[Sheepdog] Synchronization thoughput

[MORITA Kazutaka]

At Thu, 29 Sep 2011 17:21:47 -0300,
Rubens Matos wrote:
> 
> I have been performing some tests concerning the scalability of sheepdog.
> I've noticed that the throughput of write operations does not increase
> beyond 6~7 MB/s, in a two-nodes cluster. The workload is being generated to
> a VM running on a sheepdog VDI.

If barrier is enabled on your filesystem, the sync performance could
be terribly bad.  What filesystem and what mount option are you using?

Perhaps, the following thread would help you:
  http://lists.wpkg.org/pipermail/sheepdog/2011-April/000942.html

> 
> Do you have any information about the maximum synchronization throughput (in
> MB/s) obtained using sheepdog?

In my environment (124 physical machines, each has a 1 TB SATA disk),
the write throughput with 512 KB buffer size is:

-------------------
Copies   Throughput
-------------------
     1   34.92 MB/s
     2   21.82 MB/s
     3   19.95 MB/s
-------------------

'Copies' means the number of data redundancy.  Note that Sheepdog
supports only a write-through mode, so written data is always synced
to the disk.

> In the KVM Forum slides, there is a bar plot
> showing around 15 MB/s for the throughput, but I think this is the total
> rate, considering both read and write operations performed by the "dbench"
> benchmark, am I right?

Yes, the result is measured with dbench.


Thanks,

Kazutaka


> 
> Thank you in advance for the attention,
> 
> --------
> Rubens de Souza Matos Júnior
> PhD Candidate at UFPE
> MoDCS Research Group - www.modcs.org
> -- 
> sheepdog mailing list
> sheepdog at lists.wpkg.org
> http://lists.wpkg.org/mailman/listinfo/sheepdog