HETEROGENEOUS CHANNEL BONDING REVISITED

摘要

Efficient communication in distributed systems is essential for optimal system performance. As parallel computation develops, many different kinds of messages need to be sent between machines. These messages may have different requirements for network latency and bandwidth. In this paper, we revisit combining two physical interfaces into one virtual network (channel bonding). In contrast with most existing implementations, we communicate via heterogenous interfaces, and explore a number of more sophisticated scheduling algorithms than the round-robin technique used in the standard Beowulf channel-bonding driver. We explore the relative tradeoffs of interface scheduling overhead versus the additional bandwidth possible with multiple interfaces. Using both Gigabit Ethernet and Fast Ethernet Networks, we evaluate system performance using netPerf. We show significant performance gains for small messages, with incremental improvements for larger packets.