As the industry struggles with increasing data message rates with no end in sight, market data technology providers are searching for alternative solutions to the current method of throwing more boxes at the problem.
Two weeks ago, ACTIV Financial, an aggregator of data feeds and low-latency direct-exchange feed technology, announced that it's porting, its next generation market data ticker plant solution, directly onto silicon.Eyeing the potential of hardware acceleration technology, ACTIV is introducing ActivFeed MPU (Market Data Processing Unit) for low-latency exchange feed processing, which is going to run on a specialized coprocessor that can be closely coupled with the AMD64 technology.
ACTIV has been working closely with AMD through its Torrenza Project, which supports the integration of specialized coprocessors in systems based on AMD Opteron microprocessors. Specifically, ACTIV is using an FPGA (Field Programmable Gate Array) co-processor and AMD64 technology with Direct Connect Architecture and HyperTransport Technology.
"We're concerned that when you have increased data rates and you just scale with boxes to handle the increased data rates, you end up having an increasing cost that is correlated to the increasing data rate," says Frank Piasecki, President of ACTIV Financial. "The name of the game is to find out how we can beat the curve by having more compute power deployed to the problem more efficiently," says Piasecki. "We're intrigued with what hardware acceleration technology can do to the general purpose CPU model," he adds.
The traditional model in the market data industry is run a software-based ticker plant, which means software running on Linux or Windows on a commoditized chip like AMD, explains Piasecki. "All the players such as Reuters, Infodyne and Wombat, use this construct," says ACTIV's president.
Instead, ACTIV is running its ticker plant application on the FPGAs. Then ACTIV is making a connection to the chip through the HyperTransport slot that's built as part of the AMD chip set. There's one CPU and there are three slots on the HyperTransport in which ACTIV plugs in its FPGA. The chip has two more "slots on to which we could implement other FPGAs," says Piasecki, which means there is room for growth.
By converting its ticker plant technology from software straight onto silicon, Piasecki expects huge capacity gains, dramatic reductions in numbers of servers and significant improvements in latency. "It allows us to perform more instructions simultaneously to get the job done more quickly," says Piasecki.
The silicon version of ACTIV's ticker plant can process 20 times more messages that it can today, he continues. "Let's say it takes 12 devices to run a standard ticker plant on some North American feeds to day. Writing our application on these FPGAs allowed us to have those number of servers reduced to one," he says.
Latency - the time it takes ACTIV to process data - can be reduced from about one millisecond today down to less than 100 microseconds on the ActiveMPU Platform, he says. (In calculating this end-to-end latency, he includes normalization of the exchange messages, application of all the value-added calculations, distribution and then exposing the data through the customer interface or client API.)
The speed is due to the environment that AMD has built, and the ability of these FPGA chips to access the main memory and network IO at extraordinary high rates, he says. "It's a whole environment in which moving data back and forth and accessing memory all happens in the fastest possible manner without adding latency," he says.
According to Doug O' Flaherty, division manager for acceleration strategy at AMD - who leads the Torrenza Project, "The principal behind the Direct Connect Architecture is that you can have FPGA directly connected to an X86 processors. As a result, the amount of data you can move back and forth is higher," he explains.
So what does this mean for customers that are facing a tsunami of data such as OPRA rates projected to rise to 716,000 messages per second by January 2008, up from 22,000 MPS four years ago. "This allows customers to deploy a solution with head room and growth in the next couple of years without purchasing additional hardware," he says.
The application is targeted at customers that place a time sensitive value on their exchange data those that have made a substantial investment in RMDS, Reuters' market data distribution system. And is consuming more options depth-of-book where the information his have a scalability problem.
"Our own aggregation ticker plants are going to be the first beneficiaries of this technology," says Piasecki. ACTIV expects to have a product in customer hands before the end of the second quarter. Ivy is Editor-at-Large for Advanced Trading and Wall Street & Technology. Ivy is responsible for writing in-depth feature articles, daily blogs and news articles with a focus on automated trading in the capital markets. As an industry expert, Ivy has reported on a myriad ... View Full Bio