Chi-X Chooses IBM for High-Speed Messaging
The announcement today that Chi-Tech, the technology arm of Chi-X, the pan-European multilateral trading facility (an alternative exchange for equities), has signed up to use IBM's WebSphere MQ Low Latency Messaging prompted us to take a closer look at this messaging system, which according to IBM can process 90 million messages per second, even on x86 servers, and at the question of whether software-based message handling is better than hardware-based.
Most of the high-speed messaging vendors we speak with (such as Tervela and Solace) say that market and order messages have to run on hardware, otherwise it's impossible to achieve ultra high speeds or to have control over latency. Earlier this year, Larry Neumann, senior vice president of marketing at Solace told us, "With software, you end up switching control between operating systems, the applications, the network card, and so on. Every time you make one of those switches, you chew up some microseconds. When you try to do that hundreds of thousands or millions of time per second, it creates too much overhead. By baking all that into chips, all those issues go away. Data comes straight off the wire, there's no protocol stack, no operating system, even the data processing is all performed on the same chip or set of chips. As a result, we can reach higher rates of throughput with predictable latency."
However, although Neumann says Solace's routers are ten to 100 times faster than comparable software products, IBM's offering is software only and according to its tests, runs almost ten times as fast as Solace's promised rate of 10 million messages per second.
"We're faster than any of the hardware offerings out there," says Paul Michaud, global financial markets industry architect at IBM. "When you look at Solace and Tervela, they have to do at least one additional hop, it's not possible for them to accelerate the messaging at the two end points, sending applications or receiving applications, so Solace and Tervela are fast if you're comparing them to a daemon like Tibco uses. Most of the really fast messaging is daemon-less in its operation, so we completely eliminate that need to have any daemon in the middle and it's much, much faster than what they can do. They're at least twice our latency, at best. None of the exchanges use their appliances."
A daemon (there's a conveniently evil-sounding word) is a a computer program that runs in the background, rather than under the direct control of a user; daemons are usually initiated as background processes. This reminds us of a comment made at the High Performance Computing on Wall Street show earlier this week by Jarod Jenson, chief technology architect, Sun Solutions, Forsythe; he said that artificial pauses are built into many components of an ostensibly high-performant architecture, including in applications, network interface cards and operating systems.
IBM has three exchanges that have publicly said they're using its messaging system: Santiago Stock Exchange, Deutsche Borse (which will roll the IBM system out to three of its exchanges) and Chi-X, which will market the platform to other exchanges. "Suffice it to say there are others" using the system, Michaud says, that have not gone public.
This product was purpose-built for exchanges and financial transaction use cases, Michaud says. "It's not just fast, it solves a number of other problems for them, such as assuring high availability and never losing a message. We get specs that say they need six 9s of uptime. We run a hot/hot model for matching engines, ensuring messages have right the ordering, that failover, state synchronization, cold standby, are all handled by the messaging itself. That aspect is really attractive to the exchanges."
Michaud says he has 185 people with PhDs in Haifa, Israel who focus on developing advanced messaging technology. "At the end of the day, it's having all these really smart people spending time working on this," he says. The way the system handles packetization and it's algorithms for constantly monitoring the system for dropped messages are key to its speed and reliability, he says.
But as a software-only solution, is IBM's solution not at the mercy of underlying servers, network cards and other hardware components, as Neumann suggested? "If you consider a hardware-based daemon solution like Solace or Tervela, if I'm running a trading system or matching system on a gateway, that gateway is still doing all the things it would have otherwise done. The only thing hardware can control is the messaging daemon that is the appliance, it can't do anything about the application environment or the hardware that's working at either end. So we get much higher throughput and scale than any other messaging solution, including 29West, which quotes a rate of 2.8 million messages per second."
IBM's offering runs on an array of hardware platform, including Windows, Linux, Sun Sparc, IBM p series, HP-UX Itanium and Stratus VOS systems. It's achieved the 90 million messages per second rate on x86 servers (IBM BladeCenter HS22s) using Mellanox Infiniband cards with a Voltaire switch. "What limits us at 90 million is simply the fact that the Nehalem processors can't handle any more messages and we end up with a CPU bottleneck," he says.
Tweet This