Startup company Level 5 Networks has launched a go-faster NIC that it claimed can cut server CPU use by over half.
The product, dubbed EtherFabric, offers a pair of Gigabit Ethernet ports and can increase the performance of applications such as computational fluid dynamics by up to fourfold, according to marketing manager Craig Easley. The $500 product made its European debut at NetEvents this week and a 10 Gigabit product is due out next year.
Applications include high performance computing, such as scientific and creative applications, as well as high-end database servers.
It uses technology comparable to TCP/IP offload techniques, which have traditionally been the way to help solve the problem of the CPU spending significant amounts of time processing the network stack.
"This is of benefit for large block transfers but not small ones also for IP video or VOIP applications don't benefit because they use UDP", said Easley. "They also have limited scalability, while Infiniband architectures require a much bigger hardware upgrade."
"Instead, we virtualise the protocol stack and drivers so that each application has its own virtual copy. The OS scheduler spreads applications around each core or CPU, so each application takes its EtherFabric virtual driver with it to the core it's running on".
The product consists of a NIC with a Level 5 ASIC plus driver software. According to Easley: "It looks like a standard NIC. It calculates the CRC checksums and reduces latency to sub-10μs compared to 40-50μs range. Each application makes a network call, and creates a socket. We intercept that and create a virtual driver. DMA is used to transfer data to and from the card, so there's no use of CPU and it doesn't have to cross the front-side bus.
"The result is that it delivers more usable bandwidth to the application and reduces the I/O load on the CPU."
Easley added that the product would also be of benefit in a server running virtual machines, each of which traditionally will have its own network stack.
The company has received £30 million to develop the product, which is a high figure for a post-dotcom company. The brains behind the product come from Cambridge University, although the company is US-based.