Following on from an announcement at the end of last year, a consortium of IT vendors has published a white paper outlining a reference stack for high performance trading applications, and has documented a set of benchmarks conducted on commercial and open source FIX engines to demonstrate different performance characteristics of the stack.
As it stands, the consortium – led by OnX Enterprise Solutions – includes Dell, Intel, Arista Networks, Solarflare Communications, Equinix, GreySpark Partners and Edge Technology Solutions. FIX engines from B2BITS and Rapid Addition were put under test. With the publication of the reference stack, the group is now open for additional participants.
The reference stack developed includes systems software, compute and networking components, and has been designed for implementation within a co-location centre. See below:
With a reference design created, the stack was built for real at Intel’s fasterLAB in the UK, and was then optimised and tuned for performance, predictability and reliability. OnX undertook the vendor selection and integration of the components, as well as overseeing the benchmarks, and analysing the results.
In order to exercise the stack for real-life trading applications, the consortium ran a number of benchmark tests for FIX messaging, comparing commercial FIX products with QuickFIX open source offerings. Both C++ and Java engines were tested, to reflect the reality that both technologies are widely deployed by trading firms.
In all, around 90 tests were conducted, with the FIX engines’ ability to process market data and generate buy and sell orders tested at different throughput levels and over different time spans – for bursts and extended periods of several hours, to document latency and latency variance in a simulated real-life trading scenario.
Not surprisingly, the commercial FIX engines showed a significant performance improvement over their open source counterparts, although notable was the very similar performance of both B2BITS’ C++ engine, and the Java-based product from Rapid Addition.
Highlights from the tests showed the commercial products were between four and 16 times faster than the open source equivalents, with an average latency test result of 11 microseconds, as opposed to 180 microseconds. The performance gap was most evident when the test harness was tweaked to reflect faster matching, and so faster order acknowledgements.
Next up for the consortium is to put in place commercial arrangements and processes so that trading firms can purchase and deploy the reference stack as a package. It is expected that further technology vendors will join the charter group, and that other applications beyond FIX engines will be integrated, and benchmarked.
In other – related – news … FIX vendor Connamara Systems has released FIXpresso, a commercial engine based on the open source QuickFIX/J, which it claims reduces latency by 98% over the standard distribution, with a standard deviation of under one microsecond. The company spent nearly 2,000 man hours to rewrite the code, focusing on reducing garbage collection.