to failure of an instruction to execute
US20090210888A1 (en) * 2008-02-14 2009-08-20 Microsoft Corporation Software isolated device driver architecture
JP4530182B2 (en) * 2008-02-27 2010-08-25 日本電気株式会社 Processor, memory device, processing device, and instruction processing method
US9027003B2 (en) * 2008-03-31 2015-05-05 Vmware, Inc. Precise branch counting in virtualization systems
CN101593136B (en) * 2008-05-30 2012-05-02 国际商业机器公司 Method for obtaining high availability by using computers and computer system
JP5268469B2 (en) * 2008-07-23 2013-08-21 株式会社東芝 High availability system and execution state control method
US8499297B2 (en) 2008-10-28 2013-07-30 Vmware, Inc. Low overhead fault tolerance through hybrid checkpointing and replay
JP5352299B2 (en) * 2009-03-19 2013-11-27 株式会社日立製作所 High reliability computer system and configuration method thereof
US8201169B2 (en) 2009-06-15 2012-06-12 Vmware, Inc. Virtual machine fault tolerance
US8090977B2 (en) 2009-12-21 2012-01-03 Intel Corporation Performing redundant memory hopping
US11320797B2 (en) * 2010-09-27 2022-05-03 Fisher-Rosemount Systems, Inc Methods and apparatus to virtualize a process control system
US8984318B2 (en) * 2011-01-03 2015-03-17 Ca, Inc. System and method to avoid resynchronization when protecting multiple servers
CN104503861A (en) * 2012-04-20 2015-04-08 华为技术有限公司 Abnormality handling method and system, agency device and control device
US9720712B2 (en) 2013-06-03 2017-08-01 Red Hat Israel, Ltd. Physical/virtual device failover with a shared backend
CN103581177A (en) * 2013-10-24 2014-02-12 华为技术有限公司 Method and device for managing virtual machine
EP3129903B1 (en) 2014-07-01 2018-11-28 SAS Institute Inc. Systems and methods for fault tolerant communications
US9946718B2 (en) 2015-07-27 2018-04-17 Sas Institute Inc. Distributed data set encryption and decryption
US9811524B2 (en) 2015-07-27 2017-11-07 Sas Institute Inc. Distributed data set storage and retrieval
US10474819B2 (en) 2015-11-20 2019-11-12 Lastline, Inc. Methods and systems for maintaining a sandbox for use in malware detection

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US3810119A (en) * 1971-05-04 1974-05-07 Us Navy Processor synchronization scheme
EP0398694A2 (en) * 1989-05-17 1990-11-22 International Business Machines Corporation Fault tolerant data processing system

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4228496A (en) * 1976-09-07 1980-10-14 Tandem Computers Incorporated Multiprocessor system
US4456952A (en) * 1977-03-17 1984-06-26 Honeywell Information Systems Inc. Data processing system having redundant control processors for fault detection
US4521871A (en) * 1982-04-12 1985-06-04 Allen-Bradley Company Programmable controller with back-up capability
US4787041A (en) * 1985-08-01 1988-11-22 Honeywell Data control system for digital automatic flight control system channel with plural dissimilar data processing
JP2580113B2 (en) * 1985-11-15 1997-02-12 富士通株式会社 Virtual computer system
US4965717A (en) * 1988-12-09 1990-10-23 Tandem Computers Incorporated Multiple processor system having shared memory with private-write capability
US5295258A (en) * 1989-12-22 1994-03-15 Tandem Computers Incorporated Fault-tolerant computer system with online recovery and reintegration of redundant components
DE69021712T2 (en) * 1990-02-08 1996-04-18 Ibm Restart marking mechanism for fault tolerant systems.
US5155729A (en) * 1990-05-02 1992-10-13 Rolm Systems Fault recovery in systems utilizing redundant processor arrangements
US5157663A (en) * 1990-09-24 1992-10-20 Novell, Inc. Fault tolerant computer system
US5363503A (en) * 1992-01-22 1994-11-08 Unisys Corporation Fault tolerant computer system with provision for handling external events

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US3810119A (en) * 1971-05-04 1974-05-07 Us Navy Processor synchronization scheme
EP0398694A2 (en) * 1989-05-17 1990-11-22 International Business Machines Corporation Fault tolerant data processing system

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
PATENT ABSTRACTS OF JAPAN vol. 011, no. 334 (P-631)31 October 1987 *

Cited By (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5896523A (en) * 1997-06-04 1999-04-20 Marathon Technologies Corporation Loosely-coupled, synchronized execution
AU733747B2 (en) * 1997-06-04 2001-05-24 Marathon Technologies Corporation Loosely-coupled, synchronized execution
WO1998055922A1 (en) * 1997-06-04 1998-12-10 Marathon Technologies Corporation Loosely-coupled, synchronized execution
WO1999026133A2 (en) * 1997-11-14 1999-05-27 Marathon Technologies Corporation Method for maintaining the synchronized execution in fault resilient/fault tolerant computer systems
WO1999026133A3 (en) * 1997-11-14 1999-07-15 Marathon Techn Corp Method for maintaining the synchronized execution in fault resilient/fault tolerant computer systems
US6279119B1 (en) 1997-11-14 2001-08-21 Marathon Technologies Corporation Fault resilient/fault tolerant computing
US6473869B2 (en) 1997-11-14 2002-10-29 Marathon Technologies Corporation Fault resilient/fault tolerant computing
US7155704B2 (en) 1998-06-30 2006-12-26 Sun Microsystems, Inc. Determinism in a multiprocessor computer system and monitor and processor therefor
US6438707B1 (en) 1998-08-11 2002-08-20 Telefonaktiebolaget Lm Ericsson (Publ) Fault tolerant computer system
WO2004034172A3 (en) * 2002-09-12 2004-09-23 Siemens Ag Method for synchronizing events, particularly for processors of fault-tolerant systems
WO2004034172A2 (en) * 2002-09-12 2004-04-22 Siemens Aktiengesellschaft Method for synchronizing events, particularly for processors of fault-tolerant systems
EP1674992A1 (en) * 2004-12-22 2006-06-28 Microsoft Corporation Replicated virtual machine
US9753754B2 (en) 2004-12-22 2017-09-05 Microsoft Technology Licensing, Llc Enforcing deterministic execution of threads of guest operating systems running in a virtual machine hosted on a multiprocessor machine
EP1907935A2 (en) * 2005-06-24 2008-04-09 Syncsort Incorporated System and method for virtualizing backup images
AU2006262046B2 (en) * 2005-06-24 2011-05-12 Catalogic Software, Inc. System and method for virtualizing backup images
US8346726B2 (en) 2005-06-24 2013-01-01 Peter Chi-Hsiung Liu System and method for virtualizing backup images
CN101243413B (en) * 2005-06-24 2013-08-14 信科索尔特公司 System and method for virtualizing backup images
US9329941B2 (en) 2005-06-24 2016-05-03 Catalogic Software, Inc. System and method for virtualizing backup images
US9378099B2 (en) 2005-06-24 2016-06-28 Catalogic Software, Inc. Instant data center recovery
EP1907935A4 (en) * 2005-06-24 2009-04-22 Syncsort Inc System and method for virtualizing backup images
US9983951B2 (en) 2005-06-24 2018-05-29 Catalogic Software, Inc. Instant data center recovery
US10877852B2 (en) 2005-06-24 2020-12-29 Catalogic Software, Inc. Instant data center recovery
FR3023398A1 (en) * 2014-07-01 2016-01-08 Bull Sas METHOD AND APPARATUS FOR SYNCHRONIZED EXECUTION OF AN APPLICATION IN A HIGH-AVAILABILITY ENVIRONMENT
EP2975517A3 (en) * 2014-07-01 2016-04-13 Bull Sas Method and device for synchronised execution of an application in a high-availability environment
US9575850B2 (en) 2014-07-01 2017-02-21 Bull Sas Method and device for synchronously running an application in a high availability environment

Also Published As

Publication number Publication date
US5488716A (en) 1996-01-30

Similar Documents

Publication Publication Date Title
US5488716A (en) Fault tolerant computer system with shadow virtual processor
US6401216B1 (en) System of performing checkpoint/restart of a parallel program
US5790397A (en) Fault resilient/fault tolerant computing
US6393583B1 (en) Method of performing checkpoint/restart of a parallel program
KR0137406B1 (en) Fault tolerant computer system
US6308318B2 (en) Method and apparatus for handling asynchronous exceptions in a dynamic translation system
US7111086B1 (en) High-speed packet transfer in computer systems with multiple interfaces
US6338147B1 (en) Program products for performing checkpoint/restart of a parallel program
US6161198A (en) System for providing transaction indivisibility in a transaction processing system upon recovery from a host processor failure by monitoring source message sequencing
US5664195A (en) Method and apparatus for dynamic installation of a driver on a computer system
JP2587141B2 (en) Mechanism for communicating messages between multiple processors coupled via shared intelligence memory
US5621885A (en) System and method for providing a fault tolerant computer program runtime support environment
US5968185A (en) Transparent fault tolerant computer system
JP5258019B2 (en) A predictive method for managing, logging, or replaying non-deterministic operations within the scope of application process execution
US8281304B2 (en) High availability system and execution state control method
US8510592B1 (en) PCI error resilience
US7870296B2 (en) High availability system and execution state control method
US6622259B1 (en) Non-disruptive migration of coordinator services in a distributed computer system
EP0441087A1 (en) Checkpointing mechanism for fault-tolerant systems
US7581220B1 (en) System and method for modifying user memory from an arbitrary kernel state
US10013367B2 (en) I/O processing system including dynamic missing interrupt and input/output detection
US9672165B1 (en) Data management tier coupling primary storage and secondary storage
GB2504772A (en) Coprocessor providing service address space for diagnostics collection
US6035419A (en) Logical execution sequence error recovery method and system
US6332199B1 (en) Restoring checkpointed processes including adjusting environment variables of the processes

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A1

Designated state(s): JP KR

AL Designated countries for regional patents

Kind code of ref document: A1

Designated state(s): AT BE CH DE DK ES FR GB GR IE IT LU MC NL SE

DFPE Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101)
122 Ep: pct application non-entry in european phase
122 Ep: pct application non-entry in european phase