“Pantor Inside” could become a new tagline for the Stockholm-based company which is achieving very high performance transaction processing on the latest Intel generations of processors. It uses the Intel fasterLAB to test and modify large configurations of the Pantor Framework and Pantor Fast Virtual Machine.

Because Pantor’s technology is used by several investment firms and is on the leading edge of differentiating their trading activity – the only public reference it can use is from SunGard Front Arena which has announced it uses Pantor’s FAST Virtual Machine in its trading solution.

Designed to handle extreme spikes of data without inducing latency or dropping packets, Pantor Engineering is working with the Intel fasterLAB on continuous improvement of its already extremely high performance products. Pantor Presto, for example, encodes FAST more than ten times faster than the original 2005FAST proof of concept using benchmark data from OPRA.

“The lab gives us tremendous benefits because we can test new hardware models before they are launched,” says Rolf Andersson, CEO of Pantor. “Whenever we have a client with demanding requirements on speed, we can test the latest and greatest from Intel in engineering white boxes or from production OEMs which is very useful. And we are also able to test large configurations and combinations of infrastructure components to assess optimum contributions to the latency race. With 45nm quad core Harpertown we used six fully equipped boxes with 10Gb and 1 Gb networking and got very, very good results. Obviously we wouldn’t have seen that at home because we don’t normally have access to that many boxes in our lab.”

In addition to providing the latest processors, fasterLAB gives users the chance to work with Intel’s latest and fastest hardware, such as its 10Gb NIC, code-named Oplin (Intel 82598 10 Gigabit Ethernet Controller), and its extensive range of software tuning tools – many of which are not widely known in the market.

Without access to several computers, a firm can’t determine if its solution will work in realworld production conditions. Users have often discovered that solutions which worked well in a small lab break down in large-scale production.

“You need this to test horizontal scaling for market data or market feed systems”, explains Andersson. “You want to know that it actually scales to 30 parallel pipes and that it will remain fast if you cascade data to hundreds of users. Testing that requires a lot of machines, and ISVs rarely have the equipment to do that in their own lab.

“With Harpertown we found linear scaling in places where we didn’t think we would. Software performed better than we expected and then when we tried it on a large box with multiple CPUs we found it was linear all the way.”

The demands on systems are growing constantly. This summer, options data is projected to reach 900,000 messages per second; in three years the markets have seen a 9x increase in volume, more than 100 percent per year, he says. Messages per second is a measure, but it doesn’t tell you how data is flowing within your system.

“Things are happening within the second, so you should look at your latency period and have that as your measurement. Millisecond – you did that a year ago. Now aim for a tenth of a millisecond. In that time you have 90 messages, six frames. We can do 8 million messages per second, or 8 per microsecond. You have 125 nanoseconds to code your message. Is this relevant? Can we expect to handle several messages in 100 microseconds? Yes. One stock price change can generate hundreds of changes in options prices. As soon as you want to trade with the large market players you may have to decode a message in one-eighth of a microsecond.”

Anders Furuhed, Pantor’s chief technology officer, explains the performance improvements the company achieved at fasterLAB.

“Pantor ORDO, a high-performance, fault tolerant application server and gateway configured for a derivatives exchange mass¬quoting benchmark, was able to process and generate best bid offer messages for 300,000 two-sided quotes per second with an average client to client latency of 500 microseconds. The low latency of Intel’s recently launched 82598 (10Gb) Ethernet controllers with I/ OAT technology, combined with the X5440 processors, is key to achieving this level of performance as the fault tolerance built into Pantor ORDO is based on networked servers. Intel’s I/OAT technology has helped us to decrease the mean latency by 20 percent while simultaneously improving transaction rates by 15 percent in our ORDO framework.”

“FAST is purpose-built for market data and for high speed trading and stat-arb trading”, adds Andersson. ”It is also well-suited to hardware acceleration.”

Using a new Intel Core 2 machine, Pantor did 8 Gbit/s of market data feed with eight parallel feeds from the same machine. It nearly saturated one 10 Gbit/s pipe, he adds. Previous tests were not capable of enough throughput to meet the capacity of a 10 Gbit/s connector but the lower overhead delivered by Oplin enables unprecedented I/O rates.

“Sixty-six million messages per second was the aggregate peak rate on all parallel feeds with the same network interface. It was to show how quickly you can get data in and out of a modern machine – it was more I/O testing than CPU testing.”

(No Ratings Yet)  Loading ...

Filed under: Issue 5 - Summer 08, Technology, Trading