Mellanox Bridges Network Performance Divide
In a Forbes article from last week, Tom Groenfeldt remarked on the increasing interest from the high performance computing world in the emerging trend of big data. This covers everything from HPC servers, storage, applications and of course, the core of the latency conundrum for the big data/fast processing–the almighty network.
Groenfeldt pointed to a number of ways the HPC and big data worlds are merging, recognizing the uptick in news items that revolve around how high performance hardware and software are leaking into enterprise big data solutions.
We’ve seen this happening in the processor, server and storage space in particular with traditional HPC companies like Cray, for instance, climbing aboard the big data express. In that same vein, today, Mellanox wrapped its HPC messaging in big data packaging with its announcement of the newest incarnation of its Messaging Accelerator VMA 6.0.
This update provides more robust support for both TCP and UDP acceleration over ConnectX-3 VPI adapter cards. The company says that the solution provides low UDP latency under 1.4 microseconds and TCP socket latency at just under 1.7 microseconds.
According to Mellanox, this is an achievement over other comparable products—they claim it’s twice as fast and will appeal to those with applications in the arenas of high frequency trading, databases, in-memory caching and others. While just five years ago this announcement would ring in the halls of the highest end datacenters, the need for rapid turnaround on complex transactional data, for example, is opening Mellanox’s messages to a potential untapped market (think retail, social media-driven data analysis, etc.).
Mellanox has already wrapped some messaging around some of the frameworks to support enterprise big data, including Memcached, Hadoop and the platform from HPCC Systems. They point to the relevance of solutions outside of Ethernet for providing the latency levels required for applications running across such platforms, including VMA and Infiniband.
The HPC network giant describes its approach to big data (while giving a hat tip to the new markets it expects to find) in the following statement:
“Big Data analytics used to be the exclusive domain of big enterprises. Now leveraging parallel processing technologies, the challenge of quick analysis of massive of data that streamed-in from different sources at a very high speed, is going mainstream. Parallel processing architecture breaks a big data analytics job into smaller jobs and runs it over tens, hundreds or thousands commodity servers which keeps data center costs low and provides easy scalability.
The above trends have existed in the HPC industry for a long time already, but the recent emergence of massive network access devices has led to a new services-oriented infrastructure which requires reliable high performance networking infrastructure. A lower latency and higher throughput network infrastructure enables data centers to increase their hardware utilization and to scale up to massive capacities in an instant without having to worry about bottlenecks and bad user experiences.”
According to David Barzilai, VP of marketing at Mellanox, “VMA 6.0 enables the highest performance of any socket-based application over our leading low-latency interconnect solutions.” He says that VMA offers datacenter managers the ability to tap into “low-latency data distribution or high packet rates based applications, resulting in faster executions, lower response time and higher return on investment.”
Others in the high performance network space, including QLogic, among others are equally enthralled with the HPC and big data opportunity. It will be interesting to watch as the two realms (HPC and enterprise big data) move closer together over the course of 2012.