This week Mellanox announced the newest version of its VMA 6.0 Messaging Accelerator, which includes enhanced TCP and UDP acceleration support over ConnectX-3 VPI adapter cards. With ultra-low UDP latency under 1.4 microseconds and TCP socket latency under 1.7 microseconds, this messaging technology is reportedly more than two times faster than competitive offerings.
I caught up Mellanox’s Gilad Shainer to talk about how VMA 6.0 accelerates applications like high frequency trading.
insideHPC: How does VMA 6.0 accelerate performance in socket-based applications? Does it employ some form of kernel bypass for UDP?
Gilad Shainer: VMA bypasses the kernel by moving TCP and UDP to the user space. With that it can reduce the overheads and accelerate the latency. VMA also takes advantage of the world lowest latency NIC – ConnectX-3 to provide the lowest latency in the market.
insideHPC: How will socket-based applications like those used for high-frequency trading benefit from VMA 6.0?
Gilad Shainer: Low latency is the major item here, as well as the lowest jitter – VMA guarantees the same low latency regardless of how many processes run on the server
insideHPC: Is VMA 6.0 strictly a software solution for delivering low latency?
Gilad Shainer: VMA is a software solution that takes advantage of Mellanox adapters
insideHPC: VMA 6.0 works with ConnectX-3 adapter cards, which support both Infiniband and 10 GigE. Does the acceleration of VMA 6.0 only work
on the IB side?
Gilad Shainer: The acceleration works on both IB and Ethernet
insideHPC: How important is low-latency in this age of ever-increasing cores?
Gilad Shainer: Low latency is critical – more cores means more processes that need more from the interconnect side. The requirements are not only for low latency, but with more cores – to guarantee the same low latency regardless of how many cores/processes exist. This is a tough requirement, and other vendors are failing to deliver the desired solution.