Fast NIC-based barrier over Myrinet/GM

Abstract

An efficient barrier implementation is desirable on parallel systems to obtain good parallel speedup and to support finer-grained computation. Some modern Network Interface Cards (NICs) have programmable processors which can be used to provide support for collective communications such as barrier In this paper we utilize such a programmable NIC to provide an efficient barrier synchronization operation. This paper describes the design, implementation and evaluation of a NIC-based barrier operation as an addition to Myricom's GM message passing system. Our NIC-based barrier implementation achieved a barrier latency of 102.14 /spl mu/s for 16 nodes which is a 1.78 factor of improvement over the host-based barrier using the same algorithm for LANai 4.3 NIC cards. Using LANai 7.2 cards, which has a faster processor, we achieved a 1.83 factor of improvement for eight nodes. Our NIC-based barrier operation promises scalable fine-grained parallel computation over clusters of workstations. To the best of our knowledge, this is the first NIC-level barrier implementation on a cluster with Myrinet/GM.