HPC on Windows – Does It Still Make Sense?

Print Friendly, PDF & Email

Several years ago, Microsoft made its first serious foray into the high performance computing market, offering Windows as an alternative operating system in the Linux-dominated space. When Microsoft backed off its commitment to HPC, Windows’ prospects dimmed significantly. But could there still be a place for the OS in high performance computing environments? To find out, we caught up with Jan Wender from German IT service provider science+computing, who will deliver a talk on this topic at the ISC High Performance conference next month in Frankfurt, Germany. In the interview below, Wender outlines some of the issues surrounding Windows in HPC and previews his upcoming presentation.

insideHPC: So why do you think Windows HPC failed to gain a significant foothold in the market? Was it a matter of Linux being so well-established in the HPC market, the fact the Windows HPC was a proprietary solution, or something else?

Jan Wender, science+computing

Jan Wender, science+computing

Jan Wender: I think we need to have a differentiated view here. On one hand MS HPC did not succeed in the “classical” HPC areas like CFD or structural simulations, but one other hand in the financial services area there are supposed to be more applications of MS HPC.

There are several reasons why this did not happen. One, maybe the foremost, is that there has been no clear performance benefit from using Windows HPC. If that would have been the case, then probably we would have seen a more massive adoption of it.

Lacking this, the advantage of using Windows HPC would have to come from other factors. Looking at “classical” HPC in academia, here most users are working with the source code of an application. This has been based on a Unix-like system for several decades now, so there is a lot of dependency on the Unix environment. It is not an easy task to change this to another quite different environment.

Looking at HPC applications in CAE and related fields, most have been Unix-based and then being ported to Linux, which is a lot easier than changing to Windows.

In the end, what has been missing is a clear and large enough benefit from using Windows HPC to make changing your environment worthwhile.

From what I have seen, most real-world usages of Windows HPC are smaller installations with rather specific applications only available on Windows (and then sometimes not even using the HPC pack).

insideHPC: Do you think Microsoft’s repositioning of their HPC product, now called HPC Pack, and its interoperability with the Azure cloud platform, makes it a more attractive platform for HPC users?

Jan Wender: Using HPC in a cloud is still an emerging topic, so there will be changes. But looking at the current state of affairs, moving from a Unix-based local environment to a Windows-based cloud environment is an even larger change, so I suppose that this will not be done often.

If you are coming from an Windows environment, and your HPC applications are running on Windows, then using Azure for Windows HPC is a logical extension. This should enable using HPC especially for smaller companies who need a larger amount of HPC resources only temporarily. Then the challenge is to have a painless integration which is made easier by products such as Azure Batch.

insideHPC: What types of HPC users are most likely to benefit from a Windows-based operating system and software stack?

Jan Wender: In the end, it depends on the availability of the applications you use. If they are not running on Windows HPC, well, then obviously using Windows is not an option.

If they are available (which seems to be slowly more often the case), then using Windows HPC would make sense if you are coming from a Windows environment. Otherwise, changing not only the operating system of your HPC environment but also all the user management and so on would need much more effort. If you are already Windows-based, then using Windows HPC is a relatively small extension.

insideHPC: In general, do HPC users tend to be too resistant to adopting new types of system software – operating systems, programming languages, cluster management tools, and so on? Can you think of any of this type of software that has become successful elsewhere, but is being ignored or underutilized in HPC environments?

Jan Wender: Well, HPC users are looking for performance foremost, the name is not for nothing ;-). If there is a new technology offering clear performance benefits it will get adopted. Maybe not at once, but over a few years. Case in point: Adoption of Beowulf Linux clusters. At that time HPC was done with large Unix-based machines when a few enthusiasts started using Linux. Then it was shown that you get similar or better price/performance on Linux clusters and the rest is (almost, by now) history.

Registration is now open for ISC 2015, which takes place July 12-16 in Frankfurt.

Sign up for our insideHPC Newsletter.