Re: Off-the-wall: WRF Model and PS3's (or Cell)

25 msg	Infiniband modular switches
5 msg	RHEL5 network throughput/scalability
9 msg	size of swap partition
13 msg	User resource limits
27 msg	A couple of interesting comments
15 msg	Barcelona hardware error: how to detect
7 msg	OFED/IB for FC8
14 msg	SiCortex experience anyone?
2 msg	Size of GlusterFS installations? Hints? Alterna...
1 msg	AMD releases Stream GPGPU SDK for Linux
3 msg	local and private net on one switch?
3 msg	Mix different drives in RAID (WD4000YR and WD50...
7 msg	Supercomputing Companies in the United Kingdom
4 msg	Re: TOE on Linux
10 msg	Capacity / Capability Computing
1 msg	[p2s2-announce] Reminder: P2S2 Workshop Deadlin...
5 msg	ECC Scrub, which setting?
19 msg	Computation on the head node

Off-the-wall: WRF Model and PS3's (or Cell)
\ Gerry Creager (16 May 2008)
. \ Bernard Li (16 May 2008)
. . \ Gerry Creager (16 May 2008)
. . . \ Bernard Li (16 May 2008)
. . . . \ Gerry Creager (17 May 2008)
. . . . . \ John Leidel (17 May 2008)
. . . . . . \ Gerry Creager (17 May 2008)
. \ Brian Dobbins (17 May 2008)
. \ Cédric Augonnet (17 May 2008)

1 msg	Re: TOE on Linux? (Prentice Bisbal)

Subject:	Re: Off-the-wall: WRF Model and PS3's (or Cell)
Group:	Beowulf
From:	Cédric Augonnet
Date:	17 May 2008

2008/5/16 Gerry Creager <gerry.creager>:
> Has anyone seen a citation or reference for actually running WRF on a PS3
> cluster? I've got an itch to deploy a 16-PS3 cluster (gigabit intersinnect
> <sigh>). That said, it's too much money to squander if there's not some
> indication it'd show improvement over a "normal" cluster.
>
> gerry

Hi Gerry,

I guess this is unlikely that PS3 would be that good for such a task
while having only around 200MB available once the OS is running. In
addition to the network card that is not really adapted for
clustering, the most important aspect is how to get some performance
improvement using the Cell ?

Perhaps i'm just stating something obvious there, but a cell is
nothing but a mere (rather old) PPC which comes with a bunch of
coprocessors (SPEs), if you do not explicitely port the code so that
it is offloaded on those coprocessors, there is no use in looking at
the Cell. Perhaps there are ways to greatly improve WRF by offloading
some compute intensive parts onto the coprocessors, but it *really*
requires a lot of non-trivial work, and as far as i know (please
correct me if i'm wrong) there is no such work done yet.

However, i have seen that WRF 3.0 should support some "GPU accelerated
microphysics" as an experimental feature; so perhaps some of those
compute intensive parts can be rewrote to be explicitely offloaded on
SPEs instead of GPUs ?

Eventually, PS3 does not seem to fit clustering requirement given how
small is its memory, and the poor network. And even if you were to use
some Blades from IBM (like QS21, with high speed networks and decent
memory size) there is still a big change that WRF needs a substancial
effort to get the least benefit from running on the Cell. More
generally, in case you did not see those before, Dongarra et al. wrote
some interesting papers about PS3 clustering [1][2].

Regards,
Cédric

[1] A Rough Guide to Scientific Computing On the PlayStation 3 -
http://www.netlib.org/utk/people/JackDongarra/PAPERS/scop3.pdf
[2] Limitations of the PlayStation 3 for High Performance Cluster
Computing - http://www.netlib.org/lapack/lawnspdf/lawn185.pdf

_______________________________________________
Beowulf mailing list, Beowulf
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf