The Windows Observer--Supercomputers' Need for Speed Satisfied with Windows HPC Server '08

Newsletters	Subscriptions	Forums	Safari	Store	Career	Media Kit	About Us	Contact	Search	Home

Volume 6, Number 25 -- June 25, 2008

Supercomputers' Need for Speed Satisfied with Windows HPC Server '08

Published: June 25, 2008

by Alex Woodie

Microsoft impressed the supercomputing community last week when a beta of its new Windows High Performance Computing (HPC) Server 2008 scored a top 25 ranking on the Top 500 list of world's biggest supercomputers. The new HPC Server 2008, which is due to hit the first Release Candidate stage by Monday, has also been clocked at 2 microseconds of latency between nodes using the new NetworkDirect interface and InfiniBand cards and switches--an impressive feat for any supercomputer.

Microsoft is a newcomer to the field of high performance computing, and still looking to make its mark. The first HPC version of its Windows Server operating system, Windows Compute Cluster Server 2003, debuted in August 2006, and the company still doesn't talk about how many users it has.

Sites running either Windows CCS Server 2003 or early release versions of its follow-on, Windows HPC Server 2008, register barely a blip on the radar of the world's biggest supercomputers, which is maintained at www.top500.org. The latest list shows Linux HPC deployments dwarf Windows deployments 85 to 1 (the list had five Windows installations, compared to six on the November list). And while IBM POWER-based systems account for 33 percent of the Top 500's systems' HPC capacity as a whole, compared to 62 percent for X64-based processors from Intel and AMD, POWER-based systems dominate the top of the field, and gets the bulk of the glory.

And the funny bit about all this is: Microsoft doesn't really care (or at least it says it doesn't). That's because it has a different vision for where the HPC market is going, as Bill Gates outlined two-and-a-half years ago. The company is hoping to revolutionize HPC systems, and make them easier to use and less expensive to buy. If it succeeds in that vision--and it appears it's on its way to getting there--then the profits and acclaim will follow, according to Ryan Waite, group program manager for Windows HPC Server 2008.

"Our goal is to certainly be on the Top 500 list, but not to be at the very top," Waite said in an interview. "There are some people that joke that the top five systems in the Top 500 list are a $0 billion a year market. They're all sold at cost. They're kind of ego-systems. Microsoft wants to be successful in the Top 500, but really bring it home to everybody else. It's not about the pinnacle of computing, it's about everybody being able to use HPC resources."

With that said, the world's largest software maker is not averse to a little ego-massaging, every now and then. And it got that when it bagged the number 23 slot on the recent Top 500 list with a 68.5 teraflop cluster of Dell PowerEdge servers at the National Center for Supercomputing Applications (NCSA) at the University of Illinois at Urbana-Champaign. That system is dwarfed by the current leader--the Department of Energy's hybrid POWER- and AMD-based "Roadrunner" cluster that churned out more than 1 million teraflops--but it's still the fastest Windows system on the planet.

There's a good chance Microsoft will better that ranking in the upcoming months, as Windows HPC Server 2008 hits the RC1 and RC2 milestones and eventually is released to manufacturing, which is slated to occur by the end of the year. That's because the NCSA cluster is running a beta release of HPC Server, one that doesn't support the latest NetworkDirect Remote Data Memory Access (RDMA) network interface that is supported in later versions.

RDMA may be the biggest thing to hit HPC computing in a while. The technology, which Microsoft co-developed with Mellanox Technologies, enables nodes of a cluster to communicate at latency levels of two microseconds on a 2 GB per second line, according to Microsoft.

"It's incredibly fast," Waite says. "Imagine if you had regular Gigabit Ethernet connection, a great switch and no contention, you would probably be looking at 30 to 40 microseconds of latency." The breakthrough will enable HPC Server clusters equipped with latest RDMA technology and InfiniBand switches and cards to achieve latency levels that are slightly better than the fastest systems currently out there, Waite says.

RDMA-equipped HPC clusters will simply accomplish more work with the same processing power, by reducing network contention. "Instead of copying data from the user mode into kernel mode and then sending it over the network cable," Waite says, "RDMA is a trick where we're able to pull the data right out of user mode, right out of the application, and throw it over the network to the other side."

RDMA may be the big new feature in HPC Server 2008, but there are several other enhancements worth noting. For one, the operating system's job scheduler will now be able to support service-oriented jobs, in addition to the regular batch-job style of submitting work to HPC systems.

The new service oriented job scheduler will change how organizations use HPC systems, Waite says. "We've heard in talking with customers that they want to have a more interactive model, and they want to be able to run service oriented applications across the cluster," he says. "Service oriented models are great because you make a call to service and the thing that provides the service says, 'Please compute this value to me.' It gives you the result back. It's very fast. That's perfectly suited for these service oriented architectures."

The service oriented job scheduler fits perfectly into Microsoft's vision for HPC. "HPC is no longer this special technology that sits off in a special room where only a small group of users have access to. It's really becoming a part of the general business community, and part of general enterprise computing," Waite says.