On Friday 21 February 2003 08:25, besus wrote:
> Dear Daniel,
>
> Why do you think it unusual approach ?
> Is Microsoft Windows unsuitable for HPC?
>
> I am guessing that high-performance Computing is possible
> when many computers connected on network compute in palallell with .NET.
Fortran MPI clusters of up to 16 CPU's have shown similarly good performance,
either with Windows (2K server master, XP Pro slave), or linux. Sometimes,
the higher latency of the message passing phases in Windows is more than
offset by the willingness of Windows to let a single Fortran task, with
self-contained libraries, hog a CPU. Server 2003 (AKA .net) has not so far
equalled the performance of the current Windows versions, nor have Windows
clusters of more than 16 CPU's equalled performance of linux. No production
Windows or linux has dealt with the possibilities of task scheduling on
HyperThreaded MPI nodes, or using all the logical processors without
excessive message passing overhead. At 16 CPU's, there is a significant
additional price, monetary as well as in setup effort, in using Windows
rather than linux, and beyond that, the licensing fees for MPI also are
larger on Windows. Check the Cornell Theory Center web site; they show
Windows MPI clusters up to 128 CPU's installed, but if you can find practical
results there from the larger clusters, you're doing better than I.
Likewise, if you can find any software versions less than 2 years old
mentioned.
--
Tim Prince
|