[Beowulf] GPFS on Linux (x86)
shaeffer at neuralscape.com
Thu Sep 14 09:58:32 PDT 2006
On Thu, Sep 14, 2006 at 07:05:42AM -0600, Craig Tierney wrote:
> Karen Shaeffer wrote:
> >>>IBM GPFS
> >>>(Did I forget some).
> >>Lustre/SFS and Redhat GFS, certainly.
> >Might help put things in perspective a bit and adds a few
> >more to the list. Published about a year ago -- still fresh
> >in the context of filesystems.
> The website doesn't have any real-world experience. I would
> like to know things like 'Filesystem X has much better
> meta-data performance than Filesystem Y', or 'Don't try using
> netcdf files on Filesystem Z, because the performance will
> be awful'.
> Discussions about real-world use will help users get through
> the marketing fluff to find out what is really going on. Not
> everyone has a chance to test every single filesystem before
> choosing one.
You asked: "Did I forget some?"
I just offered up the link to Jeff's article, because it added
to your list.
I don't think the article has any marketing fluff in it, else I
wouldn't have spammed the list with the link. While it is just
general feature sets and limitations, they are very specific
to each filesystem and concise enough to not be considered
I believe it is those coarse grain features and limitations
that enable you to filter down the list to just those systems that
are well suited to your specific cluster environment. In the end,
every cluster filesystem is going to perform optimally in only one
environment -- even if they can be safely deployed in more general
circumstances. And those environments are dependent on storage,
scale, interconnect fabric, and your application's interaction with
this constellation of interdependent technologies. So, in this sense,
I think that survey is very helpful to folks who appreciate its
Of course, your point about real world reviews of filesystems would
be interesting and helpful to most folks on this list. But, then,
because of the above mentioned interdependencies on the entire
constellation of technologies involved, comparisons are invaribly
going to be about apples and oranges -- or require a massive
effort involving many sets of hardware -- or suffer a built in
bias towards a specific filesystem. A bias that that survey can
help one to appreciate up front -- before falling into such a
trap. (smiles ;)
Thanks for your comments. I appreciate them, and they have started
an interesting discussion. I hope it continues -- as I am very
Neuralscape, Palo Alto, Ca. 94306
shaeffer at neuralscape.com http://www.neuralscape.com
More information about the Beowulf