Markets hate surprises, and the same applies for the time it takes to execute trades. While latency is often taken to be the critical measurement for high-frequency trading, there are many non-HFT traders who will sacrifice speed for consistency, the networking term for which is 'determinism'.
“On the less-latency-sensitive but latency-aware and latency-cognizant sides, what we see happening in the market is a focus on determinism rather than on latency,” Dave Snowdon, co-chief technology officer and founder of networking appliance manufacturer Metamako, told Markets Media. “What that means is that you have exactly the same latency through your entire system for each and every trade.”
Metamako, an 18-month-old Australian startup, is focused on building network device for financial applications. In this respect, it believes it has an edge over big players in the market like Cisco and Arista because it’s catering to a specific market.
“When you build an Arista switch or a Cisco switch, you're building for a market that generally includes cloud computing, general data center deployments, and office networks, and corporate networks, and all these other different use cases,” said Snowdon.
Computer designers provide a mechanism to improve the performance of common tasks, but leave infrequently executed or hard-to-implement tasks to the “slow path.”
This makes perfect sense when optimizing for bandwidth, but it destroys a system’s determinism because the worst-case delay via the slow path is now very different from the best-case delay via the fast path.
“What tends to happen with off-the-shelf devices is they aren't tailored for financial purposes, but instead they're optimized for bandwidth, which means that they'll try and optimize for the “average” use case, which in an exchange or in a trading situation is actually not needs to be optimized,” Snowdon said. “You optimize for the exceptional case, because that's when you're actually going to make trades, and you're going to make money.”
At the Australian Stock Exchange, for example, the fibers are all cut to 65 meters, which equates to about a 390-second latency, said Snowdon. “If your switches are down to 4 nanoseconds like ours are, the switching component is a small component of the latency of the Exchange. The response time from the Exchange is somewhere in the tens or even hundreds of microseconds.”
IEX’s 'speed bump' is an example of determinism. IEX has located its matching engine at the Savvis NJ2X data center in Weehawken, N.J., but requires all orders to pass through the Equinix NY5 center in Secaucus. It’s created a 350 microsecond travel time between the two data centers by coiling miles of fiber at each end. “Every order takes exactly the same amount of time to traverse the fiber,” said Snowdon.
Similarly, the size of the delay in sending orders to other venues is not important, but it is extremely important that the latency is consistent and known, i.e., deterministic, so that forwarded orders can be delivered simultaneously, avoiding exposure to arbitrage.
“There’s this focus on making sure you know exactly what latency is,” said Snowdon. “That applies to all sorts of institutions, particularly exchanges, who really should be ensuring fairness in the exchange by guaranteeing that the latency doesn't change for each and every order."
