-
-
-
http://www.autobahn.db.comYou need to upgrade your Flash Player
-
- REGISTER Partial Site Access - Digital Editions - News Feeds
- SUBSCRIBE Full Site Access - Printed Magazine - PDF/Digital Edtions
As growing competition places ever greater emphasis on the fine-tuning of automated and algorithmic trading programmes, AT asks leading solutions providers for their views on the tweaks that can minimise application latency and maximise performance. With: - Ary Khatchikian, president and chief technology officer, Portware - Amir Prescher, vice president, business development, Voltaire - Matt Meinel, global director, business development, 29West - Gena Ioffe, CEO, EGAR Technology - Ali Pichvai, managing director, Quod Financial - Jonas Lindqvist, head of system development, NeoNet - Yuriy Shterk, vice president, product development, CQG - Nimrod Gindi, business development manager, Mellanox Technologies
Application Latency
Application vs. network latency: which is more important in improving algorithmic /automated performance?
Khatchikian: Improving application latency is more important without question. There is always a limit to what you can optimise regarding networks. With applications it’s more complex because they are always changing. A few lines of bad code can bring an algorithm to a crawl if you’re not paying attention. In some cases, applications have been optimised on the algo level to overcome network latencies where cost was a serious issue.
Prescher: To achieve overall endto- end latency reduction, both application latency and network latency must be addressed. Application latency has a few common causes. Code ‘bloat’ is one, but we find this to be the cause less often than most people would think. Trading firms do a lot of due diligence on their algorithmic trading applications and optimise those applications as much as possible.
The majority of the time, we see electronic trading infrastructures running out of CPU cycles (even though they are running on the latest and greatest hardware). High market volumes and the resulting high packet rates are to blame for a large part of the high CPU utilisation problem. InfiniBandbased solutions can help because they offer network processing offload technology which helps to significantly increase packet per second rates, as well as lowering the CPU utilisation that occurs when processing the high packet rates. This is very different from traditional TCP offload engines (TOEs). Since market data feed packets and general FIX packets tend to be between 60 and 200 bytes, traditional TOEs are not useful. InfiniBand-based solutions can deliver latency that is 20-500 per cent lower (depending on the applications that are being used) than what is achievable with Ethernet-based solutions. ...