CA1245327A - Path oriented routing system and method for packet switching networks - Google Patents

Abstract

PATH ORIENTED ROUTING SYSTEM AND METHOD FOR PACKET SWITCHING NETWORKS
Abstract of the Disclosure A path oriented routing system and method for packet switching networks with end-to-end internal protocols. It allows switch pairs to communicate over multiple paths without packet disordering. A distributed loop-free shortest path algorithm assigns a number to a path at the time it is created and this number remains valid through path changes. Consequently, existing traffic can be maintained on existing paths, while new traffic is assigned to the current (i.e. new) shortest paths.

- i -

Description

3L,2~53;Z~

A distributed loop-free shortest path algori-thm assigns a number (path identifier or path ID) to a path at the time it is created and this number (path ID) remains valid during shortest path changes. Consequently, existing traffic can be maintained on existing paths, while new traffic is assigned to the new shortest paths. Packet disordering is thus minimized since the packets of a given connection all follow the same path. Stable multiple path routing is thus achieved.
Briefly stated, the invention is implemented as follows. Each switch contains four tables of information: a routing table; a shortest path table; a trunk table; and a metric table.
These tables will be described later, in more detail. When a user connection is established (between two end-point switches), each end-point switch obtains the path identifier (path ID) of the then current shortest path (time) to the other endpoint. Every packet sent over the connection is tagged with an appropriate path identifier instead of the more customary identifier of the destination switch.
Because these path identifiers remain valid even after the creation of new shortest paths, the same paths can be used by a connection throughout its lifetime (barring any failures).
The routing tables at each switch are indexed by path identifiers (path ID) instead of the more customary identifier of the destination switch. In addition to a trunk identifier (trunk I.D.), each entry in the routing table contains the new path ID to be assigned to the packet. Indeed, unlike a destination identifier, a packet's path identifier ID must typically be changed at every relay

2 ,~

. ~

, :

53Z~

(or tandem) switch.
Stated in other terms, the present invention is a path-oriented routing method for packet switching networks, wherein the network is comprised of a plurality of interconnected packet switches, the method characterized by: each destination packet switch broadcasting to all its adjacent neighbours the preferred path identifier to use to send messages to it; and each successive packet switch, moving monotonicaly in a direction away from the destination packet switch, broadcasting to all its adjacent neighbours, the preferred path identifier to use to send messages to it destined ultimately for the destination packet switch.
Stated in yet o-ther terms, the present invention is a path-oriented routing system for packet switching networks7 wherein the network is comprised of a plurality of interconnected packet switches, the system characterized by: a path identifier being associated with each packet at the source packet switch and carried by the packet as it traverses through the network to the destination switch, the path identifier being updated at each packet switch traversed.
Brief_Description of the Drawings The invention will now be described in greater detail with reference to the accompanying drawings, wherein like parts in each of the several figures are identified by the same reference character, and wherein:
Figure 1 is a simplified block diagram of a packet switching network, depicting some possible physical trunks between 4S3~

the switches;
Figure 2 is similar to Figure 1 but depicts the logical interconnections between the switches after an initial path assignment to destination H;
Figure 3 is similar to Figure 2 but depicts the logical interconnections between the switches after one path update;
Figure 4 is similar to Figure 3 but depicts the logical interconnections between the switches after two path updates;
Figure 5, containing parts a to d, is a stylistic representation of the information tables stored at a switch in the switching network;
Figures 6, 7, and 8 are routing tables for switch C in the conditions of Figures 2, 3 and 4 respectively; and Figure 9 is a shortest path Table for switch A in the condition of Figure 4.
Figure 1 depicts a packet switching network 10 comprised of packet switches A to H, inclusive, and showing all the potential interconnections between the various switches A to ~1.
Before the theory of operation is described, it may be of value to first consider, in more detail, the philosophy behind what is happening. To do this, refer to Figure 2.
Figure 2 depicts packet switching network 10 from the perspective of the switch which is the destination switch (which in Figure 2 is of course switch H). In other words, when drawn from the perspective of switch H being the ultimate destination of all the other switches A to G, all paths lead to switch H.

.

12~5i327 Note also that each switch A to H carries a path ID
indicated as 0, 1, or 2. In Figure 2, each switch A to H has its path identified as path 0. This means that if switch A wishes to send a packet to switch H, it sends it on path 0 (to switch C), switch C
sends it on path 0 (to switch E), etc., until it gets to switch H. At the destination switch H, the packet's path ID is mapped to a special end-of-path indicator and further routing information in the packet is used to locate the connection's endpoint.
Now compare Figure 2 to Figure 3. Traffic between switch C and switch E has increased sufficiently that the shortest path from switch C to switch H is now via switch D. This information is cdlculated at switch C and is broadcast to all switches that are directly connnected to switch C (i.e. to switches A, B, D, and E in this example). In effect, switch C is saying to all switches that are connected to it, don't make any new connections to me via path 0, but rather, use path 1. Note, however, that existing connections still employ path 0 (as is shown in Figure 3) and only new connections will use path 1.
New information from switch A (destined for switch H) leaves switch A on path 1 (this rnay be on the same trunk as is path 0). When the packet ge-ts to switch C it is received on path 1, its path ID is changed to 0, and then the packet is sent to switch D.
From switch D the packet follows path 0 through switch G to switch H.
Figure 4 depicts switching network 10 after a second path assignment update. Assume that the traffic between switches E
and F has decreased enough so that this connection is now chosen as ~2~53Z7 part of the shortest path. Switch E calculates this fact and broadcasts it to all switches directly connected to switch E (i.e. to switches C, D, F, and G). This is reflected in the fact that swikch E
needs to receive on a new path (i.e. path 1) to distinguish between packets destined for switch G (path 0) and those destined For switch F
(path 1). Assume that switches A and C also find that going to switch H via switch E is now shorter, switches A and C operate on path 2 to keep those packets separate from packets on paths 0 and 1.
Note that each time a packet goes through a switch, its path ID gets updated. It may get updated to the same path ID or to a new path ID. Using path ID 2 leaving switch A as an example, it remains path ID 2 entering switch C, in switch C it is changed to path ID 1 and leaves switch C as that. It enters switch E as path ID 1 and in switch E it is changed to path ID 0, which it maintains through switches F, G and H.
Figures 5a, 5b, 5c, and 5d (referred to collectively as Figure 5) depict tables of information contained at each switch A to H. Figure 5a depicts the routing table for a switch. This table relates an incoming path to an outgoing path when the switch in question is a tandem switch (i.e. not the destination switch). As can be seen from Figure 5a it is a look-up table (stor~d in memory at the switch) which is indexed by the incoming path ID.
Figure 5b depicts the shortest path Table For a switch.
As can be seen from the Figure, it is a table indexed by the identifier of the Destination switch (i.e. Destination switch ID). If the swi-tch in question wishes to establish a user connection to switch lL2~153~7 D (for example) it looks up under switch D for the Outgoing path ID
and Trunk ID (identifier) to use.
Figure 5c depicts the trunk Table for a switch. As can be seen from the Figure, this relates destination switch ID, trunk ID, path ID, and distance to destination. Note that the shortest path table (Figure 5b) is a condensed version of this Trunk Table. That is, the Trunk Table contains information on all the possible shortest paths to a destination, while the shortest Path Table (Figure 5b) contains only the current shortest paths.
Figure 5d is a Metric Table depicting the distance of each crunk to the next switch. Note that the use of the term "distance" in the present context does not mean physical distance but rather is a term well known in this art that is a function of the time delay incurred by a packet moving from one switch to another. A
normalized value of distance is typically used (ranging, for example, from O to 2,0~8 for a single trunk). This concept is also sometimes referred to as "cost" or "delay".
The use of the Tables in Figure 5 and their interrelationship will now be described.
Let's assume that the network of Figure 2 is being established. The destination switch (i.e. switch H) broadcasts to its neighbours (only switch G in this example) a message that says: "if you want to send a data message to me, sent it on a path having path ID - O". This is then stored in the Trunk Table at switch G and it is also stored in the Shortest Path Table and the Routing Table at switch G if indeed this path is shortest for switch G.

:~LZ4532~

Swi-tch G then broadcasts a message to its neighbours (i.e. switches D, E, and F) that says, in effect: "If you want to send a message to switch H via me, send it on a path having Path ID =
0". Switch G can then use its Routing Table to transfer any message it receives coming in on a path with path ID = 0 to being sent on a path with path ID = 0. It can do this without having any further information about the message. Note that the path ID may or may not change as it passes through a switch. Specific examples of Routing Tables for Switch C can be found in Figures 6, 7, and 8. A specific example of a Shortest Path Table for Switch C (of Figure 4) is depicted in Figure 9.
The remainder of the switches in Figure 2 are initialized in an analagous manner. Note that the information contained in the Routing Table is used when the switch is acting as a Tandem switch for packets; under these circumstances, the shortest Path Table (at that Tandem switch) is not needed. The Shortest Pa-th Table is used only when that switch is originating a user connection for the destination switch (or when receiving updated routing information).
Suppose the situation changes, and switch E becomes too busy to handle more traffic from switch C bound for switch H (see Figure 3). In effect, switch E has sent a message to switch C saying:
"My distance to switch H on path ID = 0 has increased". Note however, that switch E continues to handle existing traffic on path ID
= 0; it is only new traffic~that switch E will not handle.
Switch C now chooses a new shortest path by examining `, `

~ILZ'~S3Z7 its Trunk Table and chooses a path via switch D and alters its Shortest Path Table and its Routing Table accordingly.
However, switch C has to distinguish the traffic it receives from switch A into previously existing traffic (still to be sent to switch E on path ID = 0) and new traffic to be sent via switch D. It does this by sending a message to switch A to send new messages, destined for switch H, with path ID = 1. Switch C then receives a message with path ID = 1, converts the pa-th ID to 0 and sends it to switch D.
After a second update the situation is as shown in Figure 4, and the Shortest Path Table for Switch A representing such a situation is shown in Figure 9.

APPENDIX
Included as appendices to this document are the following:
A) Paper to be presented by the inventors in September 1985.
B) High level programme.

APPENDIX A
~L2'~327 A P~Tfl-~RIENTED ROUTING STRATEGY
FOR PACRET SWITCHING NETWORKS
WITH EN~TO-END PROTOCOLS
R AUBIN - P NG
BNR P O Box 3511 St~tion C
Otta-ra Ont~rio KlY 4H7 Canada A,STR~CT tic~tive of the improvement possible one withmixed trunk speeds and the other with sparse traf-fic matrices and hi8h network connectivity Opti-A path-oriented routing strategy is proposed for mal multiple path routing aehievod respectively paeket switching network~ lith 2nd-to~end internal 122 percent and 186 percent better network protocols It ~llows switch pairs to eommunicate throughput than single path routing (under a cer-over multiplo p~tho (or better network through- tain delay con~traint) tlixed trunk speeds put) while maint~ining knowledge of u~er con- l~parse traffie matrices and high connectivity are nections at the network's endpoints only The mo~t precisely typical of the trends seen in packet significant aspeet of this strategy lies in its switching networhs flow assignment method A distributed loop-froe shortest path ~Igoritha~ assign~ a number to a path Thi~ paper propo~e~ a path-oriented routing method st the time it i~ croated and this number rem~ins for packet ~witching natworks with end-to-end valid throu8h ~hortest p~th changeo Con~oquontly internal protocols. Ti is method permits the es-exi~tng tr~ffic can be ~aintain~d on oxisting tabliahment of multiple path~ betwoen any two path~ while now trafeie ir as~igned to the eur- ~witches in a notwork in a fashion that allows the rent ohorte~t p~ths St~ble multiple path routing p~ckets of the s~me connection to follow the saDe is thus achievet without paeket di~ordering. Ab- path; Packet disordoring i3 thus mininized and normal condit}on~ such ;1~ trunk failuro and r~cov- tr~cing of user connoetionr is possible. So-o ory and trunk congertion aro doalt with by ta2ging arlier ~tratogio~ such as i~olated load-routing updateJ ~ith update cau~oi. Simulation of ~plitting' wore not path-oriented They had tho this routing ~trategy shows that Isaximu~ notwork ti~advantige of reaeting DLowly to ti~tant ovont~
throughput (undor a cart~in cong~tion eonstraint) -and increa~ing paeket disordoring. t10ro rocently can be incr~asod ub~tanti~lly eo~pared to ~ ~in- a path-orionted method waa d~veloped or the glr path routing ~trategy datagram-baoed SITA notwork' Novortholo~s tho ~tatic appro~ch to routing takon by thi~ ~tr;togy applieo only to trafic that is oaoily prodict-lAlTPfOOVCTlOR
Tho fundamontal property of ond-to-end arehitoe-ture~ ean be prooervod ovon if routing i~ path-Tod~y'~ packet ~wltching notuork~ are ehsract~r- oriented Namely Jtste inforr~ation about user ized by an ineron~ing diver-ity and utilization of eonnection~ csn still bo kept only ~t tho net-~witching ~nd tr~n~ ion ~quipmont haking op- worl;'s ontry 3nd exit l~ntitios that porfor~ the tir~l u~e of ~ueh divor~o and burdon~d re-ouree~ intornal er~bedding X 25-liko protoeol User eon-is ~ ch-llonge ;It time whon the ~orvieo ~aturo~ nection~ can be Iwitehod from one path to another ~nd networS op~rator~ boeome moro eost eonoeiou~ uithout oxplieit roeonnoct procoduro~; tho doci-s ion to do ~o resides ~trietly with the eon-Routing strategies have ~ ~ajor role to play in neetions' endpoint~ and i~ basod on ognifie~nt optimizing the uJo of network resoureos Yet eh~nging network eonditions Furthermoro tran~-eurront routing ~othods have boen eritiei7ed for aetion~ involvo no more than the network ' s entry failing to ~oretimos even approach opti~ality' and exit point~
In particular ~witch pairs often eom~unie~te ovor ingle patho when optimal routing would typieally Path-oriented routing i~ prosented in the eontext require multiple path~ in ordor to oven out traf- of a eorlpreh~n~ive experimental routing sy~tem fie over tho network. Thi~ offoet was me~sur~d by called THESEUS The p~per st~rt~ by giving the Seligman" Two oxperimont~ wore particularly in- general ~odel underlying THESEUS Tho methods Permission ~o copy wi~hout fee all ùr parl of Ihis material is ~ran~ed put~tion are then outlined anThirhrt't path eo~
provldcdthallhccopicsarcnotmadcordistributcdrordircct morG olaborato explanation of multiple path flow commcrcial advantagc thc ACM copyright noticc and the titlc of thc a~si~nr~ent Finally the reoults of a ailoulation publication and ils date appcar and notice is given that copying is by aro pre~ontod pcrmission of the Association for CoAnputing Machincry To copy otherwise o~ to republish requircs a re- and/or Spccirlc pcrmissioA

~ / O

, Z4~327 ROUTI~G h~70f~ ~Th7~C ~VA~UAT10~
A routing scheme is normally characterized by its THESEUS adopts a tractable and approximate netric performance criterion its path computation formula that takes into acccunt processing trons-method and its packet forwarding strategy. Hore mir~ion and propagation delay as well as any precisely given a performance criterion one can necessary uaiting delay. It is proportional to identify a routing metric that i~ a function m8rg;nal de13) which is the first derivative of that measures the lengths of network trunks with average network delay with respect to trunk respect to performance optimization. This metric throughput; this is a neces~ary condition Eor op-can then be used to compute optimal paths. timality. If propagation delay is ignorod then Finally the knowledge of optimal paths allows one network throughput is maxir~ized otherwise net-to assign traffic flow. These three steps are il- work delay is minimized'. Choosing ~ metric is lustrated in Figure 1. not 3ufficient, however; one ~lso has to decide how freguently to evaluate it and generate routing A dimension not visible il~ Figure 1 is systom re- updates.
sponsiveness. It is cu:tomary to identify three increa~ing levol~ of routing responcivene~s: THESEUS proposes that every switch of a network static, quasi-~tatic and dynamic. Mo~t ty~tems evsluato the parameters o tho metric (i.e.
react dynamically to the failure of a network processor and line utilizaton pacxet size and trunk. However re~cting to change~ in traffic i~ packet throughput) periodically for each trunk. A
more controversi~l. With static routing the new ~etric value results from t~king the exponen-feedback loop is performed off-line while quasi- tial average of the current value and the instan-static routing control~ traffic on-line. Both try tancous measured value. However routing updates to track ctatistically significant trends but at are not necossarily gener~t~d every timo the met-a quite different pace. Dynamic routing is a ric is ovaluat~d. THESEUS favors uperiodic ~ener-limit caso suitable to trafic with no prodictable ation of routing updates at the risk of patterns; it trie~ to adapt to the instantaneous convergine more slouly and less accurately. 7n network state. practice control traffic due to frcquent periodic updating has been observed to increase network The method of computing routing tables is yet sn- congestion7 this phenomenon i8 not factored into other dimension of the routing modDI. Computation poriodic routing optimization ~othods. THESEUS
c~n be centralizod or distributet. This question ~endc routing upd~tes only when cert~in thresholds csn rc~listic~lly be po~ed Sor quaai-~tatic rout- sro crossod. These threshold~ depond non lin~rly on trunk congestion as captùrod by trunk utiliza--~ ~ ~ -~-~~~~~ - ~-- tion. Typical thresholds rd~y be 6û 75 ~nd h2 traffic porcont or an increase in utilization and 43,NET~ORK 69 ~nd 79 percent Sor a decre~se. Hysteresis is applied for stability.
~ith this approach THESEUS is increasingly aday-control observ~tion tive ~s load goer up. At very high load howev~r when trunk utilization exceeds a certain Is~xirdum threJhold congestion control tako~ prtcedence r~ - - _ over routing as will be oxpLainod later.
flow P~th Qetric assignment .~_ computation .~_ ovaluation SHOPT~SJ PATN COM,UTATIOH

Figure 1. Bs~ic Routing Modol Thc noxt step after d~ciding on ~ ~otric i~ the~oloction of ~ method for computing optimal path~.
Cla~Jical distributed shortost path algorithm~ of ing only; indeed t~tic routing is normally cen- tho Tajibnapic family " ~re prono to transiont trJlized and dynamic routing i~ norm~lly routing loops during convergonco aftor a trunk distributod. Iongth increa~e. As will be ~een in th~ next ~ec-tion THESEUS proposes to romo~bor psth~ ac ~oon THESEUS ollows a quesi-st~tiC distributed ap- a~ they are croated. Cons~qu~ntly, in ordor to pro~ch to routing t~bl~ ccmputation. Better per- ~void the e~t~blith~ont oS infinito paths THESFUS
ormance is thu- expected. In thir context the requiros a shorto3t psth algorithm that provides following section~ expand on the thr~o b~ic rout- 1000-free paths ot e~/ times.
ing component~ for THESEUS: metric evaluation ahort~st p~th computation and flow a~ignm~nt. THESEUS has adapted an algorithm proposed by J~ffe and MosJ' that not only m~intains loop-froo p~ths ~t ~11 ti~es but ~1~3 giv~r the bert f~ilure r~-covery performance. Further~oro it i~ partic-ul-rly well suited to ~p~riodic updating. Anothor // ~ , ~ ~2~27 ~ . .
Ioop-free algorlth~ by herlin and Segall' i- bet- ~ psc~et's path Ident~fler must typlcally b~
ter suited to periodic upda~ing. The enhanced ch~nged at every relay switch akin to ~ logic~l velsion of the 3affe-Moss algorithm used by channel number since path identifiers are inde-THESEUS can be summarized ai follows. The basic pendontly sel0cted by every switch. At dostina-procedure i~ similar to the Tajibnapis algorithm. tion the packet's path identifier is mapped to a For decreasing trunk lengths 3nd for trunk recov- special end-oE-path indicator and further routing ery only this basic procodur~ is used. Houever inEormation in the packet is used to locate the for increasing trunk lengths and for trunk Eail- connection's endpoint. If the path originally set ure the Tajibnapis algorithm is preceded by a Eor a pac~et terminates prematurely e.g. because more coordinated phase that prevents loops. of a failure the packet Is discarded and the end-to-end network protocol ensures recovery.
Let the shortest paths from all sources to a given destination be represented by a spanning tree rooted ~t that destination (an arc in this tree represents one direction of a full-duplex trunk). 8~sic ~tension to the Shortest PJth Alaorith~
When a length increase occurs on a trunk special routing updates propagate to the switches uptree of this trunk as well as the neighbors of these THESEUS extends the distributed shortest path al-uptree switches. The recipients of the special gorithm given in the previous section to incorpo-updates change only their distance to the destina- rate the establishment of paths. The basic idea tion without modifying the ~tructure of the span- i- to let routing updates carry a unique path nin8 tree. Upon completin~ this first phase no identifier in addition to a destination ~nd ~ di~-uptree neighbor of ~ switch can be immediately tanco (for the Tajibn~pis phase only). Sending and chosen by this ~witch to be part oE a new horto9t receiving these new updatos givos ris2 to the ol-path becauLe none pro~ents tho shorteJt di~tance lowine new evont~:
to the destination. No IOOPG c~n then be formed and the second, Tajibnapis phase can s~fely be Suppose switch S receives ovor trunk t a rout-started by the switch that initiatod tho first ing update for destin~tion D uhich contains phase. The Jaffe-hoss ~Igorithm rocoverr fro~ path identifier p. Thi- id2ntifier is re-trunk failure in O(h) ~teps where h is tho height cord2d by S along with the u~ual diftance from of the sp~nning subtree rooted at the point of S to D via t.
f~ilure.
Suppose switch S ostablishos a now short-~t path to D vi~ t. Then, S selects a path iden-_ _ _ tifier q (see below for how q is chosen) ~nd F~O~ ASSlGHM~hT updates its routing tabl2 eo that any packot arriving with path idontifi~r q will be ~ent over t with path identifior p. Finally, switch The shortest path algorithm 31ways delivers a sin- S sonds over its trunks routing updates con-gle path for overy destination. One siQple way to taining the new shortost distance to destina-attain Qultiplo path routing is to assign now usor tion D as ~/oll as path id2ntifier q.
connoction~ only to this neu path and koep oxist-ing connoctions on ~xiJting paths. Such a ~ethod ln order to triggor th~ algorithQ when a trunk although not optimal, conv~rges to a n~ighborhood coQes up, ~ switch sends over that trunk routing of th~ optimuQ and thon oscillatos uithin that updat~ containing its current shortest distance neighborhood'. ~nd path identifior to overy do~tin~tion. A ~witch always has at loast a valid path idontifior to it-solf.
Pccket Forwcrdinq Princip7e This gonoral path ~ssignmont ~ochanissl is b~st ~n with an example. Figure 2 dopicts a cert~in THESEUS identifi~s oach path at tho tiQo it is path assignQent thrUgihUt E just before an update croatod. When a usor connection i~ ostablish~d, t~kes place. The '~' indicato~ that J path is un-each ondpoint obtains the idontifior of tho cur- defin~d or l~ds to a destination othor than E:
ront shortost path to tho other ondpoint. Evory tho '~' is ~n ond-of-path indicator.
p~clet sent over the connection is taggod with tho appropriate path idontifier insto~d of th~ d~rti- Suppo~e that ~s a re~ult of ~ routing upd~te n~tion ~witch idontifior. Bec~u~o tho~e idontifi- ~itch A has d~tached itsel froQ it~ succoJ~or ~rs romain valid oven Dft2r tho cro~tion of now ~witch B in the ~ini~al ~panning tre2 and ~tt~chod shortost paths th~ ~QO paths can be u3ed by ~ it~elf to D new succos~or ~witch D ~SOO Figuro 3).
connoction throughout it~ Iifotir~o (barring any f~iluro). Switch A still recoivn- with path idontifior I tho p~ck~t~ that should go down th~ old path through Routine tJbles ~l~o aro indox~d by path id0ntifi- B Howev~r it neot~ a new p~th idontifior say ~rs in~toad of d~tin~tion idontifior~. In ~ddi- 2 for tho packot~ th~t ~ro to 8 down th; n~u tion to a trunk identiior o~ch o~try contain~ ~horto~t p~th through D. h~anwhilo, both ~witch-~
tho new p~th idontifior to be ~ignod to tho B snd D ~till o~poct to rocoivo p~ck~t- ~ith p~th p~ckot. Indoed unlik~ ~ de~tination identi~ior ;!

~- ~~-~ ~~~ - Fi~ure 4 show5 the stato of the patha af a re-ult oE ~nother porturbation aft~r ~wltch D has st-I Z 3 ¦ t-chod it~elE to ~witch C and switch A has reat-_ _ _ ¦ tached it~olf to switch D At thi- point three

3 _ _ ~ path identiiers are in use in switch A
¦ The exact ~ethod for selecting now path identifi-¦ ers ~t ~ switch for use in routing updatei w~s I left open In reality it divides into four < ' cases Suppo9e that switch 5 ha- chos~n trunk t or a new shortest path to defitination D which happens to carry path identifier p:

I the distance frorl switch S to dostination D
l 2 3 1 2 3 1 2 3 i8 infinite or i p i~ the undefinod path _ _ _ _ _ _ _ _ _ identifier then S select~ the undefined path_ _ 2 _ * 2 ~ 2 * * identifier (No traffie can bo csrriod on a _ _ _ _ _ path labeled with the undefined identifier ) > ¦ -< I If there i5 already a path identifizr q ~ap-E ping to p in the routing table oE switch 5 then S reu~os path identifier q Otherwise iE there is ~ path identifier q *¦ tl ~¦ that is unassignod in tho routing table of ~witch 5 then 5 chooses thi~ path identifior Figure 2 Initial path assign~ont o Otherwiso path identifier- are oxh9ustod at 5witch S; in this case S U50- the undofined p~th idontifier identifier~ 3 and I ro~peetively boc~use they ~ero unaffected by tho routing update Con~o- A notwork should be configurod with onough path quontly ~witch A ~u~t change a packet's p~th idontifiors to avoid the latter kind of con-identifier 1 to 3 beforo forwarding it to B ~nd g~tion It must be notod to tho eredit of A A
~ [~1 r~c ~ D r ~ < I L_~ ¦ . < _~

~ 3~
Figure 3 Path ~signment ~etor first updat- Figure 4 Path assignment aftor second updat-~ng ing . . . ___ , ,_ __ _ _ ~
r~u~t al~o ehange a paekot ~ path identifior 2 to I TNESEUS that ~ince it is ba-ed on ~p~nning troos b~fore Eorwarding it to D it r~quireJ only O(n) path idontifi~rs whore n i;
the nu~bor of ~witehes in a network Source-/~~ :

,,, . . ~ , -de~tination path assignment such as routes in the Sl~AT10~ R~SUlTS
Systems ~etuork Architecture (SNA)' demands O(n') id~ntifiers THESEUS was validated by extensive 6imulation The sample networh studied contained 32 switches interconnected with 57 trunks and repre-ented a Further ~ensions to the Shortes~ Path Al~ real operating network The simulator generated trafEic at the user connection rather than packet granularity The eotablishment of user con-The strategy of assigning new traffic to new paths nections was Poisson For every source the and keeping old traffic on old paths must be ad- choice oE a de~tination switch was random and bi-justed to handle abnormal conditions auch as trunk ssed occording to the actual traffic matrix of the Eailure and recovery and trunk conge~tion Thc sample network, which was Eairly ~par~o ~nd unb~l-general ~ppro~ch takon by THESEUS i~ to t~g rout- anced All simulation run~ were c~rried with the in8 updates with ~n update cause Such ~ cauie i; same oEfered traffic pattern deter~ined at the point of perturb~tion by the 5witch re~pon~ible for goncrating th2 fir~t rout- The ~imulator reacted immedi~tely to th~ mean con-ing upd~te~ The cau~e i~ then ~imply inherited nection throughput For example if ~ 12û b/-from one upd~te to the ncxt A ~witch c~n receive connection wa~ e~t~blished from source A to desti-updates with one cause interleav2d with update~ nation D through ~witche~ 3 ~nd C, then the tr~f-bearing another cause without conusion So~e ic over the A-B 9-C and C-D trunks was typic~l upd~te c~u~e~ ~re con~idor~d in the r~- increa~d immedi~tely by 120 b/~ Furthorr~or~ if m~inder o} thi~ ~ection In gcneral update ~ utilization thre~hold wa~ cro~ed as a re~ult oE
cau~e~ ~re not limited to those thi~ increase, routing updates were genor~ted im-modi~tely Routing update~ were ~cheduled a~
Trunk Failure: When a trunk fail~, messages are event~ in the simulator and intorleaved with ~ll ~ent to erase ~11 the paths using this trunk and the other events free up their identiEiers The ucer connoctions generating traffic on the~e path~ are orced to Sor~e oÇ the cimulation p~ramoters ~re t~bulated stop or a while In parallel, routing updates bolow are disseminated to establish ~ new shortest path However these updates are tagged with the 'trunk packet size exponential with mean - 400 bits f~ilure c~u~e This is an indication that the connection packet throughput exponential u~er connections ~Efected by the trunk ailure with r~ean - 0.3 packets/s of ~ean size mu~t be switched to tho new path Furthermore, 'trunk Eailure' updates can be given higher prior- source-destination selection: random, accord-ity than other kinds o updates ing to the traffic matrix Trunk Recovorv: Anothor po~ible upd~te c~u~e i- connection est~blishment r~te Poisson with 'trunk recovery' When a switch creates ~ new rato ~ ô connoctions/s shortest p~th on the basic oE a routing updat~
carrying thi~ causo, THESEUS proposes to move per- trunk proressor~ capacity 300 packots/~
mancnt virtu~l circuit~ to this path but not es-tablished virtual callc In general, permanent trunk utilizaton thresholds 60X, 75X and virtual circuits should be ~ovod to a new path 82X for an increase 43X, 69X and 79X for a cre~ted ~s a rosult oE a trunk length decrease; decrease thi~ ensures that they converge back to their in-tended paths following ~ trunk recovery The main ~spoct studied by tho simulation was max-imum network throughput (under a congo~tion con-Trunk Congo~tion: ~hen trunk utilization exco~ds str~int) The ro-ults or a notwork with a certain ~aximuru throshold e g , around 80 por- homog~neous 56 kb/s trunks ~how th~t THESEUS de-cont or r~ore THESEUS end~ routing upd~t~s uith livers 46 percent better throughput th~n a inele th~ 'trunk congestion' c3use ~o new usor con- path routing rtratogy whos~ ~etric is invor~ely nections c~n b~ ignod to ~ ~hortoJt p~th built proportion~l to trunk c~pacity (Eor ~t~bility) fro~ such updat~ A drop in trunk utilization For both ~tratogie~ no connoction~ were rele~sed can lator trig8er new routing updates and recreat~ and the ~imulation was tdrmin~ted as soon ~c a an open ~hortost p~th Thi~ ~implo on/ofE con- trunk in the network was congosted i e , over 82 gestion control mechlni-m sllows for tht detection perc~nt utili~cd snd control of bottlon~cks in a netuork It is r2mini~cent oE chokd p~cket~'; howovor, THESEUS At the ti~o th~ ~imul~tion ~toppod, an ~ver~ge oE
coQpl~tely int~gr~t~s thi~ warning to th~ hortost 3 75 paths tand a Qaximum of 12) h3d beon cre~tod p~th ~Igorithm por ~witch p~ir, out of uhich ~n avor~go of 1 72 p-th (and 3 rl3ximum oE 9) w~ c-rrying tr~fEic Av~r~gc network del~y w~s lower even when theso extr~ psths involvod moro hops bec~u~ Ion8 u~it-in~ timd~ wero avoided Figur~ 5 plot~ th~ n~t-~ / ~

24~3Z7 uork ~verage number o~ hops again~t the ~ppliod Be~ides maximum throughput the othor ~jor s~p~ct lo~d uhere the average number of hop~ is th~ ra- ~tudied by the ~imulation wa~ ~to~dy ~t~te tio of internal to external netuork throughput. throughput. Thi~ was achiev~d by loading the si~-ulated network with connection~ of exponential du-A version oE the network with mixed trunk spe~ds ration with various means. The result9 are 6hown was also studied. Switches located in the same in Figure 6 for mean connection durations of 5 city were conslected with 112 kb/s trunkc while the 10 and 15 minutes. ~n steady ~tatc, network speed of inter city trunks was maintained at 56 throughput should oscillatc around:
kb/s. ~ith single path routing the higher speed trunks tend to attract too much traffic at the ex- connoction establishment rate pense of the lower speed trunks. A~g~in THESEUS ~ connection mean throughputavoids this by diverting traffic over alt~rnate * connoction mean duration paths. In such a network THESEUS achieved 99 percent better throughput than single path rout-ing, when the simulation was stopped as soon as a Finally regarding control information the excess trunk was congested. trafEic due to routing updates was found to be negligible (less than 0.1 percent) and did not af-~ore stress was placed on THESEUS by pushing the foct the simulation results. Furthermore the av-simulation until 70 percent of all trunks were orage numb~r of path identifiers used by ~ switch conge3ted in the homogeneous n~twork. A well-known of the homogeneous network at the end of the ~imu-phenomomon was then obsorved: network throughput lation was 120, with a mzxlmum of 159. Out of became worse for multiple path routing. This is this nu~bor an averago of 55 identifiors only because the traffic committed at lower applied were ~s~igned to p~ths carrying traffic with a lo~ds over p~th~ with extra hop~ prevent~ addi- maximum of 72. This indicntos th~t the d3mand on tional traffic frottt t~ntering the notwork whon p~th identifiers is not ttxco~sivc, oven though highor load~ ~r~ oEforod. But such l~vols of con- thore is a case for rouslng the identifiors of gortion ;tro highly improbable ~nd if 9 ~ystem idlc or qua~i-idlc paths.
ehould be ~blo to cope with improb~blo ~ituations, it should not neceslarily be optimizod for thets.
Note th~t uhen 35 percont of the trunks were con-ge~ted multiple p~th routing w~s ~till ~uporior.
Finer ~p~ct~ of TH~SEUS wore ~I~o ox~in-d. For oxample, the number of threshold~ ct~rt;tinly doter-_ .. .. ....
2.70 Slnobp~h p ttOO t Smlnutt~ ~
D Mull~p~th / ct_ _10m1nuto~ /
2.65 - / 700 _ 15 m1nut~ /
2.60 _ ~ tiOO /~ i 2.55 _ / ~,500 /~ 1.
2~ 2.50 ~ ~CO //
L~

O ~O~t 4~10 tSOO t~OO 1000 12~ 1~00 0 500 1000 1500 2000 2500 ~ppll~Ct to tt ~t~
Figure 5. Tr~Efic Sproad Figure 6. Ste~dy State Throughput.
~ines how active the routing ~yrtem is it pre~um- C
~bly ~Iso ~ffect~ perforrtanco. The simulation indic~tod th~t by incro~sing tho nuDtber of thro~h-olds (e.g. from 3 to 7), only c slight ir~provtl- The worx doscribod in thi~ paper oogan with the nent in nt~twork throughput ~ ob~orvod ~e.g., obj-ctive of incro~sing thr ~axi~u~t throughput of lo~e than 4 porcent). p~ck-t 3witching nettork- with ttixed trunk 9poodt~

/S
!

,~,/ ~, sp~rse traffic matrices and high connectivity. 3. Bertsekas D.P. "Optimal Routing ~nd Flow Th~oretically a nece~sary condition ~as known: Control Methods for Communication Network~."
multiple path routing had to be used in order t~ ln Ano1~sis dnd Optimi~ation of Systems, pp.
even out traEEic over the network. The problem 615-643. dited by F. Eensoussan and J.-L.
was to Eind a practical ~pproach that would con- Lions. Eerlin: Springer-Verlag 1983.
ver&e roasonably close to the optimum for quasi-statlc tr~ffic conditionc while proserving an 4. D~vies D.W.; Barber D.L.A.: Price W.L. and end-to-end architecture and without introducing Solomonides C.M. Computer ~etworks ~nd Their adver~e effects ~uch a~ packet di~ordoring and ox- Protoco1s. Chichostor: John Wiloy 1979.
ce~sive control information.
5. ~affe J.M. nnd Mo~s F.H. "A Respon~ive Dis-The corner~tone of the solution developed in the tributed Routing Algorithm for Computer Net-foregoing sections is a management scheme for Qul- works." IEEE Tr2ns. Communic. C0~-30, 7 (July tiple p~ths that is completoly integrated to ~ 1982): 1758-1762.
Ioop-free distributed shortest path ~Igorithm. In the context of this scheme two design decisions 6. Kaliszcwski J.M. "The New Routing Algorithm were made to cut down on the ~mount of control in- on the SITA ~ata Tr~nsport Network." In Per-formation required ~t the risk of converging less formcnce of Computer-Communicution Systems, ~ccur~tely. The first one s~id th~t routing up- pp. 415-431. Edited by H. Rudin ~nd W. Bux.
d~te gener~tion would be aperiodic ~nd based on Amsterdam: North-Holl~nd, 1984.
trunk utilization thresholds. The second one con-cerned Elow assignr~ent. The simplest scheme was 7. Rleinrock L. "Principles ~nd Le~sons in ~dopted for which new user connections would be Packet Communications." Proc. IFEE 66, 11 ~ssigned to current shortest path~ while exirting (November 1978): 132û-1329.
u~er connections would be m~intained on existing paths. 8. Hajithi~ J.C.: Irl~nd M. Grange, J.-L.
Cohen, M.- and û'Donnell, C 0. "xperir~ents SiQulation has shown th~t these prsctical choices in Congestion Control Techniques." In F10w were not overly simplifying. The increase in Contro1 in Computer Networks, pp. 211-234.
throu~hput obt~inod for ~ particular notwork undor dited by J.-L. Grange ~nd M. Gien. A~terd~m:
a congestion constr~int was i~pr~ssive compared to North-Holland, 1979.
~ single p~th routing strategy i.e. up to 100 percent. However, it is not known how far from 9. M~rlin D.M. and Seg~ll, A. "A Failsafe Dil-th~ optimum the botter network throughput li~. tribut;d Routing Protocol." IEEE Trons.
It w~ further observed th~t just ~ few threshold~ Communit. C0~-27, 9 (September 1979):
wer~ needed to achieve this level of improvom~nt; 1280-1287.
adding more thresholds yielded only ~ small mar-ginal g~in. 10. Seligman D.R. "Traffic Routing in ~ Co~puter Network." Performnnce 7, 2 (April 1984):
Inasr~uch ~s the results Oe the simulation c~n be 59-64.
tr~nsposed to an oper~ting network, this path~
oriented routing trategy is stable, increases 11. Tajibn~pis W.D. "A Correctness Proof of nrtwork throughput ~ub~tanti~lly, and does not re- Topology Information M~inton~nce Protocol for quire ~n excessive ~ssount oE control inform~tion Di~tributed Computer Networkr." C. AC~ ~0, 7 to be ~tored ~nd ~xch~nged. (July 1977): 477-485.

ACKllOlr~ EDGFMENTS
We would liko to th~nk W S. L~i for his technic~l advice ~nd F. ~ellor for his const~nt upport throughout this projoct.

REFf REI~C~S
1. Ahuj~ V. "Routing and Flow Control in Sys-tems Network Architecture." IEr~ Syst. J. lP
2 (1979): 298-314.
2. Aubin R. "Including Processing ~nd Prop~-gation Del~y to Routing in Packet Switching Networks." Proc. Int. Conf. Communicstions Chic~go 1985 pp. 107-111.

/~

~2~32~

APPENDIX B
Shortest Path Algorithm Case 1. Switch S receives UPDATE (DESTINATION-ID, DISTANCE, PATH-ID) over trunk T:
a. Set TRUNK TABLE (T9 DESTINATION-ID) to (PATH-ID, DISTANCE + METRIC(T)).
b. Reevaluate to shortest distance and the best outgoing trunk for DESTINATION-ID in the shortest path table, in the light of the UPDATE received.
c. If the shortest distance to DESTINATION-ID has changed:
(i) update the shortest path table by assigning the TRUNK ID, the DISTANCE, and the PATH ID of the best outgoing trunk to the entry corresponding to DESTINATION-ID
(ii) select a new PATH ID (using the New Path ID
Selection Algorithm) and update the routing table by assigning the TRUNK ID and the PATH ID
of the best outgoing trunk above to the entry corresponding to this new PATH ID
(iii) send UPDATE (DESTINATION, new shortest distance, new PATH ID) over all trunks outgoing from switch S.
Case 2. The value of METRIC(T) changes at switch S:
For all DESTINATION-ID, do:
a) Increment or decrement the DISTANCE field in TRUNK TABLE
(T, DESTINATION-ID) according to change in METRIC(T).

` ~2453Z7 b) Do l.b and l.c above.
New Path ID Selection Algorithm Assume switch S selects T and P to be the best outgoing TRUNK
and PATH ID for a destination D, then a new PATH ID is selected by S as follows:
Case 1. If the distance from S to D is infinite or if P is the undefined PATH ID, then S selects the undefined PATH ID.
(No traffic can be carried on a path labelled with the undefined PATH ID).
Case 2. Otherwise, if there is already a PATH ID Q mapping to P in the routing table, then S reuses Q.
Case 3. Otherwise, if there is a PATH ID Q that is unassigned in the routing table, the S chooses Q.
Case 4. Otherwise, S uses the undefined PATH ID.

~/ ~ ~

Claims (4)

THE EMBODIMENTS OF THE INVENTION IN WHICH AN EXCLUSIVE
PROPERTY OR PRIVILEGE IS CLAIMED ARE DEFINED AS FOLLOWS:

1. A path-oriented routing method for packet switching networks, wherein said network is comprised of a plurality of interconnected packet switches wherein at any given period of time each said packet switch can be functioning as a destination switch, an originating switch, a tandem switch, or any combination of the preceding, said method characterized by:
each given packet switch broadcasting to all its adjacent neighbouring switches a preferred single path identifier to use to send messages to it; and each successive packet switch, moving monotonically in a direction away from said given packet switch, broadcasting to all its adjacent neighbouring switches, the preferred single path identifier to use to send messages to it destined ultimately for said given packet switch.

2. A path-oriented routing system for packet switching networks, wherein said network is comprised of a plurality of interconnected packet switches, said system characterized by:
a single path identifier being associated with each packet at the source packet switch and carried by the packet as it traverses through said network to a given destination switch, said single path identifier being updated at each packet switch traversed.

3. The path-oriented routing system of claim 2 wherein each said packet switch maintains information concerning the shortest path from itself to a given destination switch, and in addition maintains information for relating the incoming single path identifier to both a single outgoing path identifier to be appended to a packet and a trunk identifier to be used locally by said packet switch.

4. A method of routing packets of information on a packet switching network comprised of a plurality of interconnected packet switches, said method characterized by:
assigning to each said packet as it is being assembled, a single path identifier indicative of the path said packet is to follow on the way to its destination;
updating periodically said single path identifier that is to be assigned to the packet at the originating switch, wherein packets corresponding to new connections may be assigned new path identifiers, while packets corresponding to existing connections continue to employ the same path identifier as did earlier packets in the same connection; and routing each said packet, at each said packet switch, to the path indicated by said single path identifier.