xCelor Readies Low-Latency FPGA ‘Ticker Plant in a Switch' on Arista Hardware

Chicago-based high-performance data technology vendor xCelor is testing a version of its hardware feed handler and normalization technology running on Arista switches to eliminate switching latency, dubbed TPS (Ticker Plant in a Switch).

Rob Walker, chief technology officer at xCelor, says he expects to release TPS within weeks. The product is currently working well in testing with Nasdaq data, and the vendor will port its other FPGA feed handlers -- for NYSE Arca, Direct Edge and the BATS Exchange -- to TPS shortly after launch, with an FPGA handler for CME data currently in development. Subject to demand, the vendor can develop FPGA feed handlers for other markets in around 90 days, he adds.

"Some people stick our FPGA feed handlers -- which are faster and more deterministic because they are in hardware -- straight into servers running their trading logic... while other groups have a ticker plant to centralize market data normalization within a co-location facility," Walker says. However, sending data through multiple hops and switches between an exchange, ticker plants, network card, and trading engines, can slow down data, he adds.

"The market data has to travel through a switch to get to the trading servers regardless... so why not combine our FPGA card with that switch? So we are normalizing on the switch itself. Arista's Fulcrum chip... takes around 500 nanoseconds to perform its switching function. But we can switch directly on the FPGA in just 240 nanoseconds, so we have 260 nanoseconds to perform the normalization process," Walker says. "TPS will be complementary to our existing FPGA feed handlers. It does the same as the cards but in a different way -- for example, if you want to perform normalization centrally rather than on your trading servers."

In addition to performing data normalization, TPS can also filter large datafeeds to assign specific symbols to each port, and cut down feeds overall, enabling the data to travel faster. "You configure TPS so that although the raw feed goes in, what comes out is only data on the symbols you're interested in. Trading servers are better off trading rather than churning through redundant data," Walker says. This ability to filter data should also open up the vendor's services to a broader client base, he adds.

"Most of our existing clients are very sophisticated high-frequency trading firms operating in single-digit microseconds, and we are targeting tier-one and -two firms for whom every last microsecond matters. But that's a niche product for a limited number of potential clients, so we don't necessarily need a big sales force... so for the phase of growth we are at right now, we are trying to partner with third parties to get our name into the marketplace and bring clients to us," he says.

Walker says xCelor will soon announce a formal partnership with Arista, which will enable xCelor to leverage the vendor's sales force to read a larger potential client base without having to expand its own resources. "Arista already has a good footing in the HFT market... and is looking to incorporate more functionality onto its solutions," he adds.