[Beowulf] MS HPC... Oh dear...

Vincent Diepeveen diep at xs4all.nl
Mon Jun 12 08:55:21 PDT 2006


>From: "Jim Lux" <James.P.Lux at jpl.nasa.gov>
> I'll just be waiting for the .NET Framework 2.x "cluster computing 
> namespace" with all the message passing primitives.

Oh such primitives will undoubtfully get casted onto the tcp/ip library 
using
the swapfile.

Actually before they have something like that it's helpful when windows
first takes care that allocating shared memory doesn't work via the swapfile
swapfile, but takes care that you can allocate REAL fast RAM that can be
SHARED between processes and removes another obvious bug.

Basically if your system has 2 GB ram, with 1GB for swapfile
and you try to start 4 processes, each process allocates and manages itself
a shared memory block of 400MB, and attaching to the 3 other 400MB blocks,
then that's gonna fail.

Reason: the OS calculates that 4 * 1.6GB = 6.4GB and that's bigger than 2GB 
ram = 1 GB swap,
so it can not possible attach all segments to each other as your system 
doesn't have 6.4GB according
to windows.

Besides, reading and writing to those shared memory blocks is extremely slow 
regurarly,
like a factor 1000 slower than it should be, for whatever reason (perhaps a 
bad memory manager?).

Nowadays you can't enforce windows to physically put that shared memory in 
RAM with
VirtualLock() function, which worked so great for NT 4.0 / windows2000 and 
no longer works
for XP, so that sucks too.

Note that microsoft can't say they do not know about this problem.

They simply won't solve it.

Windows is not so perfect for very well scaling applications that use quite 
some RAM too.

Note that multithreading using some RAM and 8 cores is not easy under 
windows either.
Basically hits a similar problem within windows.

Vincent

----- Original Message ----- 
To: "SIM DOG" <steve_heaton at iinet.net.au>; <beowulf at beowulf.org>
Sent: Monday, June 12, 2006 2:30 PM
Subject: Re: [Beowulf] MS HPC... Oh dear...


> At 06:20 PM 6/11/2006, SIM DOG wrote:
>>G'day all
>>
>>Sorry if this turns out to be a dupe post but MS has just released their 
>>HPC clustering kit.
>>http://www.microsoft.com/windowsserver2003/ccs/overview.mspx
>>
>>While I've tried to approach this with an open mind... it didn't last 
>>long. I'll refer anyone to ClusterMonkey's article about why Linux is a 
>>better platform for clustering.
>>
>>Maybe the equiv MS site should be called ClusterMoney? =)
>
>
> We had an interesting discussion about MS's clustering ideas a few months 
> back on the list (Nov 2005).
>
> Being the big boy on the block, MS has to have a presence everywhere, if 
> only to support some decision maker who's running down the list of check 
> boxes (Hmm.. Security, Check. Shared Drives, Check. Database server, 
> Check. Cluster computing, Check.).  Doesn't have to work as well as other 
> solutions, just has to exist.
>
> And, if you have an ALL Microsoft shop, AND you have tasks that 
> Microsoft's solution works for, then why not use it.  Who knows, someone 
> might come up with clever stuff to use their implementation in an 
> effective way.
>
>
>
>
>
>>Whatever you do, don't read this article as the level of marketing-speak 
>>and Ra Ra may well impede synapses firing...
>
> Such articles are fairly common in the weekly trade rags.. While I haven't 
> actually seen one, I wouldn't be surprised to see something like:
>
> Gizmotron Widgets today announced the model 1303 entertainment adjunct, 
> based on Gizmotron's patented "side to side orbital motion" swing ( US 
> Patent 6368227 ). Fred Harris, Gizmotron's CTO, announced that by using 
> their recent proprietary research methods into the principle that moving 
> objects tend to remain in motion and stationary ones require a force to 
> start moving, they have developed a new twist on an age-old children's 
> pastime. Seeking their first round of funding, Bob Smith, CEO, said "We're 
> really excited about this, and as soon as we work out the legal issues 
> surrounding swinging rights management for our content providers, we'll be 
> licensing our technology to several top-tier manufacturers."
>
>
>
>
>
> James Lux, P.E.
> Spacecraft Radio Frequency Subsystems Group
> Flight Communications Systems Section
> Jet Propulsion Laboratory, Mail Stop 161-213
> 4800 Oak Grove Drive
> Pasadena CA 91109
> tel: (818)354-2075
> fax: (818)393-6875
>
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit 
> http://www.beowulf.org/mailman/listinfo/beowulf
> 




More information about the Beowulf mailing list