A large, worldwide hedge fund company needed to reduce its server memory latency to keep investors from experiencing...
By submitting your personal information, you agree that TechTarget and its partners may contact you regarding relevant content, products and special offers.
costly trading delays. It found the answer from the small Oregon-based software startup RNA Networks and is using their technology to replace aging Tibco software.
The New York-based hedge fund company needed to increase memory capacity in servers quickly and, eight months ago, decided to give RNA Networks Inc.'s memory virtualization technology a try. According to the company's CTO, the hedge fund has experience significant performance gains.
"Latency is a big issue with trading – everyone is trying to beat each other, and even a second of latency can cost millions," the CTO said. "As our volume went up, latency went up and messages would be delayed, which is very bad in the trading world."
The hedge fund company supports its trading platform on standard x86 servers with Intel Xeon chips and about 32 GB of memory per box, running Linux. Before giving RNA a shot, the hedge fund company used Tibco messaging software, which provides a message bus for enterprise application integration. It used the software for messages, risk analysis and 23 other applications, but it "had its limits," and the company needed a technology that could move a high volume small messages around with lower latency than Tibco software provided, the CTO said.
RNAmessenger appears to have been the answer. "We save 1.7 milliseconds per message using this technology – which doesn't sound like a whole lot, but it is. In trading, time is everything," the CTO said.
RNA prices its technology on a per-node and per capability basis, so pricing varies. In some environments, the cost might be below $2,000 per node with 2 to 5 TB of data, Cook said. "The cost is about 60% less than buying equivalent amounts of physical memory to what we provide," he said.
For the hedge fund, RNAmessenger is "probably a little cheaper" than adding physical memory, but cost wasn't a top priority, the CTO said.
"We really needed the lowest possible latency and reliability," the CTO said. "We liked the speed, plus [RNAmessenger] records all the messages, so if something goes down, the messages are not lost. When you do lots of volume, having the messages recorded on a cache server is very important to prevent outstanding orders."
How RNA's memory virtualization technology works
The company describes its technology as memory virtualization, but in reality, it doesn't virtualize anything. Rather, the technology combines existing memory resources throughout a data center into a memory pool that server nodes on the same network can access.
This technology solves a major problem for IT pros. Over the past few years, Intel and AMD have advanced processor performance and efficiency at a much faster pace than memory manufacturers have, so the latest chips end up spinning their wheels waiting for memory to complete data requests. Adding more physical memory to compensate for this gap is cost prohibitive.
"The advances in memory capacity has not kept pace with CPU advancements, so often, servers are replaced simply to add memory to the system. The cost of that is very high, and memory drain on power is also very high," said Clive Cook, CEO of RNA. "We are allowing you to add performance using the resources you already have in the data center."
RNA's technology works by detecting all the available dynamic random access memory (DRAM) memory from servers across a network and stitching that memory together into a pool of up to 10 TB of storage. That pool appears to operating systems and applications as a single unit of memory, Cook said.
Installation is fairly simple; a kernel driver is installed on servers that can either contribute resources to the memory pool or access it. Any Unix operating system and applications can access the pool, which can be used as a high-speed cache, a messaging layer, or a shared memory resource for CPUs.
This approach to adding memory resources to servers is much different from yCisco's Memory Extend technolog in Cisco's upcoming Unified Computing System, Cook said. "That memory is still limited by the physical footprint within the chassis," whereas RNA technology enables servers within a chassis access memory outside that enclosure.RNA's CTO claims other companies have done memory virtualization as it does, and that may be the case. But several software products use the idea of server resource pooling, including Egenera's Processor Area Network (PAN) Manager software and VMware's Distributed Resource Scheduler. RNA Networks introduced its second product, called RNAcache, on June 22.
Tibco was contacted for comment on Friday but did not respond by press time.
Let us know what you think about the story; email Bridget Botelho, News Writer.