December 16, 2004 4:00 AM PST
Making big iron from lots of little iron
Executives from Katana Technology--to be renamed VirtuOS Computing next year--say the company's software will let computer administrators link low-end Linux servers into a single powerful machine.
Think of it as making a big iron server out of a bunch of little iron. But unlike many approaches to ganging smaller machines together, VirtuOS promises that higher-level software won't have to be changed to benefit from the approach.
Some analysts, however, are skeptical that the technology is mature enough yet. "It's clearly the direction of the future," said Gartner analyst John Enck. "But do I think this stuff is fully baked? No, I don't."
The VirtuOS software lets numerous software jobs run on a group of as many as 128 dual-processor systems, with more processors being added or subtracted to individual jobs as workloads rise or fall, said Scott Davis, co-founder and chief technology officer.
"We can take sets of commodity servers and create a virtual data center. We carve that up into a set of virtual computers," Davis said. He likens the idea to what's happened with storage, where programs communicate not with real disk drives but with an abstracted storage system whose underlying components can be more easily changed.
A flexible computing infrastructure
That kind of flexibility is a key part of a major technology shift advocated by the world's biggest computing companies. IBM, Sun Microsystems, Hewlett-Packard, Microsoft and others want customers to be better able to match equipment to business priorities and to avoid having computers lying mostly idle.
The VirtuOS software runs on top of servers with x86 chips such as Intel's Xeon or Advanced Micro Devices' Opteron, while Linux runs on top. Davis said the company eventually plans Windows support as well.
Many in the computing world are headed in the same general direction as VirtuOS. For example, Oracle aggressively markets the Real Application Clusters, or RAC, version of its database, which can span a small group of servers. But in that situation, operating systems run on each server instead of spanning several, requiring that higher-level software be adapted to the cluster foundation.
VirtuOS, based in Acton, Mass., has 32 employees and four patent applications. Since its founding in March 2003, it's raised $20 million in two rounds of venture funding from Highland Capital, Matrix Partners and Goldman Sachs.
It's not the only new entrant willing to take on the four giants of the server market, IBM, Dell, Sun and HP. Azul Systems is betting it can speed up Java servers, and Enck said he knows a handful of other start-ups planning on joining the fray.
VirtuOS' technology also will compete with virtualization technology from Microsoft and EMC subsidiary VMware.
VirtuOS plans to formally launch in February, with release version 1.0 of its software due in April.
Earlier in his career, Davis ran Digital Equipment Corporation's VAXClusters program to link several servers into a single machine. Most recently he was chief technology officer of Internet software company Mangosoft. His VirtuOS co-founder is Alex Vasilevsky, who was an engineer at now-extinct supercomputer maker Thinking Machines and most recently was chief technical officer of home media software maker Ucentric Systems.
Large servers use a design called symmetrical multiprocessing, or SMP, to link their chips together so a single operating system can span the chips. VirtuOS uses its own proprietary software and high-speed InfiniBand networking links from TopSpin Communications to create what Davis called "virtual SMP."
Others have dabbled with the idea before, including, in 2001, Dell, which scrapped its SMP-over-InfiniBand project before it ever came to fruition.
Davis said the virtual SMP technology is limited not by VirtuOS but by how big a server Linux can use today, about 16 processors.
If the VirtuOS software works as advertised, it could offer compelling price advantages. For example, a setup with eight reasonably powerful IBM x336 servers with dual Xeon processors costs roughly $32,000, whereas, according to IBM, buying that many processors in a single SMP system, a 16-processor x445, costs about $102,000.
But InfiniBand simply isn't fast enough to reproduce the performance of the high-speed interconnects that link processors and memory in a multiprocessor server, Illuminata analyst Jonathan Eunice said.
The delay, or latency, between when a processor tries to read data from memory and when that memory actually arrives is about 200 nanoseconds, or billionths of a second. InfiniBand latencies are about 8,000 nanoseconds, he said.
"The claim that you can get SMP over InfiniBand I don't think is supported by what the technology turned out to provide," Eunice said. "Latency matters."
Davis acknowledges that there's a performance penalty, but he counters that the performance of the input-output system that InfiniBand uses is improving faster than the performance of memory. And he adds, "this is a lot of where our intellectual property comes in."
In the long run he's confident the performance penalty will be worth the advantages of virtualization.
"Initially people say, 'I'm not going to pay that kind of performance hit,'" Davis said. "Then the hardware catches up, and nobody knows how we did without it."
2 commentsJoin the conversation! Add your comment