Everyone on Wall Street wants to eliminate data latency. But data doesn't sit in one place for long. It passes among market participants, through myriad network switches, servers and applications.
Where in your IT infrastructure will a technology upgrade or change have the biggest impact on data latency? Will colocating servers to an exchange make all the difference? How about installing InfiniBand? Is buying a data caching solution that provides faster access to local data or installing hardware accelerators in routers, workstations and servers the solution? According to Tom Price, senior analyst at TowerGroup, colocation, faster internal networks and a state-of-the art client-side messaging transport have the biggest impact on reducing latency.
At the SIFMA show this week, a wealth of hardware and software vendors are telling a slightly different data-reduction story.
On the networking side, Cisco Systems (booth #2401) and Voltaire (booth #1762) are showing their InfiniBand switches. In March, a report put out by the Securities Technology Analysis Center in Chicago compared running Wombat market data feed software on a Cisco InfiniBand network versus a gigabit Ethernet. According to the report, the use of InfiniBand reduced mean latency by about 63 percent; reduced standard deviation by about 34 percent; and reduced outliers caused by request bursts, cutting roughly 30 to 35 milliseconds off of spikes.
The key feature of InfiniBand that helps it reduce latency, according to Pramod Srivatsa, senior product manager at Cisco, is remote direct memory access (RDMA), a concept whereby two or more computers communicate directly from the main memory of one system to the main memory of another. As there is no CPU, cache or context switching overhead needed to perform the transfer, and transfers can continue in parallel with other system operations, this is useful in applications where high-throughput, low-latency networking is needed, he explains. Cisco offers 20 gigabit-per-second InfiniBand.
Voltaire teamed with HP, Intel and Reuters to offer a joint low-latency trading solution at the show that uses its Grid Director InfiniBand-based switches (see related story, page 8). The four vendors announced benchmark results yesterday for the new combined solution in STAC tests in which latency was held constant and messages per second were increased 50 percent versus a similar setup with an Ethernet switch (rather than Voltaire's InfiniBand-based switch).
One key aspect of this announcement, according to Nigel Woodward, director of financial services solutions marketing at Intel, is that the combined solution has been proven (by STAC and in the vendors' labs). "Wall Street firms want products that have been tested," he says.
Also in the network infrastructure space, Solace introduced at the show an intelligent routing card — an enhancement to its VRS/32 network router that lets data travel directly from exchanges to a client without passing through a ticker plant, which the company says reduces latency by up to 100 times.
"Latency is becoming such a differentiator and competitive advantage in a lot of markets, algorithmic trading being one, that companies are looking very hard for the next technology that's going to make them a little faster than their competitors," says Peter Ashton, VP at Solace. "A lot of forward-looking investment banks are seeing that they can do message routing 10 to 100 times faster in hardware and get latency that's one-tenth what it would be in a software solution, which offers a huge competitive advantage to whoever gets that into their network first." Solace customers, Ashton says, see on average 400 microseconds of latency across the router, regardless of the overall volume.
The new card Solace is rolling out, the VRS/32-08 Assured Delivery Module, fits into the router's chassis and provides high throughput and low latency for "persistent messaging," guaranteeing that messages get to right place, for functions like order routing and funds transfer where messages cannot be lost, according to Ashton.
ACTIV Financial (booth #3114), an aggregator of data feeds and a provider of low-latency direct-exchange feed technology, is demonstrating the performance of its market data ticker plant software on an Xtreme Data Field Programmable Gate Array — a hardware accelerator that works with AMD's Opteron chip (through the Torrenza initiative). The company says that using the accelerator can reduce data latency to 100 microseconds.
Some Wall Street firms, however, have been hesitant to implement specialized hardware accelerators like FPGA boards because they have to be specially configured and physically installed, and software has to be modified to take advantage of their processing power. "We don't like the idea of someone in a garage with a soldering iron on a custom card, and neither do our customers," notes Frank Piasecki, president of ACTIV Financial.
Today, AMD provides the Torrenza platform into which FPGA boards can plug. And Intel has been working with FPGA makers such as Altera and Xilinx on using their products with the Xeon chip and recently demonstrated a Xilinx board plugged into the Xeon. Intel also is developing an accelerator abstraction layer that would help developers port portions of applications to accelerators. In addition, Intel and IBM are working on a hardware acceleration standardization initiative, based on PCI Express, called Geneseo. Some accelerator providers, including ClearSpeed, already use PCI Express.
Because of this standardization work going on and the crying need for more processing power without increasing server footprints and energy consumption, such specialized coprocessors have a bright future, Piasecki says. "If we stayed on our current platform and continued to parallelize our applications and split them across more boxes, the future looks grim," he says. "In three years we'd have a huge hardware footprint, and it would take a lot of people a lot of money, power, cooling and rack space to process market data. The total cost of ownership of these infrastructures has got to have some new paradigm around which people can work, and we think hardware acceleration is the only credible one available today."
For instance, Piasecki asserts, using hardware acceleration, a North American futures, equities and options ticker plant can run on 12 CPUs instead of 50. "One of the major problems facing direct-feed, low-latency providers is the cost of running these systems," he says. "And hardware acceleration is the holy grail around which we finally see the light."
Two vendors — GemStone (booth #2101) and Gigaspaces (booth #1419) — are showing data grids, solutions that eliminate latency in the access to in-company data by providing a huge memory cache. GemStone is introducing a joint solution with IBM and Intel called GemCache. "The idea here is you can deploy a cache in one fell swoop, so if you need a caching solution, you don't have to go and procure the hardware, operating system and our cache on top of that — we just do it as one nice package," says Mike Stolz, VP of architecture at GemStone.
GigaSpaces announced integration of its GigaSpaces eXtreme Application Platform with Microsoft Excel 2007, Excel Services and Windows Compute Cluster Server 2003. "[Traders and quants] can create any model they want, [and] they're familiar with it," says Geva Perry, CMO of GigaSpaces. "The problem that creates for these organizations is that you have multiple versions of the truth, because if each one maintains a model and keeps some data on his computer, you get silos of information and logic. If you try to create a central repository for the data and the business logic, you'll face a latency and a scalability problem. The joint solution between GigaSpaces and Microsoft provides a central repository with in-memory cache that will synchronize with the local data on machines using Excel."