Network Subsystems in Mpps: Where Did All the Performance Go?

In this work we describe our results on identifying the most important overheads in the network subsystem of massively parallel processors (MPPs), specially the Intel Paragon. We show how poor implementation techniques currently prevent performance of applications using TCP/IP protocols to communicate between supercomputers connected by high performance networks from achieving data rates close to the capacity of the medium. We identify some of the possible solutions to the problem, and discuss what can be expected in future systems. In particular, we present our results on evaluation of some of those solutions, including a user-level protocol implementation which scales with the number of concurrent connections and performance superior to the current system.