Home Messages Index
[Date Prev][Date Next][Thread Prev][Thread Next]
Author IndexDate IndexThread Index

Re: Windows in High Performance Computing

On Sep 30, 5:05 pm, Roy Schestowitz <newsgro...@xxxxxxxxxxxxxxx>
wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> ____/ ness...@xxxxxxxxxxxxxxxxxxx on Tuesday 30 September 2008 23:05 : \____
>
>
>
> > <Quote>
> > I am 'one of the few' people who has actually "used" Windows Compute
> > Cluster Server 2003 in a work environment (for running Computation
> > Fluid Dynamics programs).
> > Here are my experiences:
> > Firstly, basically it is just Windows 2003 server for AMD64. There is
> > an additional CD with the 'Compute Cluster' pack that must be
> > installed.
> > There is a graphical queue manager MMC snapin etc.
>
> > There are some current problems with Computer Cluster Server:
> > * The kernel is old, and as usual, there are troubles finding the
> > correct 64-bit driver for the on-board SCSI in our brand new Dell
> > nodes. Windows requires a floppy disk. None of the nodes have a floppy
> > drive, so we have to hook up a USB floppy, and locate a floppy disk!!
> > * Default update behaviour. You must be careful to make sure that
> > Windows update is disabled, because it will reboot that machines in
> > the middle of the night, killing hundreds of hours worth of jobs, and
> > driving users into a searing rage. Updates that require reboots are
> > far too frequent for production use systems. Updates at present
> > require additional IT management to manually install.
> > * Jobs randomly crash - avoid virus checkers. I think because our IT
> > department's amazing virus scanner decides to eat up all the non-paged
> > pool (i'm still investigating this one). Installing virus scanners on
> > compute nodes, where intensive disk access is the norm borders on
> > insanity in my opinion.
> > * Infiniband drivers - We have beta drivers from a major provider of
> > infiniband cards (OpenIB stack). They are, in their current state, not
> > usable. The compute cluster node management software faults repeatedly
> > in the user mode DLL of the OpenIB device drivers. We have never been
> > able to run a job with it over more than 2 nodes.
> > Generally, we use HP MPI, which is pretty solid over ethernet, but
> > because of bad drivers, can't currently be used over Infiniband. We'll
> > have to get updated drivers when I get time to look at this.
> > Saying this, we've also had issues in the past with Linux Infiniband
> > drivers, but these are usually resolved quickly by vendor, or by a
> > kernel update.
> > * Lack of a decent scripting language that is common on all platforms
> > requires that cygwin/bash/python be deployed to all nodes to allow
> > platform independent scripts to execute.
> > * Management of shared drives on Windows has always been a problem.
> > Some Windows programs just don't like UNC paths. Drive letter mapping
> > gets mixed up with domain policy, update scripts, user's personal
> > mappings, and our source control system's mappings. What a mess.
> > Windows NEEDS an automounter. Consequently, we tend to copy files
> > about. This is better for the performance of the programs that we are
> > running in many cases anyway.
> > * There are two Terminal Server licenses per node for administration,
> > so if your users need to be able to login to compute nodes (as they
> > frequently do with some programs), you will have to fork out for
> > expensive terminal server licenses.
> > * Few HPC applications actually support Windows compute nodes.
> > Save yourself some effort, and just deploy Linux on HPC nodes.
> > </Quote>
>
> >http://blogs.computerworld.com/where_windows_is_2_to_linux
>
> > Ballmer recently said that High Performance Computing is one area
> > Microsoft aims to have a large presence in.
>
> He also boasted about search ambitions for like half a decade. And where are
> they? Still declining, despite /paying/ people to use the thing.
>
> My colleagues tried Windows high perfor~1 and it sucked so badly that it was
> embarrassing. Like many other such clusters, there was an escape from it,
> despite Microsoft 'freebies'.
>
> http://groups.google.com/group/comp.os.linux.advocacy/browse_thread/t...
>
> - --
>                 ~~ Best of wishes
>
> "Linux is a very complete and sophisticated operating system. And there is a
> lot of work being done to improve it in and of itself, particularly to make it
> easier to use and easier for people to set up on their personal computers."
>
>                 --Paul Maritz, senior vice-president, Microsoft
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.4.9 (GNU/Linux)
>
> iEYEARECAAYFAkjivs4ACgkQU4xAY3RXLo6dLQCeNuX1qMebbIgZTg9NMw1bMQ18
> yE4An04Nn+BVP+fn3Ssu0ffE+fA3bgRu
> =g1W3
> -----END PGP SIGNATURE-----

Ballmer has actually made quite a few specific predictions lately----
someone should keep track of them, as most of them will no doubt turn
out to have the same quality as Gates' promise to eliminate spam
within 2 years (was that 2004?).  Among others, he plans to take a
major presence in search---in spite of the fact that Microsoft is
descending ever lower into the single digits in search.  And to kick
Apple in the smart phone market---good luck.  What's interesting is
that Ballmer is clearly very concerned about competition from Apple,
that is, real competition.  He made some statement about expecting the
old tried and true Microsoft methods to work, ie extend, extinguish
etc.  It's like he doesn't get it, that when he doesn't control a
monopoly, he's got to compete on the basis of quality.  It seems to be
an alien thought around Microsoft.

[Date Prev][Date Next][Thread Prev][Thread Next]
Author IndexDate IndexThread Index