Exploiting multiple heterogeneous networks to reduce communication costs in parallel programs

Abstract

The different types of messages used by a parallel application program executing in a distributed system can each have unique characteristics so that no single communication network can produce the lowest latency for all messages. For instance, short control messages may be sent with the lowest overhead on one type of network, such as Ethernet, while bulk data transfers may be better suited to a different type of network, such as Fibre Channel or HiPPI. In this paper, we investigate how to exploit multiple heterogeneous communication networks that interconnect the same set of processing nodes by dynamically selecting the best (lowest latency) network for each message based on the message size. We also show how to aggregate these multiple parallel networks into a single virtual network to further reduce the latency and increase the available bandwidth. We test this multiplexing and aggregation on a cluster of SGI multiprocessors interconnected with both Fibre Channel and Ethernet. We find that multiplexing between Ethernet and Fibre Channel can substantially reduce communication overhead in a synthetic benchmark compared to using either network alone. Aggregating these two networks into a single virtual network can further reduce communication delays for applications with many large messages. The best choice of either multiplexing or aggregation depends on the mix of message sizes in application program and the relative overheads of the two networks.