Low Latency News

At a well-attended low-latency trading event today at Credit Suisse’s Flatiron district office, hosted by 29West, Wall Street executives noted that their focus on lowering data latency in their trading environments has not softened, despite the economic climate, and shared some of their latest efforts to reduce latency further.

For Credit Suisse, and especially for its CrossFinder dark pool, “Latency is our differentiator,” says Alex Roitgarts, director at the firm and the person responsible for latency. “Latency is a barrier to trading strategies. As a result, the key differentiator is how fast you can process and how much volume you can process. Logically, you don’t have to be zero latency, you just have to be faster than the other guy. Since you don’t know what the other guy is doing, you need to tune your operation like a Swiss watch every single day.” Even customers who don’t express any direct interest in latency are concerned about the performance of their trading algorithms. “The algo itself is becoming latency sensitive,” he says.

While two to three years ago latency was measured in the tens of milliseconds, today Credit Suisse’s round-trip trade order latency is within 300-350 microseconds. “I wouldn’t be surprised if within a few years people start measuring latency in nanoseconds,” Roitgarts says.

To reduce latency, the Swiss firm is installing new, faster network routers that will cause only three microseconds latency, versus 20 microseconds for the current technology. It closely monitors its trading applications by putting time stamps within trade messages and watching performance against certain thresholds. The firm doesn’t try to measure latency on every trade, though. “We try to take a top down, practical approach,” Roitgarts says. “If you try to measure latency on every signal that happens, the process of measurement may slow you down.” Instead Credit Suisse looks at logical check points. As soon as a lag is detected, “we immediately try to figure out why it happened,” he says. If nothing seems amiss with the application, the routers and network appliances are investigated.

At the National Stock Exchange, CIO Saro Jahani tries to take a holistic approach to latency. “We have to make sure that not only our matching engines, FIX engines and network are fast enough, good enough, and stable and controlled enough, we also have to make sure we have policies in place that help our clients to do smart colocations,” he says. “I have assigned a team of people within my organization to to make the system as low stress and low latency as possible.” Many times, customers’ latency is actually caused by their firewalls and workload balancers, he says.

Sean Gilman, CTO of foreign exchange ECN Currenex, points out that what his organization offers is a “fungible asset” — clients can trade wherever they want to. “When they come to us, they expect to get the price and market data first and to be able to hit that price first and faster than their competitors.” For instance, Currenex had a hedge fund client performing arbitrage based on trading models that frequently complained it was losing money because it was putting in orders that weren’t getting filled. “We found that another customer was hitting that same price, each time. The models are the same between these different hedge funds, but one was faster by one millisecond or a few microseconds. It doesn’t matter what the granularity is, you’re fastest or you lose. It really is a race.”

To keep up, Currenex upgrades its core servers every 16 months and its network every two years. The exchange also separates simple orders from complex ones, which are handled on a separate matching engine, allowing the common orders fast throughput.

For latency monitoring, Currenex finds SNMP [Simple Network Management Protocol] useful for basic devices such as routers and switches and boxes. “But the real problems tend not to be there,” Gilman notes. “That’s not usually the type of thing that bites us.” Instead, the forex venue focuses on applicational latencies between different services, storing statistics and charting trends to detect possible causes of latency, such as a code change.