Slashdot Mirror
Windows Compute Cluster Server 2003 Released
grammar fascist writes "According to an Information Week article, on Friday Microsoft released Windows Compute Cluster Server 2003." From the article: "The software is Microsoft's first to run parallel HPC applications aimed at users working on complex computations... 'High-performance computing technology holds great potential for expanding opportunities... but until now it has been too expensive and too difficult for many people to use effectively,' said Bob Muglia, senior vice president of [Microsoft's] Server and Tools Business unit, in a statement."
and what about the site licence needed for this baby, huh? For us mere basement-cluster builders, there is a cheaper, open source alternative: The OSCAR Project ( Open Source Cluster Application Resources). Yes, it runs on Linux, but it is a nearly step-by-step system of setting up HPC-level clusters. It is being used on many 100+ CPU High Performance Clusters around the world, and it is free without those pesky site licences.
"...but until now it has been too expensive and too difficult for many people to use effectively..." According to their licensing model EVERY machine costs 469 dollars... Meaning a 20 machine cluster would have a 10,000 dollar overhead just on the OS alone. Not to mention the fact that you'd be compelled to buy it again as Longhorn Cluster Ed. in just a couple of years... It seems like a little work setting up a free OS cluster would be a vastly preferable option, is there really any need or reason for this (at this cost anyway)?
You can build an HPC from random PCs but it will be crap because the PC to PC interconnects will be too slow. Real HPC needs highspeed, low latency internal interconnects and these are expensive. But I fail to see how paying a "Windows" tax will make matters cheaper, or easier.
Not to many are using Fedora or Slackware on some white box with parts from Best Buy to do HPC. They have been altered to specifically run on hardware that was made specifically for this, and even then management of it is not exactly simple. Not that I believe that 2003 Server will suddenly change that but just using Linux somewhere does not automatically make it the cheapest way.
And I believe the correct answer to your question is Traditionally it has been done by tuned versions of commercial Unices which added to the base cost of the OS over and above the very expensive custom built hardware. Recently Linux has become able to do many of these tasks by similarly being modified at a significant cost running on the same expensive custom hardware. The recent HPC installation using mostly off the shelf parts (they didn't use Ethernet) was the one at Virginia Tech and that ran OS X, not Linux.
"I use a Mac because I'm just better than you are."
You can still have real HPC with slow interconnects. It all depends on the application for the HPC. If your data has a high scatter rate that requires large amounts of data transfer all the time, then you need fast interconnects. On the other hand, if your data can be sent off to a node to be crunched on for 2 hours, then a bog standard gigabit ethernet interconnect will do you just fine.
Difficulty, therefore, is NOT a significant factor in all of this. Ok, what about expense? Well, you're right that Linux is free. So is OpenMOSIX, OpenMPI (and many other MPI implementations), PVM (another messaging library), Lustre (a very high-performance network file system), many scientific and mathematical applications for clusters, etc. There are clustering patches for PoVRay, and it's always possible to write a script to have multiple machines render parts of images anyway. I'm sure there are other applications out there that I'm not thinking of right now, and it's only a matter of time before more "mundane" applications can take advantage of clustered environments. They already do, on Plan 9, to some degree. Oh, Plan 9 is also free.
Cost would appear not to be a major problem either, then. Optimizing is the only thing that is in any way difficult, and a GUI system that doesn't let you get to the really fine detail won't help there. More time, effort and money is spent on optimizing than on anything else, and I simply can't see any possible way that an OS that is designed for ease-of-use by hiding the intricacies can in any way help in that.
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
Ummm...I know of several clusters on our campus (VT) that are made of white boxes running Fedora, Gentoo, or Suse. One is a 200 node (400 CPU) Opteron cluster with a Myrinet interconnect named Anantham, and built by Dr. Varadarajan's graduate students. There are other smaller clusters ( 16 - 32 nodes ) of various design that are running on GigE. All of them built of white boxes and other off-the-shelf components ordered from mail-order companies. In the case of Anantham, all parts were ordered separately, i.e., RAM, motherboards, processors, cases, etc., and the system integration was done onsite. So, when you say, "Not to [sic] many are using Fedora or Slackware on some white box with parts from Best Buy to do HPC," I'm guessing you are referring to those in the TOP500 List? If so, yes, there aren't many that submit to the TOP500 List (from large sites) that are using a non-commercial version of Linux, i.e., RHEL. Many of the larger sites are going through first tier vendors (Dell, HP, IBM) for a turnkey cluster solution, but they are paying a premium for those systems for the sake of time to production. They could just as well buy white boxes, but they would be spending a great deal of their own time weeding out problem nodes and components that could be better spent on doing science and supporting users. Academia can afford to take the time to do this, DOE labs cannot, although, that paradigm is quickly shifting as academic budgets tighten and competition in the Computational Science and Engineering arenas heats up among research institutions.
Clusters (the topic of this original post) are not "traditionally [...] done by tuned versions of commercial Unices [sic]". Clusters are traditionally built with off-the-shelf components with Linux and specialized APIs and drivers for the interconnect being used. If you want to talk about HPC BEFORE 1998, then you are looking at large monolithic systems of a custom built nature.
System X does have a GigE network, but it is primarily used for management and job startup within the cluster. We have had a few users with specific MPICH2 needs that have used the GigE network for message passing, but the GigE network was not designed for that task. Our primary communication fabric is IB. We are currently running Mac OS X (10.3.9) on the system, but are evaluating alternatives.