Scientists Unveil Low-Cost Latency Monitoring Feature For Routers
August 24, 2009
Scientists from the University of California, San Diego and Purdue University said they have developed a cost-efficient mechanism for measuring latency down to tens of microseconds and losses as infrequent as one in a million, according to a research paper presented at the SIGCOMM networking conference in Barcelona last week.
The feature, dubbed the Lossy Difference Aggregator, would be incorporated within routers and provide an alternative to external latency monitoring boxes currently installed by technology vendors, according to the research team.
“The router providers who would embed this are Cisco, Foundry Networks, among others—anyone who supplies routers to the data centers, investment bank and traders,” said George Varghese, computer science professor at the University of California San Diego Jacobs School of Engineering, one of the paper’s four co-authors.
Kevin McPartland, senior analyst with Tabb Group, questions whether the proposed feature would "get quite to the granular level that the ultra-low latency trader would require."
“It’s interesting from an academic point of view, but it depends on exactly who we are talking about,” he continued. “In the financial services, ultra-low latency world, you need tools that are really geared towards and know how to handle the volume of messages and latency we are talking about. It is an interesting academic finding but not yet something that has any potential to really change the landscape of financial services just yet.”
Donal Byrne, CEO of Corvil, a provider of latency monitoring and managing software, said Corvil “continuously monitors the research work that happens in the academic institutions and continues to look for novel, breakthrough ideas.”
However, his firm would not utilize the proposed technology nor would he imagine that any other serious player in high-frequency trading would use it for various reasons, he noted.
“[With this proposed solution,] I can’t tell you what the peak latency is—and of course, that is the most important thing in high-frequency trading,” he said. “You need to be able to detect and measure spikes—this technique can’t do it. Second, it doesn’t overcome the issue of how you measure latency between or multiple points because you have to solve the whole clock synchronization problem, which this does not address in any way.”
In response, Varghese said he believes when traders consider the tradeoffs--box versus no cost—“they may be very happy to settle for a delay that can be made arbitrarily close to the maximum [peak], latency,” Varghese said noted that his team’s proposed feature measures standard deviation. "By taking the Average plus a small number of standard deviations, we can come very close to measuring the maximum ever seen by 99.99% of packets [messages]."
Regarding clock synchronization, most of the delays are within the routers, said Varghese. “The latest routers (e.g., Cisco's DC-2) have solved the clock synchronization problem within routers by adding a hardware bus,” he said.
