Metamako's MetaApp32 under STAC-N1

Mean half-round trip latency through aggregation and layer 1 of just 0.06 microseconds.

27 July 2015

STAC recently performed tests to assess the mean latency and jitter of the Metamako MetaApp 32 device running the metamux-0.5 aggregation application. MetaApp 32 includes a layer 1 switch, a four-core server, and an FPGA for running pre-packaged or custom modules. The metamux application aggregates streams into a single output, such as interfacing multiple trading servers with an exchange or microwave connection, while return traffic can take advantage of the low-latency layer 1 switching capability. For example, trading applications can send orders via the aggregation path and receive responses and market data via layer 1.

To perform this assessment, we compared STAC-N1 benchmark results for the same technology stack when using the MetaApp 32 and when using no switching technology at all. In light of the expected low latency, the configuration with the MetaApp 32 forced each message to make multiple trips through the metamux application and through the layer-1 switch path. The total mean latency of these trips was divided by the number of trips to arrive at the mean latency per round trip (through both metamux and layer 1).The accuracy of the test harness only allowed us to measure the mean latency and standard deviation (jitter) of each SUT, since these have a smaller error than individual observations due to the large sample size.

A comparison report analyzes the latency between the two configurations and contains links to individual reports for each one. We tested each SUT at 6 different message rates and 4 different messages sizes.  For the baseline message size of 264 bytes, we found:

• Round trip in the Metamako devices (one trip through aggregation plus one trip through layer 1) introduced a mean incremental latency of 0.1 microseconds across message rates.
• Per-hop latency (i.e., half-round-trip, a common metric cited for switches) was 0.06 microseconds. This is the average of the aggregation hop and L1 hop in each round trip. It differs from half of the round trip number above due to rounding.
• Any incremental jitter introduced by the roundtrip through the Metamako devices was undetectable by this methodology at all message rates except 800K msg/sec, at which the mean difference was 0.04 microseconds.

Results for all message sizes, along with detailed configuration information, are available to STAC Benchmark Council members at the same pages as the reports.