[Beowulf] New FhGFS article

Andrew Piskorski atp at piskorski.com
Tue Jul 16 13:35:25 PDT 2013

On Mon, Jul 15, 2013 at 11:33:52AM -0400, Douglas Eadline wrote:

>   FhGFS: A Fast and Scalable Parallel Filesystem
>   by Tobias Goetz
>   http://goo.gl/C3XlK

That says that if you take 20 fairly high-performance each doing do
1.3 GB/s local disk IO, FhGFS scales linearly across them, giving you
about 25 GB/s total to your parallel application.  Interesting.

What about the more bargain basement case?  Say you have a cluster of
10 to 100 compute nodes each with only one or two SATA disks.  Can you
usefully run FhGFS across those thin nodes?  How well (or poorly) does
that work compared to using a smaller number of fat storage servers?

Andrew Piskorski <atp at piskorski.com>

More information about the Beowulf mailing list