| |||||||||||||
HPC computingHPC computingPosted Jul 20, 2006 10:11 UTC (Thu) by drag (subscriber, #31333)In reply to: HPC computing by vblum Parent article: Free Software Sets the Computing Agenda
I don't know.. I never done any real HPC stuff, just played around it at home.
If I can setup a Linux cluster then it can't be that difficult.
Also I've been told that for clustering projects especially, and HPC in general, that the best thing a operating system can do is simply get out of the way. The less impact the OS has the better.. You just need something to manage I/O (ie, fast drivers for the interconnects), setup the disks, initialize the hardware, bring up the node, and the rest is all managed at that application level with whatever custom thing they've setup.
Not that I know a lot about it personally.
As I understand it it's mostly fortran with some MPI libraries that people use for beowolf stuff. With Linux this is very easy. I expect a administration can probably setup a little ramdisk with everything you need on it to setup a node and load it over a network in a blink.
I couldn't imagine having to do that with something like Windows 2003. Just strip it down to a kernel and a few bare utilities and libraries... Is that even possible? I would have to devote a half a gig of ram and have a disk drive just to run the OS.
I know that Microsoft just getting into HPC is a bit of a farce. They've been working with institutions for years to get Windows-based clusters into the Top500 for many years now.. And they've consistantly been able to do it. The first time that I aware of was with a NT-based cluster in 1999. They've had various W2k machines in the Top500. Pretty much every top500 ranking since 2002 or so has had a Windows cluster in it. I haven't seen more then a specific university or two build them and those are specificly working with Microsoft for Windows cluster projects. Building clusters to prove that they can build clusters. Or so it seems to me at least.
(Log in to post comments)
HPC computing Posted Jul 20, 2006 14:36 UTC (Thu) by jschrod (subscriber, #1646) [Link] HPC means not only to run FORTRAN / MPI programs, but also to run applications like LS-DYNA (car crash simulation), NASTRAN (wind tunnel simulation), and others. It also means support of proprietary/modern/fast interconnect technologies like Myrinet or Infiniband. (My company does consulting in this area.)
While Windows is very bad in this area, Linux is better, but not really good either. I have seen several Linux HPC cluster projects in the automotive industry that got abandoned, at different companies, because the admins there were not able to get the stuff in a stable state. And these were admins with several years of experience managing HPC clusters. (Mostly AIX or other Unices like HP-UX or Solaris.) I know of other Linux clusters that were well publicized in the press, but that did not fulfill their promise during production, mostly due to problems with stability and throughput.
While you can easily get a small system to work at home, a real Linux cluster with a few hundred or even a few thousand nodes, _fast_ Interconnect technology (i.e., *not* GB-Ethernet), and good SAN access (EMC powerpath in Linux, anyone?) that runs stable over an extended period in time is hard to get by. And stability is demanded by our customers, when the design of a new car depends on modelling jobs being computed over night, every night.
Cheers, Joachim
HPC computing Posted Aug 10, 2006 23:04 UTC (Thu) by csamuel (✭ supporter ✭, #2624) [Link] Often that's the case of people selecting (or being forced to select) thewrong configuration for the codes they want to run. For instance you're not going to get very far with NASTRAN or Gaussian running RHEL with ext3 on a box with just 2 drives, they're IO hogs and need a fast filesystem.
LS-DYNA on the other hand is pretty well behaved, as long as you remember
|
|||||||||||||