FAST TCP in High-Speed Networks: An Experimental Study

We conducted various experiments using FAST TCP on two high-speed networks and uncovered some of the hidden problems affecting high-speed TCP implementations in practice. While FAST TCP achieved good performance in many scenarios, there were also a number of cases, where adverse conditions, both host related and network related, prevented it from obtaining the expected throughput. Furthermore, tests with other protocols confirmed the existence of performance bottlenecks, some of which are still unknown to us. While various high-speed protocols such as FAST TCP will continue to improve with time, it is clear that in order to realize the vision of high-speed remote data access in Grid computing, the protocol researchers, the OS implementers, the network providers, and equipment manufactures need to understand and eliminate as many performance bottlenecks as possible. To become immediately deployable today, a high-speed TCP protocol not only needs to have robust performance under all scenarios, but also has to have the ability to provide as much information as possible to pin-point performance bottlenecks.