Software, hardware solutions rush to fill demand

Simon Garland chuckled as he examined charts showing the spikes in quote volumes on the New York Stock Exchange that began in August. “It was like Christmas,” said the CTO of Kx Systems, a Palo Alto, Calif.-based provider of high-performance, in-memory database technology and streaming event processing. “It was magic.”

But not for everyone. Garland says some firms lost ticks because their systems couldn’t handle all the data coming from the exchanges. Systems didn’t crash, but at the end of a day with 475 million quotes, 470 million might be in the database. “That’s okay for a historical database,” he notes, “but if you want to trade through the storm, the last thing you want is to lose ticks just when it is getting interesting.”

Speed has long been an invaluable asset in capital markets, but the growth of algorithmic trading and new best-execution requirements have created hefty challenges for market data processing and analysis.

Larry Tabb, founder of New Yorkbased Tabb Group, cautions that the low-latency arms race is limited in scope. He says there are maybe 100 firms worldwide that can, or will, pay for the best technology.

As Garland points out, those with the most money can buy the best systems. Not only can the top firms store and use weeks of historical data in real-time, pre-trade processing, they know how much analysis can be done within the processing windows of the firms in the next tier down, giving them an advantage in trading against the competition. A similar tactic exists in aerial combat: By operating inside the enemies’ OODA loop—short for observation, orientation, decision, action—a fighting unit maintains the initiative and keeps its opponents off-balance.

Tom Price, senior analyst for capital markets at Needham, Mass.-based TowerGroup, says the need for the latest technology will divide Wall Street firms into the haves and have-nots. “If you don’t spend, you will get left behind,” says Price. “You will lose out in the hunt for liquidity, you will not get to the pools that are fragmented across the spectrum, and you will constantly be eating leftovers.”

He adds: “Not only do you need content—market data infrastructure with speed—but you also need analytics and strategies that allow you to look at light and dark liquidity and generate internal organic liquidity. If you don’t have the horsepower, you can’t be in electronic trading.”

Hardware vs. Software
Vendors are responding with both hardware- and software-based solutions, usually touting their company’s as the best possible approach.

IBM Corp. in November introduced version 2.0 of its WebSphere MQ Low Latency Messaging software, which can support the processing of millions of messages per second with microsecond latency, according to Chae An, VP of financial sector industry solutions in the company’s software group.

Reuters is using WebSphere to extend the multicasting capabilities of its Reuters Market Data System (RMDS). “The performance and interoperability of WebSphere MQ Low Latency Messaging will give our clients even more messaging choices for RMDS 6.0 at greater speed than ever,” said Peter Moss, global head of enterprise solutions at Reuters, in a statement. “For the first time, developers will be able to build new applications using Reuters Foundation API on top of IBM messaging software to access content from RMDS 6 and to distribute it into other parts of their organizations.”

IBM says that WebSphere has demonstrated very high throughput, one-tomany multicast messaging, which can deliver approximately 1 million 120-byte messages per second on Ethernet, close to 3 million on InfiniBand, and more than 8 million smaller messages per second, all on common x86 servers.

New York-based Tervela, a venture capital-backed company that counts Goldman Sachs among its investors, is one of several start-up hardware providers that believe that software solutions have reached their limits. Tervela claims it has developed the first hardware-based messaging platform, which it describes as a hybrid of networking and middleware technology that can integrate and accelerate the entire messaging ecosystem.

J. Barry Thompson, Tervela’s founder and CTO, compares market data systems to routers, which were software-based in the 1980s before moving to hardware as they expanded from work groups to the enterprise. He says that hardware can provide a 25-time greater reduction in latency than the best software solutions, a claim that is contested by companies such as New York-based Wombat Financial Software. Tervela has tested 20 million events per second—it says the hardware costs to test beyond that would strain the company. But Thompson is confident that it can scale infinitely.

The volume of Options Price Reporting Authority (Opra) messages “will reach 900,000 or 1 million per second next year,” says Thompson. “That data comes in and gets distributed through a feed handler. Some applications eat it and republish the data, so you are at 2 million; then you cache and redistribute and you can get to 3 or 4 million. You start marching through a million events per second and the numbers cascade as each message comes in and flows throughout the infrastructure. Without a message switch, your systems become overwhelmed.”

The Tervela appliance “can take all of Opra into one box and distribute the data to 500 concurrent users,” explains Thompson. By contrast, InfiniBand, which offers highspeed connectivity, is usually limited to communications between a few points in an enterprise. “InfiniBand is fine for one end point, but if you want to talk to four end points it breaks down,” Thompson says.

TowerGroup’s Price notes that banks may be hesitant to add another piece of hardware to their data centers. But Thompson says that Tervela can replace 500 servers for order routing and provide tremendous acceleration for FIX. “We do the middleware and a bank’s applications run on the fabric and they can radically reduce their box count,” he says. Because Tervela provides the single source for data, middleware and applications that run on its fabric, users can also monitor operations easily.

“Today, if an order takes a minute to get to the exchange it could be middleware, the application or the network, and as a result you get a lot of finger-pointing,” says Thompson. “Tervela gives complete visibility. You can see a machine in Tokyo driving a stream that is causing problems in London or a machine in Tokyo dropping data that London needs, and you can roll back the clock while the system is running to see what happened.”

Tervela says it has ten customers, some of them top-tier financial firms, in varying stages of production with its hardware.

Data Centers
Find Solace Ottawa-based networking equipment provider Solace Systems says its distributed messaging and application infrastructure boosts persistent messaging performance to unprecedented levels. Solace, whose founders include engineers from Nortel Networks, Cisco Systems and Alcatel, has applications share a single infrastructure that resides on the network.

“That’s a sharp contrast to today when you buy one specialized kind of middleware for low latency or algorithmic trading and something different for guaranteed delivery,” says Solace SVP of marketing Larry Neumann. Solace employs field-programmable gate array (FPGA) co-processors that, unlike all-purpose processors from Intel Corp. and Advanced Micro Devices (AMD), are programmed to perform a specific task. Solace engineers code the FPGAs’ logic gates for messaging and persistent messaging.

“Disks haven’t gotten any faster, so persistent messaging that relies on disks hasn’t gotten faster,” says Neumann. “We eliminate the disk but guarantee delivery so we can do 70,000 messages per second.”

Like Tervela, Solace offers a way to reduce the boxes in a data center; one of its devices can replace 15 to 30 middleware distribution servers. The hardware solution also provides consistently low latency, adds Neumann.

“That is hugely important for algorithmic traders,” he says. “If you usually have 1 millisecond latency and that spikes to 10 milliseconds because the operating system gets busy, it can put the whole trade in jeopardy. Because our system is in hardware, there is no operating system to create problems. You can do many things in parallel without having to wait ... in queue.”

FPGAs come with challenges—they are difficult to program and it is not a common skill in financial services. “You can’t take a Java programmer and teach him how to build an FPGA,” says Neumann.

Schaumburg, Ill.-based vendor XtremeData last month released an accelerator module that incorporates dual highperformance FPGAs from Altera Corp. and runs on AMD quad-core Opteron processors. The module is designed to speed up compute-intensive processing in financial services data analytics, bioinformatics and video transcoding.

“The combination of the XtremeData XD2000F In-Socket Accelerator with the quad-core AMD Opteron processor allows our Activ MPU [market data processing unit] to keep its competitive advantage when it comes to low latency and overall throughput,” noted Mike Dunne, CTO of early adopter Activ Financial, in a statement.

According to Jim Benbow, chief hardware architect at XtremeData, the XD2000F accelerator delivers up to five times the memory bandwidth of earlygeneration products.

Low-Latency Lab
Intel has taken a simpler approach to low latency—in April it opened a highperformance lab outside London to help financial firms and technology vendors take full advantage of the power inherent in new releases of the company’s processors.

“We have faster architecture,” claims Nigel Woodward, Intel’s financial services director for the U.K. “The challenge is to get people to understand how to use our features to improve the speed of their processing.” Intel will charge a fee to cover costs but has no plans to enter the consulting business and make the lab a profit center. Thus far it has attracted leading independent software vendors and several major investment banks. One vendor achieved a 30 percent increase in speed by upgrading processors and tuning the application for better performance.

Rick Jacobsen, financial services marketing director of Intel in New York, says the company initially has focused on faster trading systems and risk analysis in financial services. Chicago-based Securities Technology Analysis Center reported that Reuters Market Data System achieved 2.8 million updates per second on a 16-core Intel Caneland server.

Intel is also planning to start a “quick assist” program to develop an abstraction layer that lets users deploy FPGAs more easily, if they opt to go in that direction.

The demand for low latency in the face of higher volumes should not be a surprise, says Brad Bailey, senior analyst at Bostonbased Aite Group. “People had been predicting it, because messaging rates grew four to five times from last year with more electronic connectivity and more techniques like algorithmic trading that require more messaging.”

According to Bailey, firms have to look at their existing systems as a starting point and use a combination of hardware and software to adapt them to new demands. “People will look at the myriad different systems they have already and try to put in systems that can be built in parallel or on top of what exists,” he says. “I think they will look at ways to use standard hardware.” He also expects solutions such as Tervela, InfiniBand and different data fabrics to see increased usage.

The summer’s trading volumes are providing firms with an impetus to make changes. “Very few firms did not have problems,” says Bailey. “When you prepare for 20 percent to 40 percent growth and you have 400 percent to 500 percent, that is a big factor. We are going to see a major convergence of efforts. Software firms will look at ways to leverage hardware when they see the throughput and capacity that are needed.”

International Securities Exchange (ISE) VP Greg Maynard agrees that both hardware and software approaches offer great promise. He calls Tervela and FPGA “fantastic products” that “can read the messages, do some processing, even put in some business logic while the data is going through, and they can do it in one-tenth of 1 millisecond.” With the exchange’s Opra traffic running at 500,000 messages per second and heading to 1 million, a single Tervela box could handle all the messages and pass them to multiple recipients.

But Maynard is keeping an open mind about the available solutions—he is also enthusiastic about software approaches that can run on standard Intel boxes. Software vendors “tell us we don’t really need a hardware solution, and while it maybe isn’t as fast, it is more flexible.” ISE is looking at three different approaches—very focused software developed in-house, powerful software from outside vendors, and hardware solutions.

©2008 SourceMedia, Inc. and Securities Industry News. All rights reserved. Used with permission.