[Linux-cluster] stress-testing GFS ?

[Michael Will]

Jure Pečar wrote:
> On Thu, 16 Mar 2006 13:39:17 -0500 "Boris Ostrovsky" <baostr at gmail.com> wrote:
>   
>> I'd be very interested to hear comments on the numbers and hopefully some
>> tuning suggestions.
>>     
>
> Rest assured that you're not the only one dissapointed by gfs performance.
>
> If redhat (or anyone else) put up some article explaining what kind of performance we can expect from clustered file systems, we'd at least know what to aim for.
> But without such info, we can only compare against local file systems and such comparisons only show gfs in a very bad light.
>   
At the bare minimum you can compare yourself to an unclustered single 
NFS server rather than a local filesystem.

In addition there are plenty commercial cluster filesystems out there 
that you can compare yourself too:
1. polyserve matrix server
2. ibrix
3. ibm gpfs
4. lustre
others?

On the open source side there is
1. ocfs2
2. pvfs2
3. older versions of lustre?
others?

Now the benchmark cases you would pick would be interesting as well. 
i.e. I have seen machines with local
sata raid that where benchmarked with bonnie++ and seemed to stream 
happily at close to 100MB/s
but when you actually where extracting a tar-file from across the 
network, it would slow down to 10MB/s
whereas copying that tarfile across the network into one large file 
would give you close to 100MB/s as well.

Conclusion: lots of unique file-creating makes the local filesystem 
performance tank in comparison to bonnie++ or other large
file benchmarks.

Michael Will
-- 

Michael Will
Penguin Computing Corp.
Sales Engineer
415-954-2822
415-954-2899 fx
mwill at penguincomputing.com