Voltaire Splits Nanoseconds With Grid Director
You probably wouldn't think that spending a thousand dollars to save a second of productivity would be a worthwhile investment. Nor would your CFO. Yet the makers of a device costing $34,000 hope you'll spend that much to save a few hundred nanoseconds. The company is Voltaire, and its device is the Grid Director 4036E, a high-performance InfiniBand-Ethernet switch and IP accelerator that it claims delivers lower network latency than competitive products and gives securities trading applications
January 28, 2010
![Network Computing logo Network Computing logo](https://eu-images.contentstack.com/v3/assets/bltde8121fc52c5c8f3/blt3f3d0318f746b1c2/65a530e4187606040a1d8b8c/placeholder.png?width=1280&auto=webp&quality=95&format=jpg&disable=upscale)
You probably wouldn't think that spending a thousand dollars to save a second of productivity would be a worthwhile investment. Nor would your CFO. Yet the makers of a device costing $34,000 hope you'll spend that much to save a few hundred nanoseconds. The company is Voltaire, and its device is the Grid Director 4036E, a high-performance InfiniBand-Ethernet switch and IP accelerator that it claims delivers lower network latency than competitive products and gives securities trading applications a decided edge. The Grid Director 4036E will cost about US$1000 per port.
Set to begin shipping in March, the Grid Director 4036E is a 1U rack mount device with 34 InfiniBand ports, each capable of 40Gb/s sustained transfer rate, plus a pair of Gigabit Ethernet ports of either 1 or 10 Gb/s. Total throughput is rated at 2.72 Tb/s, according to the company, with InfiniBand port latency at less than one hundred nanoseconds, and hardware-based InfiniBand-Ethernet bridging in less than two microseconds.
Concerning the unit's target market, Asaf Somekh, Voltaire's vice president of Marketing, says, "In investment banks, exchanges and trading organizations, the trading strategy has been transferred from the brains of traders to algorithms and is implemented in very fast ways. For them, latency is everything. If you can offer a lower latency, they can be better than the next bank or trading company." Algorithmic trading systems, he said, issue simultaneous orders to multiple exchanges for the same security. "Whichever firm comes back first wins the business." So a few hundred nanoseconds can translate to hundreds of thousands of dollars in business for a given transactions, putting millions of dollars per year at stake.
Backing up that claim is Feargal O'Sullivan, head of Enterprise Software Americas for NYSE Technologies, which develops, hosts and deploys data centers and trading systems. "The speed of light counts for our users. They do everything they can to reduce the time it takes to make trading decisions within their own servers," he said in a phone interview today. According to O'Sullivan, these systems contain three main parts: the market data-feed handlers that process raw market data, a middleware distribution layer with a range of tools for making trading decisions and market access gateways, which send trades and orders back to the markets.
At the heart of such trading decisions, O'Sullivan explained, is the matching engine, putting buyers with sellers, and its physical location can be a major factor. "If you're 20 miles away or 200 yards away, being that close guarantees the lowest possible latency. They want the order to get in quickly because someone else might be next to them trying to get the same trade."Somekh said the ability to bridge InfiniBand and Ethernet networks is unique in the industry and is primarily for connecting market data feeds, which typically run over Gigabit Ethernet with a trading organization's InfiniBand-connected servers and network-attached storage. "Typically they have a 10Gb NAS. They'll use the 4036 to accelerate a particular application to access the storage through the and eliminate the need for 10Gb adapters on the servers that are being accelerated."
Part of the Voltaire solution, its ability to accelerate multi-cast traffic, rests in the software. "Because of the speed of the traffic and the way most trading platforms are architected, we needed to process traffic from tickers and market feeds--Reuters or Bloomberg--in a unique hashing algorithm we came up with," part of which runs on the server, he said. "A software layer sits between the application and the InfiniBand adapter. It bypasses the operating system, leveraging the low-latency of adapter while transparent to trading app. There is no need for customer to change anything in their software to be accelerated by orders of magnitude." NYSE offers a similar solutions of its own, which uses remote direct memory access (RDMA) to bypass the IP stack. "The application gets told that information is in memory and it can read right out of that memory," O'Sullivan said. "This gives people the absolute lowest latency within their co-located rack server."
Another target market for the Grid Director, Somekh said, is high-performance computing. "Government labs, universities, oil and gas companies and the automotive industry all have been using InfiniBand in high-performance computing applications for years." On the storage side, companies currently employing a NAS also could benefit, he added. "File servers that have 10Gb Ethernet don't need to use another network for storage. They can use their high speed clustering and use the same adapter for storage directly through the 4036," he said, and claimed there's no need for any special libraries or drivers.
You May Also Like
Maximizing cloud potential: Building and operating an effective Cloud Center of Excellence (CCoE)
September 10, 2024Radical Automation of ITSM
September 19, 2024Maximizing Manufacturing Efficiency with Real-Time Production Monitoring
September 25, 2024