2 page Case Study - Posted 6/23/2009
Views: 401
Rate This Evidence:
University Finds Windows HPC Server 26% Faster than Linux, More Cost-Effective
When researchers at the NSF Center for Autonomic Computing at the University of Arizona investigated the technology to set up a cluster for autonomic computing research that could largely monitor, manage, and repair itself, they faced a choice between Linux and Windows® HPC Server 2008. They chose Windows. Benchmark testing has confirmed that choice, showing Windows HPC Server to be 26 percent faster than Linux.
Business Needs
According to Dr. Salim Hariri, Director of the National Science Foundation (NSF) Center for Autonomic Computing (CAC) at the University of Arizona, the increasing size, diversity, and complexity of computer systems, combined with the relative shortage of skilled IT workers, has resulted in systems whose control and timely management exceeds human capacity.
Hariri and his colleagues want to do something about that. Their NSF CAC is dedicated to advancing the science of autonomic computing; that is, IT infrastructures and applications that largely monitor, manage, and repair themselves.
One of the first questions that Hariri and his colleagues faced at the center was choosing the platform on which to conduct their research. Linux, the traditional answer for large, distributed parallel computing systems, seemed the obvious choice—but perhaps not the right one.
“Autonomic computing requires significant amounts of control and management functionality to be added to a typical computing platform,” says Hariri. “Linux may still be thought of as the faster system, but it doesn’t offer the integrated services that would give us a leg up on building and testing autonomic systems. Moreover, if we want to have the greatest impact on Industry and government, we need to base our research on systems popular in those sectors—platforms that can be adopted without a major need to ‘rip and replace’ what those users already know and use.”
Solution
To meet these requirements, the NSF CAC researchers at the University of Arizona built their autonomic cluster based on the Windows® HPC Server 2008 operating system, Microsoft’s high-performance computing platform. “In Windows HPC Server 2008, we have a computing environment that provides significant elements of functionality needed for autonomic computing, so the effort to build out the computing environment is significantly less,” says Hariri.
The center is testing its theories about autonomic computing on a compute cluster system with 48 cores, including an 8-core HP 380 head node and five additional HP 360 compute-node computers with 8 cores each. The compute nodes were configured from Windows HPC Server templates created on the head node and pushed out to the compute nodes using the highly visual Windows HPC Server console.
The Windows HPC Server 2008 compute cluster was deployed in a day, with the university’s existing Active Directory® services structure extended to give researchers appropriate single sign-on rights to the new computing resource.
Two notable applications running on the Windows HPC Server cluster are a 3-D high-performance parallel electromagnetic (EM) simulation software system and a Kepler workflow application. The EM application is designed for distributed parallel computations of large EM problems such as those involved in aircraft design. The Kepler workflow application is popular with scientists and engineers, who can create large-scale applications by using block diagrams instead of writing line-by-line code.
To confirm its operating system choice, the NSF CAC benchmark tested the Kepler workflow application on Windows HPC Server against a similarly configured Linux system. The Kepler workflow application ran an average of 26 percent faster on Windows HPC Server than on Linux. “There’s a widespread perception—or misperception—that Linux is faster than Windows,” says Hariri. “Showing that Windows is faster than Linux for this widely used application is very important. It shows that researchers can be more productive on a system that is also more cost-effective to run.”
Similarly, the power of a Windows HPC Server cluster was demonstrated by testing the EM solution against a single-node system. The benchmark application ran in 18 hours, versus a much smaller application (just one-sixteenth the size of the larger application) that took 30 hours to run on a single-core system.
Benefits
“Researchers want their applications to run as quickly as possible, and our benchmark testing showed that for high-performance, distributed parallel computing, Windows HPC Server outperformed Linux,” says Hariri.
The advantages of Windows HPC Server extend to the setup and administration of the cluster, as well. “I have extensive experience with Linux,” says Leo Enfield, Computing Manager, NSF CAC, “and we couldn’t have set up a similarly sized Linux cluster in a day. Windows HPC Server gives us deployment and administration tools not available in Linux, and they’re the familiar tools we already know and use in Windows Server® environments. Basically, we did everything we needed to from the Windows HPC Server console, instead of having to go out to the individual computers. The console really helped—and we wouldn’t have had it with any other technology.
“We also benefit greatly by being able to use the same Active Directory services that we already have in place, and our researchers benefit because they don’t have to sign on specially for the compute cluster. If they’ve signed on to the network, the compute cluster acts as just another resource.”
Windows HPC Sever also beat Linux when it came to the availability of documentation and resources, making it easier to address problems that arose in the system. Enfield estimates that a similarly configured Linux cluster would require at least 20 percent more administrative time than he spends on the Windows HPC Server cluster.
The faster and easier administration of the cluster also changes the nature of the researchers’ work, according to Enfield. “You can experiment, decide something’s not right, and go back and change it without worrying how it will change the base image. That means researchers can work more freely and try more iterations of their work.”
Some benefits are less staid. “I like seeing Microsoft getting into the high-performance space, handling the same challenges we typically handle in Linux” says Enfield. “This is one of the more fun projects I’ve done.”
