Lehman, NYSE, CME, Forex Capital Pursue New Latency Killers
Streamlined Applications
Beyond the networks themselves, another potential source of latency is software applications. "The ideal developer in this business understands business applications and business goals, how to write functionally good applications, and how to write efficient applications," says Rubinow. "It's in this last category that you don't find people growing on trees, because in most businesses, people are not asked to do this." A good Wall Street programmer will find the fastest -- rather than the easiest -- method to run data sorts, for instance.
A major trend on Wall Street, however, actually may add latency to applications. To our surprise, Lehman's Mitsolides contends that by adopting services-oriented architectures, firms introduce new points of latency into their applications. "The fact that we're moving to a very heavily services-oriented architecture is creating complicated dependencies between services," he explains. "Once you have services calling other services that are calling other services and so on, latency compounds."
To reduce the problem, requests must be made in tandem, Mitsolides says. "But there are certain requests that are not naturally done in parallel," he notes. "For example, say you want to convert a deal and then you want to price a deal. How can you start pricing the deal before the conversion is done?"
According to Mitsolides, grid computing can help. Lehman's grid computing software from Platform Computing enables Mitsolides to allocate high-end computing resources to high-priority Web services and applications, minimizing latency times, he explains.
Compacting Data
But even the fastest applications are challenged to handle the mushrooming data volumes on the Street. As the Chicago Mercantile Exchange's market data volumes grow (due to market and regulatory factors, as well as its merger with the Chicago Board of Trade), the CME Group -- now the world's largest exchange -- has been striving to reduce latency while improving scalability. In 2007, the two exchanges handled 11 million contracts a day: 7 million at CME and 4 million at CBOT.
The CME started preparing for escalating data volumes two years ago, when it built and rolled out a multicast transport mechanism that breaks market data into product channels (e.g., an equity products feed, an FX feed, an options-only feed), allowing customers to receive just the data they need and enabling the exchange to scale its data delivery more easily. The next step in this progression is for customers to take market data directly from the exchange instead of going through data aggregators, such as Bloomberg, Reuters and Thompson. "One of the reasons we went multicast was we expected an increase in the number of direct consumers of our data -- and that's been happening," says Matt Simpson, associate director, electronic trading architecture, CME Group.
In its most recent effort to speed up market data delivery, Simpson notes, the CME adopted the FIX Adapted for Streaming Data (FAST) protocol for market data compression in October 2007, helping reduce latency to as little as a half-millisecond, he says. "It doesn't make sense to send data out in an uncompressed format," Simpson asserts. "You can get 5 million messages compressed in a second, such that you're spending a microsecond or less compressing a message." That microsecond is less than the amount of time it would take to put the uncompressed data onto the circuit, he contends.
If it takes 10 microseconds to put a 100-byte message on a 10-megabit circuit, for example, then taking that same message and compressing it in 1 microsecond down to 20 bytes (and decompressing it at the other end in another microsecond) enables you to transfer it onto a circuit in 2 microseconds, saving a net of 6 microseconds, Simpson explains. "FAST is able to reduce the transfer latency of sending market data, and at the same time reduce the size of uncompressed data by 70 to 80 percent or more so that you can use smaller lines and get bandwidth costs under control," he adds, noting that the CME is providing open-source code to its customers to enable them to decompress the data.
In January, all Chicago Board of Trade (CBOT) products were due to move over to the CME's GlobeEx platform, which will provide combined market data feeds. "Even though we're increasing the total volume by 40 percent," relates Simpson, "our market data latency will be well under what it would have been had we not been using FAST."
Simpson believes the CME was the first derivatives exchange to use FAST. But other exchanges that use the compression format include NYSE Archipelago, which uses it for options; the equities options exchange ISE; OMX's Eugenium and legacy platforms; and Eurex. OPRA, a utility that consolidates equity option market data feeds, plans to exclusively use FAST in early 2008.
Still, not everyone is bullish on data compression. "Data compression has some overhead associated with it because when you're sending the data you have to compress it, and when you receive the data, you have to uncompress it," the NYSE's Rubinow points out. "That takes compute resources." Rubinow notes that the NYSE decided to upgrade to a higher-bandwidth network that can accommodate InfiniBand and 10 gigabit Ethernet, making compression unnecessary -- for now.
Tweet This