Speaker: Jamshid Mahdavi
Making TCP Work for High Performance NetworksDate: June 13, 1996
Abstract: The PSC has long been a proponent of high speed distributed applications. Over the years, we have worked on distributed applications using the NSFnet, the NECTAR testbed, local HIPPI networks, and most recently, the vBNS. TCP is the preferred transport layer for distributed applications, because it allows for easy programming and widespread portability. However, making TCP work for these applications requires taking advantage of several little known features of TCP/IP and careful tuning to get the maximum possible performance. Even so, there are many unsolved problems in making TCP work with these applications.
In this talk, I will discuss standards for MTU discovery, TCP large windows, and Selective Acknowledgment. Some of the discussion will be operational in nature, focussed on throughput expectations in today's high speed WAN environments. I will also discuss optimal packet sizes for high speed networks. TCP congestion control algorithms also have a large impact on overall performance. Slow-start can be a big performance hit, and I show some examples of distributed computing applications which demonstrate this. Finally, I'll discuss Reno's implementation of TCP congestion avoidance and how this hurts in high-speed environments.