We’ve been tracking the developments in the high frequency trading technology space since the launch of Low-Latency.com in 2006 or thereabouts. Now retired and living near Heathrow – or rather, reincarnated as Intelligent Trading Technology – she watches with interest as the superfast aspect of our marketplace finally approaches the endgame, where physics finally says “Enough!”
Over the past decade or so, we’ve been plotting the gradual migration of trading logic toward the execution venue’s matching engine. We started with fast networks – with efforts still continuing with microwave – which brought our trading algorithms closer in terms of messaging times. We then shifted our execution systems to colocation and proximity hosting facilities, bringing them closer physically.
Once happily ensconced just metres of fibre from the matching systems we upped the compute power of our execution servers – putting in 10 servers where once there was one, sharing the messaging load via local switch, the better to optimise fast access to the marketplace.
Finally, we transferred those cumbersome feedhandlers – prone to spilling so many microseconds – from slow slow software to fast fast FPGA. The hardware approach minimised the delay between receiving market updates and sending buy/sell orders to market.
It’s been a long journey, and today only the most stoic remain. FPGA is a difficult and expensive technology to master. FPGA experts are difficult to find and difficult to evaluate even if you can find them. The industry seems to have reached the conclusion that only those who know what they are doing should dare to dabble. A multimillion dollar investment could bring a lot of heartache, and no guarantee of a return.
But FPGA is not the end of the road for ultra-low-latency market connectivity for high rollin’ HFTers. The journey is not over.
In 2012, Rob Walker and his team launched xCelor, a company specialising in FPGA that is seeking to democratise what heretofore has been a technology beyond the budgets of all but the most well-endowed HFT shops. xCelor introduced a rental model for FPGA access, and since its launch has helped many an HFT shop adopt FPGA. But it hasn’t all been about speed. Walker reckons a key attribute of FPGA is to address determinism, what he calls the No. 1 challenge facing those building high-speed trading infrastructures.
Determinism refers to the need for consistency of access. A high-speed connectivity set-up may offer access measured in nanoseconds in laboratory conditions. But at the market open, nanos turn to tens of microseconds and worse. Trading algorithms can’t be sure of what access delay to expect. The result is that their performance can vary wildly from what was expected in lab tests.
Certainly, determinism will feature strongly at our Intelligent Trading Summit in London in November (just thought I’d slip that in; date to be determined very shortly). Walker reckons the performance of many high-speed systems can be so inconsistent between market open and lunchtime as to resemble the difference between a 30-minute ride into town and a commute during peak hours. Night and day.
Which brings us to the endgame. FPGA is as fast as we’ll get, and it’s also good for determinism as it precludes messages from queueing up ahead of the application. But it remains expensive and somewhat difficult to manage. The release of the now-defunct Arista 7124FX switch a few years ago opened new vistas by allowing practitioners to add logic to the switch at the colo site.
This, combined with FPGA feedhandlers, allowed simple trading algorithms – pre-trade risk-checking for example – into the hardware environment, reducing response times while adding some smarts. It was frontier-town stuff; simple logic running at superfast speeds.
Fast-forward two years and xCelor has reimagined the FPGA switch. The result – the xCelor XPM² – allows firms to develop trading algorithms in C++ that run on a regular x86 processor inside the switch, with FPGA logic talking directly to the exchange handoff. It is possible to program algorithms to take advantage of the speed with deterministic consistency.
Walker points to clients using this kind of set-up to handle feeds from Nasdaq, BATS, CME and others. Maybe this is the end of the road, finally. Trading logic has moved as close to the matching engine as it’s going to get. Determinism is achieved even in superfast trading environments. Intelligent trading in fast markets.
Is this the low-latency endgame, or is there more to come?
Join us in November to move the discussion forward. Stay tuned for more information soon.