EMP: Raw Benchmarks

The following plots show the usual raw numbers for a back-to-back configuration of two hosts. Adding a switch in the middle increases the latency by 5us but does not affect the bandwidth. Note that all the numbers below are for reliable transmission, the only mode of operation of EMP, and the time for the NICs to generate and process acknowledgements is thus included in the figures.

Latency is measured by timing a round-trip message and dividing by two. Each test is actually performed 10000 times and the entire time to complete all the little messages is used to generate the average shown in the graph.

Bandwidth (throughput) is measured using a unidirectional send of 1000 message, with a single return message signalling completion.

Bidirectional throughput measures the case where both of the two hosts are busy sending data to each other. Note that the theoretical maximum here is twice the link speed for full duplex media.

Return to EMP.