Exploiting Spatial Parallelism in Ethernet-Based Cluster Interconnects

In this paper, the authors examine the implications of building a single logical link out of multiple physical links. They use MultiEdge to examine the throughput-CPU utilization tradeoffs and examine how overheads and performance scale with the number and speed of links. They use low-level instrumentation to understand associated overheads, they experiment with setups between 1 and 8 1-GBit/s links, and they contrast their results with a single 10-GBit/s link. They find that: their base protocol achieves up-to 65% of the nominal aggregate throughput. Replacing the interrupts with polling significantly impacts only the multiple link configurations, reaching 80% of nominal throughput.