Support for MPI at the Network Interface Level

Abstract

Commodity components and high speed interconnects allow to build fast and cheap parallel machines. Several research works demonstrated that the raw performance of MPI ports on top of user level network interfaces are close to the hardware specifications. However, most of these implementations offer poor possibilities of communication/ computation overlapping because the application thread on the host processor is responsible of a large part of the protocol. Thus, application performances are limited. One solution to this problem is to move the critical part of the protocol at the network interface card level. We implemented such a communication system that relies on the embedded processor of the Myrinet host interface. We present the problems moving part of the protocol at the network interface level introduces, our design choices and we also give performance measurements.