DTJ Number 3 September 1987 - Digital Technical Journals

Performance Analysis and Modeling of Digital's Networking Architecturepaths usable by a source than there are packetsper burst, then burstiness will have little effecton either the mean or the variance of waitingtime. On the other hand, only two or three alternativeequiprobable paths are enough todecrease the bursty-packet waiting time fromone-half to two-thirds for the first hop. Thepacket bursts will tend to spread apart as theypropagate, so that the improvement in subsequenthops will be somewhat less.Transport Layer PerformanceSeveral studies have been published on the performanceof the transport layer in the DNA structure.18·19·20·21 One of the published studies is ontimeout algorithms. We found that under sustainedloss, all adaptive timeout algorithmseither diverge or converge to values lower thanthe actual round-trip delay. 18 If an algorithmconverges to a low value, it may cause frequentunnecessary retransmissions, sometimes leadingto network congestion. Therefore, divergence ispreferable in the sense that the retransmissionsare delayed.HOSTNODETERMINALNODEAPPLICATION TRANSPORT TRANSPORT APPLICATIONLAYER LAYER LAYER LAYERFigure 5CREDIT = 1Eight Transport Level PacketsOne key lesson we learned from the timeoutalgorithm research was that a timeout is also anindicator of congestion in the network. Therefore,not only should the source retransmit thepacket on a timeout, but it should also takeaction to reduce future input into the network.There is a timeout-based congestion control policycalled CUTE (congestion control using timeoutsat the end-to-end layer) that manages theseactions. 19Among the new features of DNA Phase IV arecross-channel piggybacking, acknowledgmentwithholding, and larger flow-control windows.These features were introduced as the result of astudy that concluded that straightforward terminalcommunication over a DECnet networkwould be slow. This conclusion lead eventuallyto the development of a new local area transportprotocol, called LAT, for terminal communications.These enhancements were also addedto the DNA transport protocol . This study isdescribed below.In the DNA structure, each transport connectionhas two subchannels: one for the user, andone for control. The user subchannel carries userdata and their acknowledgments, called acks.The control subchannel is used for flow-controlpackets and their acks. Protocol verification canbe easily achieved if the two subchannels areindependent so that information on one channelis not sent on the other. In studying terminalcommunications over a IAN, we discovered thateach terminal read took eight transport protocoldata units (TPDUs) , as shown in Figure 5. Eachunit consists of two application level packets: aread request, and a data response. Each packetrequires a link service packet from the respectivereceiver; this service packet permits the senderto send one packet. The remaining four units aretransport level acks for these four packets.Given the CPU time required per packet, wecomputed that communication for remote terminalstakes four times as much CPU time as that forlocal terminals. Therefore, our goal was toimprove performance by a factor of four. We proceededin three ways to solve this problem. First,we modified application programs to utilizelarger flow-control windows; seco·n d, wesearched for ways to reduce the number of pack:'ets per I/0 operation; third, we tried to reducethe CPU time required per packet. The first goalwas achieved by multibuffering, discussed laterin the section "Application Layer Performance."30Digital Tecbnical]ournalNo. 3 September 1986