RNA Networks Delivers High Throughput, Low Latency via Virtual Memory
25 Apr 2010
RNA Networks is leveraging memory virtualisation technology to tackle memory bottlenecks that degrade data intensive trading applications. The company’s middleware platform is aimed at improving application performance, eliminating the capacity and cost constraints of server memory, and making memory a scalable networked resource.
The company was founded as a start-up in Portland, Oregon, in 2006 by ex-executives from Intel, Cisco and the supercomputing space. Its initial intention was to take the concepts of high performance computing and apply them to the enterprise data centre, solving complex computational problems that use huge amounts of data. The company looked at how organisations such as Google and Facebook handled large-scale data problems and considered solutions, such as the use of objects and applying tick data in the oil and gas markets.
As RNA Networks developed its memory virtualisation software solutions, its intent was to target the commercial enterprise market, offering support for more users, applications and virtual machines with less resources; faster and better analytics; accelerated file and data services; low latency application platforms; and, ultimately, an improved user experience.
The company’s work also attracted the attention of financial services institutions looking for high data throughput and low latency. According to Clive Cook, CEO of RNA Networks, “Interest came from areas that we had not envisaged, such as the use of our technology for market data feeds, to support input to algo trading engines and to use algo trading results to drive execution engines.
“Memory virtualisation supports ultra low latency messaging because you no longer have to pass messages through network sockets or messaging server layers. For trading floors this means micro-second delivery of execution orders which is mandatory for automated trading. Also, the algo analytics themselves get a huge boost in speed because memory virtualisation pulls DRAM memory together to create a multi-terabyte memory cloud. This enables in-memory performance for analytics, enabling near real-time access to data for algo programs and empowering traders to build more sophisticated, data-rich models.”
RNA Networks’ memory virtualisation platform is the basis for its RNAmessenger product, which accelerates ultra-low latency, high volume messaging applications, and RNAcache, which drives faster results from business critical, memory intensive applications.
The company’s first customer for RNAmessenger was a US-based, global multi-billion dollar hedge fund that signed up in early 2009. It described the software as ‘unmatched by any alternatives’ and said it ‘delivered dramatic performance gains’ for the business.
Cook says the software solutions were driven by early hedge fund clients, adding: “Many hedge funds now use our platform for guaranteed delivery of messages at very high rates. The software does not dictate a single connection into, say, Nasdaq or NYSE, instead users can go to multiple liquidity pools.”
Despite products such as Intel’s Nehalem processor adding more memory, RNA Networks claims its distributed but parallel model for memory virtualisation can scale commodity hardware to performance levels greater than any other system. The software is best suited to servers linked using remote direct memory access (RDMA) networking, an aspect that could be favourable to RNA Networks as RDMA moves beyond Infiniband connectivity and is implemented using 10 gigabit Ethernet with network cards from providers such as Intel and Chelsio.
Another opportunity for the company may be Basel II, which is expected to increase pressure on investment banks to improve and prove risk analysis policies. As Cook concludes: “Expect to see some compelling applications from us and our partners as banks will be required by regulators to do more risk modelling and this will be a big headache for any that can’t handle the vast amount of data involved.”
RNA will be pre-viewing the capabilities of an upcoming 2.5 release of the product at the A-Team Insight Exchange event in New York this week. RNA MVX 2.5 is introducing what is best described as a memory cloud. RNA MVX combines memory from many servers into a distributed memory service. To date customers have used the memory cloud for low latency messaging and as a cache for fast access to data. MVX 2.5 expands the memory cloud for two new capabilities: extending memory space for large memory applications and providing data storage right in the compute tier for data intensive applications.
Today, when applications run out of memory, they swap and performance dies. This places a practical limit on the size of models and analytics that can be built for algorithmic trading. The MVX memory cloud breaks that barrier, mobilizing memory from many servers to create a multi-terabyte high-speed memory space. Now memory hungry analytics and models can be run on existing mid-sized systems, assembling multi-terabyte memories without requiring expensive very large memory servers. The immediate result is that applications such as risk analytics, market simulations, and correlation analysis can suddenly run much more comprehensive models, at much finer detail, dramatically faster without changes to application code.
Other algo applications need frequent access to large data files as part of their processing. For these applications, the memory cloud becomes a place to store this data, keeping data close to the application inside the compute tier. This can be used to supply information from one stage of analysis to another or for cloud storage of data that is accessed hundreds to thousands of times by different applications. Many organizations are working to increase the number of models they run, the frequency they run them, and the amount of information fed into the analysis, all to better manage intraday risk and gain trading advantage. Early users of the new MVX capabilities are finding the memory cloud enables large data sets, supports greater processing frequency and delivers a significant performance boost.
RNA MVX 2.5 is a new type of virtualization — a memory visor — that manages memory as a virtualized resource. In addition to the business impacting performance benefits, a side benefit of MVX is that it leads to higher utilization of compute resources. By leveraging the memory cloud to eliminate memory bottlenecks that affect processing and I/O bottlenecks that impact storage, the result is that server cores become better utilized, storage systems are offloaded to become more responsive, and IT can acquire capital more incrementally and get greater reuse from existing investments.
The new capabilities of RNA MVX have been available in early release to select customers and general availability of MVX 2.5 is expected to be announced in May.
