Mellanox VMA 6.0 juices high freaky trading
Making Linux sit up and make money
If you use Connect-X host bus adapters from Mellanox Technologies for your Ethernet or InfiniBand networks and you use Linux, then there's a turbo boost mode available for $300 per server.
The software used to be called Voltaire Messaging Accelerator and came to Mellanox through its November 2010 acquisition  of networking rival Voltaire. Now simply called VMA, it was originally bundled with the Grid Director 4036E quad data rate InfiniBand switch announced in January 2010  and was used to map multicast market data feeds to the switch chip, offloading it from the server to the switch and lowering the latency of the transactions in the feeds.
Somewhere along the way, the VMA networking offload software was tied to the ConnectX line of Ethernet and InfiniBand adapter cards and cut loose from the switches. The prior VMA 5.5 release from May 2011 was tuned to the then-current ConnectX-2 adapters, and the new VMA 6.0 release only works with the next-generation ConnectX-3 adapters that was announced in June 2011  and started shipping last fall.
There are PCI-Express 2.0 and 3.0 variants of the ConnectX-3 adapters, and the idea – although no-one at Mellanox will ever admit this – was no doubt to time the PCI-Express 3.0 versions of the ConnectX-3 cards to be available late last year when Intel was widely expected to launch its "Sandy Bridge-EP" Xeon E5 processors, the first server chips to support PCI-Express 3.0 peripherals.
As we all know, Intel has not formally launched the Xeon E5s, but is shipping them to customers  on a selected basis; the x86 chip maker has not been tied down to a launch date yet, saying only that the Xeon E5s will debut formally in early 2012.
Gilad Shainer, senior director of high-performance and technical computing at Mellanox, says that the new VMA 6.0 announced today will work with any switch so long as it is talking to the server through a ConnectX-3 adapter card.
The VMA software running on the server in conjunction with the ConnectX-3 adapter is a dynamically linked user space Linux library that offloads the processing associated with TCP, UDP Multicast, and UDP unicast traffic to the adapter card - instead of running it on the operating system kernel and network stack on the processors inside of the server. VMA has been implemented for Linux and any application that uses BSD sockets can use the VMA library created by Mellanox without having to make any changes to the application.
All the server will know is that it has more cycles left over to do more real work, and all the applications will know is that UDP latency will be under 1.4 microseconds and TCP socket latency will be below 1.7 microseconds. That's about twice as fast as competitive offerings, says Shainer, and also about double what was possible with prior editions of VMA running on Voltaire switches. This will get the attention of high frequency trading operators and hedge funds, that live and die by the microseconds.
By the way, the VMA 6.0 spec sheet (PDF)  says this sustainable latency can be delivered under a load of 3.5 million packets per second on the server running the UDP protocol on loads using 12 byte messages, which is about 3.5 times what a Linux OS can do on its own.
Boosting the message size drops the VMA throughput to three million packets per second and the Linux OS-only message rate holds steady at about one million. With messages weighing in at 1,460 bytes, the VMA offload drops to around 1.6 million packets per second and the Linux OS all by its lonesome drops to around 800,000 packets per second.
When servers start using PCI-Express 3.0 peripheral slots, the latency and bandwidth gains from VMA will be even larger, Shainer tells El Reg.
VMA 6.0 will work with the 10GE and 40GE Ethernet ports and 56Gb/sec Fourteen Data Rate (FDR) InfiniBand ports on the various ConnectX-3 adapters. (The prior VMA software did not support 40GE or FDR InfiniBand.) You can use VMA on servers running the original Red Hat Enterprise Linux 6 and the 6.1 update as well as RHEL 5 Update 6 or Update 7. Attachmate's SUSE Linux 11 SP1 can also run the modified Linux library from Mellanox. VMA does not work with Windows – at least not yet. The VMA 6.0 software has about 30 per cent lower latency than VMA 5.5 from last May.
The VMA software costs $300 per server, and with each server generally only having one ConnectX-3 adapter, it works out to being one VMA license per ConnectX-3 adapter. ®