DTJ Number 3 September 1987 - Digital Technical Journals

---Performance Analysis and Modeling of Digital's Ne tworking ArchitectureA study of the system usage behavior at six differentuniversities showed that a significant portion(about 30 percent) of the user's time isspent in editing.23 This conclusion led us to useour text editor (EDT) as the key user level benchmarkfor network performance studies. The studyresults also led to the transport layer performanceimprovements discussed earlier.A study of network traffic at M.I.T. showed thatthe packets exhibit a "source locality."7 That is,given a packet going from A to 8, the probabilityis very high that the next packet on the link willbe going either from A to 8 or from 8 to A. Theseobservations helped us improve our packet-forwardingalgorithm in bridges. The forwardingdecision is cached for use with the packets arrivingnext. A two-entry cache has been found toproduce a hit rate of 60 percent, resulting in significantsavings in table lookup.The principal cause of source locality is theincreasing size of data objects being transportedover computer networks. The siz es of dataobjects have grown faster than packet sizes have.Packet sizes have generally been limited by thebuffer sizes and by the need to be compatiblewith older versions of network protocols. Transferof a graphic screen could involve data transfersof around two million bits. This increase ininformation size ·means that most communicationsinvolve a train of packets, not just onepacket. The commonly used Poisson arrivalmodel is a special case of the train modelJThe two major components of a networkingworkload are the packet size distribution andthe interarrival time distribution. ]. Shoch and]. Hupp made the classic measurements of thesecomponents for Ethernet traffic. 24 Their testshave been repeated many times at many places,including Digital. The bimodal nature of thepacket size distribution and the bursty nature ofthe arrivals are now well accepted facts; we willnot elaborate further on them.The utilization of networks is gen erally verylow. Measurements of Ethern et traffic at one ofour software en gineering facilities with 50 to 60active VAX nodes during normal working hoursshowed that the maximum utilization during any15-minute period was only 4 percent. Althoughhigher momentary peaks are certainly possible,the key observation, confirmed by other studiesas well, is that the network utilization is normallyvery low. While comparing two alternatives,say H and L in Figure 6, some analystswould choose alternative H, which performs betterthan L under heavy load but worse under lightload. Our view of this choice is quite different.We feel that, while high performance at heavyload is important, it should not be obtained atthe cost of significantly lower performance atnormal, light load levels. Therefore, the choicebetween L and H would also depend upon theperformance of Hat low loads.w()z