Using CMT in SCTP-based MPI to exploit multiple interfaces in cluster nodes

Abstract

Many existing clusters use inexpensive Gigabit Ethernet and often have multiple interfaces cards to improve bandwidth and enhance fault tolerance. We investigate the use of Concurrent Multipath Transfer (CMT), an extension to the Stream Control Transmission Protocol (SCTP), to take advantage of multiple network interfaces for use with MPI programs. We evaluate the performance of our system with micro-benchmarks and MPI collective routines. We also compare our method, which employs CMT at the transport layer in the operating system kernel, to existing systems that support multi-railing in the middleware. We discuss performance with respect to bandwidth, latency, congestion control and fault tolerance. © Springer-Verlag Berlin Heidelberg 2007.