[PVFS2-users] Re: PVFS2 vs. NFS performance
rross at mcs.anl.gov
Tue Jun 29 12:35:50 EDT 2004
On Tue, 29 Jun 2004, Robert Latham wrote:
> On Tue, Jun 29, 2004 at 08:21:05AM -0700, Charles Shen wrote:
> > Overjoyed with successful PVFS2 installation, I decided to run some
> > simple test to compare it with NFS. My machines are PCs with 2.6 GHz
> > Celeron CPU, 512 MB memory, 100 Mb Ethernet connection, and Fedora
> > Core 2. I used Linux "dd" command for file read and write, and
> > "time" for duration. The file size is 1GB to avoid files being
> > cached.
> > Apparently with the single client test conditions, PVFS2 is much slower
> > than NFS, 30% for write, and 65% for read.
> Two factors here: dd isn't very kind to PVFS until the block size
> reaches 1MB or more, and NFS is much more agressive with caching than
> PVFS2 is.
It's not just the NFS is more aggressive with caching, it has completely
different consistency semantics. We do have some ongoing work here that
will implement similar consistency semantics for PVFS2, at which point
we will be playing on a level field, so to speak.
> > I know PVFS2 is for parallel computing. But, is there anything I can
> > do to make it run faster in my setup?
> Use a bigger block size for dd and see if that makes things better.
Also, if this is the sort of workload that you expect, and the NFS
consistency semantics are adequate for what you're doing, then you should
just use NFS. There's no point in throwing PVFS2 at a serial workload,
particularly one with small accesses.
More information about the PVFS2-users