Throughput analysis between high end workstations across an FDDI network

Abstract

Recently developed high speed networks are capable of transmitting data at rates of 100 Mbps or more. One such network protocol is Fiber Distributed Data Interface (FOOl). This network has a physical. transmissiop rate of 100 Mbps. Analytical and simulation studies have shown that t:le FOOl protocol should provide actual throughput of HO% to 95% of this physical rate. Can the end user expect to see this kind of performance? If not, then what kind of throughput can actually be expected and where are the bottle necks? To answer these and other related questions, two areas were studied: First, a perfomance parison between a 40MHz SPARCstation 10 workstation and a 50MHz SPARCstation 10 workstation was conducted using the Neal Nelson commercial benchmark tool. Next, a well-known network measurement tool, ttcp, was used to obtain data transfer rates while varying several tunable operating system and network parcll11eters. The parameters varied were: Target Token Rotation Time, TCP/IP window size, NFS asynchronous threads, Logical Link buffer siu and Maximum Transfer Unit size. The results from the commercial benchmar~ analyis were used to determine if there are any differences which can affect transfer rates between the two workstations. The results from the conunercial benchmark tool clearly showed that the newer, higher speed processor is faster. The network tool ttcp showed that the TCP/IP window size had the largest impact on throughput performance. Throughput more than doubles from a window size of 4k to a window size of 20k. This is followed by having more than one workstation transmitting data simultaneously. Having two workstations transmitting nearly halves throughput This is followed by having a faster processor. A measurement of flle transfers using rep system calls showed that the largest impact on file transfer speed is the overhead of receiving the transferred file.