Precision I/O has unveiled a new high-performance server I/O architecture based on Ethernet technology.
The Palo Alto, Calif.-based company claims it’s the first to solve the server-to-network bottleneck problem using standard IP/Ethernet infrastructure. Other approaches have been proprietary solutions or new switch fabric technologies such as InfiniBand that require users to deploy a secondary switch infrastructure and make software modifications.
TCP Offload Engine (TOE) solutions are based on standard Ethernet, says Judy Estrin, Precision’s chairman and acting CEO, but “they mistakenly treat the
protocol processing, instead of the operating system overhead, as the primary performance obstacle.”
InfiniBand bypasses the OS — the key to achieving very low latency — but forces users to implement a new network fabric side-by-side with their packet-switched IP networks. “This adds untold cost and complexity, requiring new hardware and drivers as well as changes to applications,” says Estrin.
A third solution, remote direct memory access (RDMA), “adds a whole new protocol that must be deployed at both ends of the wire to achieve any performance improvement at all,” according to Estrin.
By addressing the way IP/Ethernet packets are processed on each server, rather than developing alternative protocols, Precision says it offers the advantages of other solutions without their drawbacks. The company’s technology takes the networking function out of the path of the operating system, achieving high throughput, low latency, and greatly improved CPU utilization using the high-volume, cost-effective Ethernet foundation.
Precision reports its products will require no new switch fabrics, packet formats, or changes to applications, and will operate on servers running any
number of CPUs. Because it leverages the existing IP infrastructure, the technology — unlike InfiniBand or RDMA — can be deployed at just one end of the wire, such as one server of a given pair, to boost performance at that end.
Precision’s products will be implemented initially as software solutions that support network speeds of up to 10 gigabits per second (Gbps), and later as
hardware/software solutions that support wire-rate processing of 10 Gbps. Products will be offered for Unix, Linux, and Windows operating system environments. The company plans to sell its products chiefly through system integrators, value-added resellers (VARs), and original equipment manufacturers (OEMs).
Analysts Mixed on Threat to InfiniBand
Analysts are mixed on Precision’s threat to InfiniBand, with most saying the competing technologies address different markets.
“I see the Precision position as not a direct competitor to InfiniBand,” says Vernon Turner, IDC’s group vice president for global enterprise server solutions. “IB will be able to secure the corner of the market where expensive proprietary interconnects reside today. I think that Precision is going after a broader, more balanced commercial market wanting to have latency, throughput, and utilization be addressed. I don’t think that this was IB’s positioning, and could be seen in hindsight as a missed opportunity.”
John Abbott, chief analyst at The 451 Group, says Precision has “a chance” against InfiniBand, but that IB maintains a performance edge.
“I think they have a chance if it works, because a software-only approach requires fewer modifications to the fabric and to the applications than either InfiniBand or RDMA-over-Ethernet,” Abbott told Enterprise Storage Forum. “It’s hard to tell at the moment, but it looks as if InfiniBand will continue with a performance advantage, but could lose business at the low end on price.”
“The key test will come when 10GigE comes out in a few years time,” continued Abbot, “but Precision is working on a combined hardware-software bundle for that market, which suggests there are performance limits to its software product. Still, the company has plenty of IP expertise on board.”
Mark Hoover of Acuitive, on the other hand, sees Precision as a big threat to InfiniBand.
“If the Precision I/O technology works as promised — and we’ll know in another few months — then I think the question will be reversed: Is InfiniBand really a challenge to optimized TCP-over-Ethernet? My answer to that — under the assumption that the Precision I/O technology really works as advertised — is no. The niche that Infiniband would like to fill will have been eliminated.”
However, Hoover says the debate about Precision vs. InfiniBand “misses the point about Precision. Precision offers the promise to make all networked devices operate much more efficiently, Web servers, Network Attached Storage devices, file servers, application servers, database servers, DNS servers, DHCP servers, etc., any server for which a good chunk of its effort is getting bits into and out of it. In the future, even clients might benefit from the technology when I/O becomes more of a performance issue in them.”
Precision’s ability to work on a single end is another plus, Hoover says. “What that means is that if you put their software on a server, that server will run faster independent of whether the client or server or storage device it is talking to is running the same technology,” he says. “That’s a pretty big deal. Contrast that with InfiniBand, whose similar advantages are only obtained among the group of devices attached to a usually small and local InfiniBand network.”
Precision also announced that it has raised $10 million in first-round funding, led by Advanced Technology Ventures, 3i, and Foundation Capital. Precision I/O was spun out in March 2003 from Packet Design, which provided seed funding.
Additional funding was provided by Masthead Venture Partners and Incubic.
In addition to Estrin, Precision I/O’s management team includes chief scientist Van Jacobson and CTO Bob Felderman, among others. A search is underway for a permanent CEO.
The company plans to bring its products to market mid-year.
Back to Enterprise Storage Forum