An evaluation of the Myrinet/GM2 two-port networks

Abstract

It is important to systematically assess the features and performance of the new interconnects for highperformance clusters. This paper presents the performance of the two-port Myrinet networks at the GM2 and MPI layers using a complete set of microbenchmarks. We also present the communication characteristics and the performance of the NAS Multi-Zone benchmarks and SMG2000 application under the MPI and MPI-OpenMP programming paradigms.We found that the host overhead is very small in our cluster, and the Myrinet is sensitive to the buffer reuse patterns. Our applications achieved a better performance for MPI than the mixed-mode. All the applications studied use only nonblocking communications, thus able to overlap their communications with the computations. Our experiments show that the two-port communication at the GM and MPI levels (except for the RDMA read, and overlap) outperforms the one-port communication for the bandwidth. However, this did not translate in a considerable improvement at least for our applications.

High-Performance Multi-Rail Support with the NEWMADELEINE Communication Library

“…At the lowest part, the transmit layer is in charge of interfacing NEWMADELEINE with the networking hardware and associated specific APIs through a set of drivers. NEW-MADELEINE currently provides drivers for the Quadrics Elan API [10], the Myricom Myrinet Express [12] and GM-2 APIs [17], the Dolphinics SiSCI API [9] and the legacy socket API on top of TCP/IP.…”

A Speculative and Adaptive MPI Rendezvous Protocol Over RDMA-enabled Interconnects

“…In their method, the communication overhead is first computed and then application availability is calculated using the overhead amount. While [20,21] address combined send and receive overlap, our proposed overlap measurement method in this paper, along with [8,19], target non-blocking send and receive overlaps separately.…”

Improving Communication Progress and Overlap in MPI Rendezvous Protocol over RDMA-enabled Interconnects

“…In their method, the communication overhead is first computed, and then application availability is calculated using the overhead amount. While [8,23] addresses combined send and receive overlap, our proposed overlap measurement method in this paper, along with [5,22], target send and receive overlaps separately.…”