 |
|
Focus on Latency |
|
Global Investment Technology
03/17/2008 |
|
'Predictable Reliability' Becomes the New IT Mantra For Advanced Trading as Milliseconds Matter
NEW YORK — After recent regulatory stimuli favoring best execution of trades, volumes of market data that have to be analyzed are rising sharply. At the same time, new challenges for reducing latency in trading have pushed market participants to co-locate their servers at or as close as possible to exchanges, and to develop new approaches to increase trading speed.
"It's not as much the trading styles as the amount of quote data being generated because of the regulation," says John Coulter, Vice President of Marketing at VhaYu Technologies Corporation, a financial market data capture and analysis provider. "The order sizes are coming down and there are more points of liquidity now, and all sorts of new exchanges and alternative trading facilities popping up. Traders need to identify opportunities but also at the same time, analyze quotes coming in to make trading decisions."
While the greater availability of real-time market data may not necessarily always create more latency challenges, limitations of operational networks or throughput capacity in processing real-time data may be the cause of problems and slowdowns, according to data service providers.
"Right now, you could suddenly have a peak burst of 300,000 messages per second on an active day," says Coulter. "If something happens or there's a crisis, anything that's going to impact the global economy, in the next year or two that message rate could get to 400,000 to 500,000 messages per second. Most systems can't throttle that much information. Their applications are usually not able to handle that much data. As a result, a lot of the applications start falling over when you get to those rates."
An event similar to this occurred in February 2007, notes Coulter, when a decline in China's market set off a sudden sell-off in the US. "This easily could have been avoided if the proper complex event processing technology infrastructure were in place," he says. "The Dow Jones Industrial Average lagged the market for an hour as NYSE servers were deluged by the record 2.41 billion shares that changed hands that day. This was a warning flare to financial IT managers to incorporate systems capable of capturing and analyzing millions of messages per second, with the ability to instantly spot anomalous events and take appropriate action."
Increasing the capacity of applications to handle messaging and other trading functions is crucial to lowering latency. "Real-time market data requires a tremendous amount of bandwidth, which typically saturates the throughput capabilities of traditional Ethernet-based network infrastructure," says Patrick Guay, Senior Vice President of Marketing at Voltaire Inc., a provider of data network infrastructure serving financial services and other industries. "It's the limitations of the network or the fabric that creates the latency. Reducing transaction latency may be the most important objective in the financial markets today."
Data throughput is closely related to data latency, according to Michael Lenahan, Senior Vice President at Skyler Technology, a trading systems provider. "Throughput is the amount of data that can move through the system in a given unit of time, without falling behind, and latency is the time from when a system takes in a data record, processes it or operates on it, and then outputs a message," says Lenahan. "We are seeing market data volumes increasing and peak rates spiking, causing exchanges trouble in handling the market data loads. For the exchanges, this translates into delays in the feeds they publish. We have also seen problems on the buy side, where some systems are not able to handle market data spikes and fall behind."
Latency can also mean the difference between winning in a trade or not, according to Mark Mahowald, Chief Executive Officer of 29West Inc., a messaging services provider. "If you have a clever algorithm, it's only as clever as the next one," he says. "If I get the data sooner, I'll win the trade and beat the other guy to the punch. So there's a real advantage in latency."
Increasing the speed of networks and data processing is key to reducing latency for real-time data, observes Bill Fathers, Senior Vice President of Development and Engineering at Sawis Inc., a trading solutions provider for the financial industry. "Latency is a characteristic of how real-time market data is delivered," he says. "Many companies are investing huge amounts in reducing the latency of the applications that process the data by optimizing the applications and underlying database technology, but also by investing in very fast underlying computing platforms."
The speed of transmission between systems including networks, switches, algorithmic trading and complex event processing (CEP) systems will continually be under pressure to decrease, according to Coulter of VhaYu. "There's definitely room to compress things that used to take seconds down into milliseconds," he says. "Eventually people will try to get it down to nanoseconds. It's going to be a never-ending arms race to go faster and faster."
CEP systems [see Global Investment Technology, January 7, 20081 can play a role in reducing latency, observes Ross Hamilton, Director of Client Engagements at Lab 49 Inc., a financial technology applications provider. "Complex event processing provides the means to consume multiple streams of real-time data and execute meaningful operations on the data," says Hamilton.
"This includes filtering, cleansing and normalization, as well as pre-trade analytics. This can take the load off downstream systems. If these high-performance infrastructures are not in place or not designed well, then that high-volume, high-frequency data can paralyze a system, leading to increased latency."
Real-time data highlights where latency bottlenecks exist in a trading environment, explains Hamilton. "As latency bottlenecks are removed from the systems that consume and produce real-time data, latency becomes increasingly sensitive to what people performing functions in the trading room do," he says. "If large orders, for example, are routed to traders for execution, then clearly they become the major individual contributor to latency. Automating these functions could help, but that requires significant investment in risk and control systems."
With certain algorithmic trading strategies, it is impossible to react to all ticks in market data, so stream processing becomes important to reducing latency, according to Gideon Low, Principal Architect, Business Development and Alliances at GemStone Systems, an enterprise software company that recently partnered with several CEP vendors to analyze and distribute transaction data. "Stream processing vendors are able to look for patterns in very fast-moving market data and only inform the trading system when a specific trend or complex event appears," says Low.
The biggest challenge for firms may be migrating to new low latency architectures. "Designing software so it is highly 'parallel-izable' is a challenge," says Low. "It's necessary to achieve very low latency in large-scale environments, regardless of the technology chosen to reach that goal, because 'parallel-iza-tion1 is an absolute requirement."
Aside from simply co-locating at or near exchanges, investment firms are turning to networks and "parallel-ization," as an answer for latency. "There is a new ecosystem looking to challenge the existing way of doing things and solving these problems a different way," says Daniel Chait, Managing Director, Lab 49 Inc. "That is building loosely coupled networks of low-cost commodity hardware to process computation in a distributed fashion, rather than just buying a bigger, faster supercomputer every year or two."
The end result of lowering latency should be "predictable reliability," according to Barry Thompson, Chief Technology Officer of Tervela, an enterprise data messaging provider. "People are firing more and more orders, but they're behind the market," he says. "So those who can't have low latency in those periods of peak volatility are the ones who can lose an entire year's profit in 20 or 30 minutes in high market volatility if they don't have predictable reliability. Those who can get predictable reliability are far more profitable than those who just have low latency."
New approaches to reducing latency are countering a more complex and faster market data and messaging environment with greater complexity and speed of their own, including parallel systems or systems that marry monitoring and analysis of market data and market events. In fact, the drive for low latency could eventually focus on lowering latency in peak trading periods where latency can cause the most potential loss.
|
|
| |
|
|