Performance evaluation of a Gigabit Ethernet switch and Myrinet using real application cores

purpose network hardware, thereby losing the cost benefit offered in mass market. Riding on the wave of Ethernet popularity, Gigabit Ethernet is fast becoming a commodity item. Given a scalable switching architecture, we believe it can be a cost-effective alternative to interconnect parallel systems. This study evaluates the performance of the Avici Gigabit Ethernet switch against Myrinet. We simulate a 256-node cluster running core algorithms from real parallel applications, and then compare raw performance figures such as bandwidth and latency, as well as more complex parameters such as jitter, routing, and points of congestion in the fabric. A. INTRODUCTION The idea of cluster computing is to aggregate machine rooms full of relatively cheap hardware, connected with some sort of network, and apply the combined force of the individual machines to a single calculation. The hardware employed in a cluster is generally from the volume personal computer market, so as to leverage the cost advantages of buying commodity hardware. As this architecture has distributed memory, parallel processes communicate using a message-passing paradigm. Unfortunately, highspeed interconnect is completely irrelevant for the mass market. High-end cluster users, therefore, have to rely on special purpose hardware such as Myrinet [1], HiPPI [2], or ServerNet [3] for their message passing infrastructures, losing the cost benefit offered by the commodity market. The purpose of this study is to identify a scalable and cost-effective alternative to the traditionally expensive interconnects. Riding on the wave of Ethernet popularity, Gigabit Ethernet is fast becoming a commodity item. In addition to bandwidth enhancement, the full-duplex mode of Gigabit Ethernet [4] allows switched access at full channel capacity without the limitation of CSMA/CD. Therefore, we believe that, given a scalable switching architecture, Gigabit Ethernet can be a cost-effective solution for cluster computing. The remainder of this paper discusses the simulation results of the Avici based Gigabit Ethernet switch, and the methods we used, which involve a mix of " artificial " basic tests and simulations of core algorithms from real parallel applications. We compare our results against an identical study of Myrinet, a popular interconnect technology, and tally the positive and negative aspects of each. B. INTERCONNECT TECHNOLOGIES The following paragraphs describe the network technologies we considered in the simulations that will be discussed in Section C. In each section we calculate the current pricing for a prototypical 256-node cluster, a size which is feasible for most high-end cluster builders. 1. …