GPU Computing를 이용한 트레이딩시스템

1.
한 달전쯤 홍콩의 선물옵션트레이더와 통화할 일이 있었습니다. 옵션전략을 개발하는데 두가지 기술을 구한다는내용입니다. 첫째는 GPU기술입니다. Option Pricing을 GPU를 이용하여 하겠다고 합니다. 독일 사례를 이야기하면서 GPU개발을 하겠다고 하더군요. GPU를 이용한 Pricing은 GPU 예제에 나와 있지만 국내에서 이를 업무에 활용한 사례를 찾아보지 못했습니다.

독일사례란 오픈소스 CEP로 유명한 MarketCetera를 기반으로 하여 독일 Lakeview가 GPU기술을 이용한 옵션트레이딩시스템을 개발한 것을 말합니다.

Using these performance options, common calculations such as theoretical prices, implied volatilities, implied dividends and correlations as well as risk simulations, amongst others, can be done on NVIDIA GPUs, offering speed ups of 10-200x, depending on calculation, in comparison to current commercial trading systems that run calculations on traditional hardware.

회사의 설립자인 Peter Van Cleef는 High Frequency Trader Review와 가진 인터뷰에서 FPGA와 GPU 및 거래소기술에 대해 재미난 이야기를 하였습니다.

HFTR: OK, so continuing on the technology theme but drilling down a little, you guys are doing some interesting work around FPGA and NVIDIA GPU accelerators. Can you tell us a little bit about what you’re doing in that space

Peter: Basically you have two technologies there that are used for slightly different things. GPUs usually provide scalability in terms of calculations that you can do in parallel. For example option pricing calculations, some implied calculations for correlation dividends, things like that. Any calculations that you can parallelize can give you a huge advantage. You gain massive savings in terms of electric power consumtion; computing power and heat emission. Calculating risk scenarios that might take hours in some cases can be brought down to minutes or even fractional seconds, because in the complex risk analysis you can paralyze calculations. So that’s one feature where GPUs are used. But then GPUs, at the moment, you still don’t have perfect integration with the rest of the trading infrastructure so you have some overhead in terms of sending the data from the normal CPU, from the normal bus in your computer, to the GPU, and getting it back. So there’s some latency involved with that. For the moment you can’t buy the chips alone and build your own system, but I am sure that will come.

GPU를 이용하면 좋은 점도 있지만 CPU와 GPU간의 통신때문에 지연이 발생한다는 점을 강조합니다.

HFTR: And FPGA’s

Peter: FPGA basically can give you ultra fast response time, way below millisecond response time on the market. You would use those with pre-calculated values from the GPUs uploaded to the FPGA, pre-trade, and then when the market update comes, the FPGA card basically already has a memory of what the response will be and can send out that response immediately without having to go through the bus of the computer to the CPU to the storage to the memory and back, basically doing some calculation on the fly. So you use one technology for one part and the other technology for the other parts. You use FPGAs for feed handlers or for reacting to the market with electronic eyes and trading engines and stuff like that. And you would use GPUs to pre-calculate the values that maybe the FPGAs are loaded with.

HFTR: So do you see exchanges moving towards using FPGAs for matching engines, for example

Peter: Well for me it seems ridiculous that the exchanges are still rolling out new exchange systems in software. All of them are competing and saying they want to be the fastest, so they should be serious about speed. If they want to be serious about speed they would have to do it in hardware of course. For the moment none of the exchanges (as far as I’m aware) is even planning that far.
미래에 FPGA를 이용한 매매체결시스템이 가능하겠지만 현재 이를 시도하는 곳은 없고 아직까지는 소프트웨어의 영역이라고 진단합니다.

Interview with Peter Van Kleef, Lakeview Arbitrage International중에서

2.
최근 ZeroM의 성능때문에 GPU를 이용하는 방안을 검토한 적이 있습니다. 당장은 아니더라도 GPU와 CPU를 통합한 제품이 보편화되면 고민해볼 수 있지 않을까 생각합니다. 물론 개발자의 의견입니다.(^^)

가장 GPU Computing을 보여주는 자료입니다. 참고하세요.