Credit Suisse Underscores Need for Hardware Appliances in High-Performance Trading Infrastructures

Credit Suisse’s decision to implement Celoxica’s hardware appliance-based data feedhandlers for its next-generation trading infrastructure appears to be consistent with its belief that its best opportunity for reducing overall trading system latency lies in addressing the processing latency of execution engine and the market data systems that support it.Speaking at this week’s TradeTech Architecture conference in London, Gerald La Donne, chief architect of One Bank initiatives at Credit Suisse, said the bank sees the integration of trading logic with silicon as the future of low-latency trading capabilities. La Donne is responsible for front-to-back system architecture within Credit Suisse’s private and investment bank units. La Donne told the audience that the industry’s attempts to establish messaging infrastructures that respond to market events using software alone “is no longer enough.” The industry spent the 1980s and 1990s attempting to handle the trading lifecycle using distributed environments. Today, through the use of multi-core processors and field programmable gateway array (FPGA) hardware appliances, it’s attempting to consolidate processes onto a single box. Credit Suisse sees three types of latency involved in typical high performance trading infrastructures. The first is propogation latency, which is the time it takes to deliver data – typically market event data – across a network to the data feedhandler. The speed of this step is limited by the speed of light, and collocation or proximity services are aimed at addressing this type of latency, La Donne said, with each 20 kilometres of network adding 100 microseconds of latency. The second type of latency is transmission latency, which relates to the bandwidth or throughput capability of the network connection, and the speed at which data can be converted to packets for transmission. Here, the deployment of high-performance network equipment incorporating Infiniband or 10gigE is a common way of addressing latency issues. But it’s in the third type of latency – processing latency – where La Donne sees most potential for improvement. He reckons firms are already taking steps to achieve minimal latencies in both propagation and transmission latencies. “What you do with the data once you get it [is where] most of the latency is ... and hence where there is most opportunity” for improvement, he said. La Donne reckons between 100 and 200 lines of simple code can be processed in 1 microsecond. He said that by using multicore processing with acceleration techniques to parallelise key functions can result in faster executions. Furthermore, the use of FPGAs can allow data to circumvent potential bottlenecks like having to traverse market data feedhandler adapters, which can add between 20 and 100 microseconds of latency. “I can minimise latency using FPGA on market data, GPU accelerators and multicore processors,” he said. He said that commercially available products made such a set-up possible. He denied that the cost involved in implementing such a system would make it prohibitive to smaller players, pointing out that the cost of blade technologies, high-speed networks, FPGAs and GPU had fallen dramatically, and that open-source platforms like Linux can be optimised for the best results. La Donne also dismissed suggestions that inconsistency of execution venue latency, or jitter, would obviate any benefits of dealing with processing latency. External factors like matching engine or network jitter were essentially outside of the bank technologists’ control, he said. “You can’t do much about execution [speed] within a particular exchange,” he said. “It would be nice to know which venues are fast enough to give you a fill, then you can design your algorithm accordingly.” Stressing that firms that get the fill will make money, he said that in the future it would be “desirable to insert the trading strategy into the silicon.” While declining to address Credit Suisse’s own set-up directly, La Donne’s remarks appear to underline the bank’s approach to its next-generation high-performance trading architecture. The firm has selected Celoxica’s appliance-based data feed handlers as part of its initiative. Celoxica announced last November 30 that it had forged a strategic relationship with Credit Suisse wherein “Celoxica will provide hardware-accelerated trading and market data technology for Credit Suisse’s next-generation electronic trading platform.” As part of the arrangement, according to the statement, “The two firms have agreed to implement and develop both existing market data solutions and new state-of-the-art infrastructures to help address the growing need for high volume, deterministic trade processing capabilities as global markets continue to grow exponentially.” Celoxica had a month earlier released its first execution module, which uses a hardware-accelerated TCP/IP offload engine to boost the performance of clients’ execution platforms. It continues along its market data feed roadmap, with new hardware-accelerated feeds for Bats Exchange and its Bats Europe sidekick, and for Chi-X Europe, Celoxica’s first TCP/IP-based feed. At the same time, Celoxica extended its functional coverage to include normalized as well as raw market data format, multiple feeds on a single card, and recovery and support for full depth-of-book for its feed handlers.