[Swift-devel] Swift-issues (PBS+NFS Cluster)

Tim Freeman tfreeman at mcs.anl.gov
Tue May 12 15:38:52 CDT 2009


On Tue, 12 May 2009 15:19:14 -0500
Ioan Raicu <iraicu at cs.uchicago.edu> wrote:

> Hi Tim,
> 
> Tim Freeman wrote:
> > Is that for speed or less moving parts?  I think EBS is the fastest option
> > they have for disk space (faster than local disk), fyi.
> >
> >
> >   
> Can you elaborate more on how EBS, can be faster than local disk? I know 
> nothing about EBS, but there are only 2 ways it can work. Since it is 
> persistent, it likely lives on S3 (or something similar). When it is 
> mounted, perhaps it still lives remotely, or perhaps it gets copied to a 
> local disk, and runs locally. A locally running EBS, should be 
> comparable in speed to a raw local disk, perhaps a bit slower for yet 
> another layer of abstraction. If EBS lives remotely, perhaps in a 
> completely idle EC2 cloud, a remote EBS might perform better than a 
> local disk (assuming network communication is lighter-weight than 
> SATA/PATA bus operations), but I can't imagine how this can hold true in 
> a large scale and loaded cloud scenario, where shared infrastrucutre 
> (network, S3, etc) can become congested under load.
> 
> I don't see how EBS can be faster than local disk. Can you elaborate 
> more on this claim?

They state that it is purely network based (they do not list the technologies
they use but this could be infiniband, iscsi, etc.).  It's not uncommon to see
a SAN etc. faster than a local disk...

I made my statements based on Amazon documentation:

"The latency and throughput of Amazon EBS volumes is designed to be
significantly better than the Amazon EC2 instance stores in nearly all cases." 

I just googled and found this person showing that EBS wins on medium and high
powered instances (although someone comments at the end that 'dd' tests are not
the best thing to measure why EBS is better):

http://developer.amazonwebservices.com/connect/message.jspa?messageID=125197

Tim



More information about the Swift-devel mailing list