The Low-Down

By Pete Harris, President, Americas and Editor-at-Large at A-Team Group

A-Team has been tracking the diverse collection of technology innovators playing in the low latency space. So where is low latency IT going?

I was recently asked for a definition of low latency, and as I struggled to articulate a concise answer, I realised that it’s a term that means different things to different people. The best I could offer is that it means delivering data or conducting a transaction in as near to zero time as possible.

What’s “possible” using current technology is usually expressed using an array of statistics, such as the ability to “process in excess of one million messages per second”, “fill orders in l0 milliseconds” or deliver data with “less than 150 microseconds’ latency” – which, while they may be pretty meaningless claims to a market data manager or a systems architect, at least provide a general idea of the ball park that low latency today plays in – namely milliseconds and microseconds.

It’s in this world of milliseconds and microseconds that an expanding number of liquidity providers, execution venues and institutional investors play. And so too does a diverse collection of technology innovators. It is these participants that A-Team has been tracking, through our 116-page report Faster Than A Speeding Bullet – Low Latency Architectures and Building Blocks For Tomorrow’s Trading Applications and our continuously updated portal at www.low-latency.com. By way of introduction, here’s the 101 on low latency IT and where it’s going.

Communications
It’s still a tad difficult to defy the laws of physics. Data can only travel so fast between two points, and that is still way below the theoretical speed of light. Fibre optics, Gigabit Ethernet and InfiniBand are all steps in the right direction, but the current “clever” solution to reducing communications latency is to remove the distance between data producer and consumer. Proximity services – allowing trading applications to be placed next door to trading venues – are now common tools, provided either by the liquidity venues, such as Deutsche Boerse and OMX, or third parties, like BT Radianz.

Hardware
It is no coincidence that Intel is a sponsor of our low latency report and, along with Reuters and Endace, our portal. The company that invented the microprocessor in 1971 has more recently (since about 1989, or thereabouts) been in the business of building chips for servers, the platforms that power low latency delivery software. Now Intel is locked in a battle with competitor Advanced Micro Devices (AMD), and the technology race has moved from developing faster chips to designs that incorporate multiple CPUs on a single silicon fabric, providing extremely efficient parallel processing – multithreading as it is often referred to – of computer software code.

Multi-core processors are the current leading edge in hardware deployment. To make effective use of multi-core, software needs to be designed (or rewritten) to utilise the power of the hardware platform. Put simply, the more a process lends itself to take advantage of performing multiple tasks in parallel, the more effective use of multi-core processing it can make. And the more a process requires that tasks are tackled in sequence, the less improvement multi-core will achieve. Right now, efficiently programming multi-core chips is where the software investment is being directed, and that’s going to be the case for some while.

If multi-core is leading edge, then co-processors such as Field Programmable Gate Arrays (FPGAs) are surely on the bleeding edge. The concept is to offload very specific processing tasks to very tightly coupled ancillary processors, boosting performance of operations like (in memory) database management and data field manipulation. If one can get the right balance in terms of which functions to offload (that’s the really hard part), the performance gains can be staggering.

Exegy and ACTIV Financial are using FPGAs in their datafeed handling products, while companies such as TS-Associates are leveraging them in latency monitoring products. It’s going to take some time for FPGAs to become widely used – for one, the skills to program them are rare outside of the US military – but if the requirement persists (especially when it comes to OPRA data rates), then FPGA-based solutions may be the only way to go.

Systems software
Operating systems and network drivers have been necessary evils since programmers forgot how to program in binary code, and read hardware manuals. The good news is that a lot of investment has gone into increasing efficiency of these vital components, whether they be from Microsoft or the various Linux camps. For sure, there is still work that can be done and perhaps once FPGA-class hardware becomes more commonplace there will be more of an impetus to tune this area of the software stack more finely.

Datafeed handlers and middleware
Moving away from consolidated feeds (at least as a primary source of tick data) and towards deploying integration platforms for direct datafeeds (from exchanges and other liquidity pools) is the current trend in minimising latency in data delivery. Thus, offerings in the form of Reuters Datafeed Direct, ACTIV Financial’s ActivFeed Direct and Interactive Data’s DirectPlus are likely to continue to be much in demand, whether located at client sites or at proximity service data centres.

That said, hosted services might well become attractive to some firms in the future, as more exchanges offer direct feeds, and their communications bandwidth hunger rises. For example, services that deliver to the customer just the data that interests them (eg for a selection of instruments from one or more exchanges) would solve a host of infrastructure issues.

As for middleware, whether it be in the form of message buses or data fabrics, the likelihood is that it will become increasingly commoditised, partly as a result of similar functionality being embedded into other components (such as event stream processing engines), and perhaps also as initiatives such as the open source Advanced Message Queue Protocol (AMQP) gain popularity (and also become embedded in Red Hat’s Linux, Cisco Routers et cetera). Most likely, middleware vendors will work more closely with, or merge with, providers of higher-value complex event processing/event stream processing engines.

Complex event processing
Life is going to get more complex for those in the CEP marketplace, and it needs to for those companies to survive in the long term. As the current players – such as Progress Apama, Kx Systems, Vhayu Technologies, Skyler Technology and Streambase Systems – all level out in terms of the basic functionality of their products (in terms of message rate handling, in-memory and on-disk persistence and performance of basic calculations) the challenge for them is to provide higher value offerings in terms of products and services. With the open source Esper project proving that geeks can traverse the value chain, the CEP vendors won’t want to join the middleware crowd in the commodity business.

Already, vendors such as Vhayu are broadening their asset support, to include options and fixed income calculations, while the likes of Streambase are addressing specific regulatory challenges, such as MiFID, and Skyler has incorporated order book handling into its architecture.

Put simply, the technology that drives low latency architectures will itself be driven by “better, faster, cheaper”. Along the way, we’d expect maybe a couple of casualties, some consolidation and a handful of real winners, but overall plenty of business to keep this challenging and bleeding edge niche in vogue at least as long as making money is. l

For more on low latency and for details of A-Team’s report Faster Than A Speeding Bullet – Low Latency Architectures and Building Blocks For Tomorrow’s Trading Applications, visit www.low­latency.com