Super Micro Computer last month introduced a new line of “Hyper-Speed” servers, pitching them at high frequency trading applications. The company says they work by “maximising processing power and precisely tuning hardware and firmware to attain up to 30% lower latency.” And just as importantly, the servers feature “high reliability as a primary design focus.” So what are these servers, and how does Supermicro boost their performance?
* They are based on dual Intel multi-core Xeon E5-2600 (Sandy Bridge) chips. The 2U model is pitched at space-limited co-location deployment. Crucially, they are “designed for maximum airflow and custom heat sinks provide optimal thermal distribution for mission critical reliability.” Which means they provide superior cooling for the chips and other components.
* This cooling is required because the microprocessors in the Supermicro servers are over clocked – they operate at clock frequencies higher than normal, which means they execute code faster. Which translates to lower latency.
* As an example, Supermicro tested network performance processing messages with the UDP and TCP protocols, and determined that over clocking reduced latency by 31%. Moreover, jitter was reduced by 73%.
* Over clocking of its microprocessors is generally frowned upon by Intel because the implementation of over clocking can shorten their life. But because of the cooling abilities of the Supermicro servers, the chips have the same warranty as standard products. Also, Supermicro tests all of the components in its servers – including RAM and network interface cards – to ensure they will operate with an over clocked microprocessor.
* Just to be clear, the performance of the Supermicro servers is not just down to microprocessor over clocking but also to other hardware and firmware optimisations incorporated into them.
For applications that need the fastest execution of serial code – data feed handlers are an example – over clocking should be beneficial. And Supermicro’s implementation seemingly overcomes the reliabity issues often cited for over over clocking.
For some, this approach will have merit compared to turning to specialist processors, such as FPGAs, to handle certain processing. It will mean that traditional programming techniques can continue to be used instead of requiring specialist skills. And existing mainstream code can be readily deployed, and runs faster.
It’s worth noting that Dell recently introduced its Dell Processor Acceleration Technology, an alternate approach to boosting the processing power of microprocessors, also for HFT applications.