Anthony Malakian: Getting the Jitters

How do you solve for something that doesn’t have an exact definition? Well, that’s not totally true—jitter has a definition. According to the unimpeachable Wikipedia, jitter is the “deviation from true periodicity of an assumed periodic signal in electronics and telecommunications, often in relation to a reference clock source.” There you go!

But it’s really not so simple. Chicago—where Waters’ sibling publication Inside Market Data recently hosted a conference—is home to large numbers of prop-trading shops, each of which is trying to figure out how best to manage latency.

At the conference, I spoke to several industry participants for a story focusing on how firms are now just as concerned with the measuring, monitoring and consistency of their latency across the enterprise as they are with the so-called race to zero (see page 30).

When it came to jitter, though, my sources were much more reluctant to talk. Some cited the lack of a clear definition, while others said that it would give away too much in terms of their trade secrets.

Problems
During a panel discussion in Chicago Stefan Gratzl, performance engineer at CMT Asset Management, described jitter problems as being an issue among network switch providers, whose solutions are built for the general internet, but not for low-latency trading. “I buy a $15,000 switch in my environment where I access seven destinations in Layer 3 routing, but I have the capability of holding the entire world’s internet routing table—it’s like taking the family car to a Formula One race. If I want to win a car race, I’m not going to bring the kid’s seat with me.”

Since messages arrive in bursts and not in a neatly sequenced order, some messages must be put in a queue before they can be processed.

The solution, he says, is to work in partnership with third parties and even other trading firms. In order to control jitter, firms need to work together and find partnerships, he says. “The big firms are doing this. They’re building their own devices and looking at the data and the future of this. Since it’s mostly commoditized, partnerships need to happen,” he says.

Of course vendors can help—it’s their job. But other firms? I’m not so sure. For the sake of consistency, I’m defining jitter as the variability in the latency experienced by successive messages, or trades, in a sequence of messages. Since messages arrive in bursts and not in a neatly sequenced order, some messages must be put in a queue before they can be processed. If that’s a fair definition, then jitter will inherently be a problem, and, as such, is not curable. Thus, reliable workarounds—or re-routes—are necessary, according to Donal O’Sullivan, vice president of product management at Corvil.

“The simple fact is that you cannot control jitter—you can only change the location where the queuing occurs,” he says. “There are supposedly deterministic network protocols like Infiniband, and these promise better jitter characteristics than ethernet/IP, for example. However, the way they achieve this is to minimize the queuing available in the network, and make the client and server computers perform that queuing instead. So, yes, network jitter is reduced, but the overall jitter experienced end-to-end by the trading application, for example, is just the same as before.”

Monitoring
Henry Young, CEO of TS-Associates, adds that while jitter is an unavoidable evil, monitoring technologies can help firms manage it.

“Many factors cause jitter, but the key tool for determining the cause is effective monitoring,” he says. “Once the cause has been identified, then various technology solutions can be deployed.”

Those technologies include cut-through switches and kernel bypass network interface cards, but mainly it’s about ensuring that the network engineers are close to the systems and have freedom to build the systems as they deem fit.

“The art and science of low-latency application engineering has become extensive,” Young says. “It often involves turning our backs on years of abstraction in computer science and getting closer to the machine again.”