Printer Friendly
"Yes, it's not for everybody and it's hard (to program) and in some cases it doesn't make sense," said Adam Honore, senior analyst at Aite Group, who covers emerging technologies. "Hardware acceleration is where CEP was one or two years ago," says Honore, referring to complex event processing.
"The value proposition is not just to sustain speed at peak but also a reduction in rack space at the data center," he explained. Depending on the specific application, a hardware appliance can reduce the amount of rack space by 10-to-1 or 20-to-1in certain market data and some options events, says the analyst. In order to gain traction in the financial industry, hardware acceleration vendors need to improve their platforms, said Honore.While FPGAs and GPUs are hot buzzwords in the industry, low-latency practitioners have qualms about getting involved with technologies that are proprietary. "I look at it mainly as a last resort, because hardware accelerators bolts into proprietary hardware," said Albert Doolittle, VP-IS at George Weiss Associates, a privately-held hedge fund sponsor as well as broker located in Hartford, Conn. and New York.
While Doolittle is still exploring the hardware acceleration space, he said the programming is different. Outsourcing to a vendor who does the hardware programming and runs the hardware acceleration appliances is an alternative to running the GPUs internally. One way to reduce latency is to collocate applications in a data center near the trading venues and use a vendor who accelerates options analytics and have it "ship it to me," suggested Doolittle. In general, experts are working with not one, but a whole host of technologies to speed up their analytics, data feeds and communications between systems. Conor Allen, VP and director of research and development at NYSE Technologies, the company that bought Wombat Financial Software, a low-latency data feed integrator and middleware provider, said it found hardware acceleration "interesting and intriguing. We definitely have been looking at it for while," said Allen. Noting that the NYSE Technologies supports 160 different feed handlers around the world, Allen said, "It's my judgment we don't need hardware accelerators for that." However, in places where it makes sense, we won't shy away from it," he said.
In terms of what's in a broker or money manager's low-latency toolbox, Doolittle said he is developing the next generation trading system with complex event processing (CEP). The CEP is bolted up to a low-latency feed and Doolittle said he is looking at the execution side of it. "Now I'm trying to find the balance between speed, simplicity and taking as many components out using commoditized information and keeping it small," said Dolittle.
Allen said he was focused on low-latency middleware and finding ways to move data between places. According to Conor, "High performance middleware, something that bypasses the IP stack," is important. In the high frequency trading worldyou want to run a lot of stuff as fast as possible," advised Allen. "If you can bypass the operating systems and bypass the IP stack, you can get blistering performance that is not associated with the operating system."
Looking ahead, low latency experts are eyeing advances in multi-core processing. "Two or three years ago, it was unusual to have two cores, noted Allen. Now everyone has eight. Next year, thirty-two-core systems will be available to you," said Allen.
Doolittle emphasized the need for latency monitoring tools, adding that he wanted to get a handle on latency measurement. "I'm not comfortable with having these waves or jitters and not having a way to attack that in a system that is moving very fast," said Doolittle. Troubleshooting is also an important part of the toolbox. "You have to be able to go back forensically and determine what was the trouble with it," said Allen.
