Lustre discuss - May 2006 - Client performance

Hi Vasily--

On 8/24/2005 10:38, Kostryukov Vasily wrote:
> 
> Does anyone achieve or know, if Lustre client can get real file system
> read and write bandwidth equal or more than 4GBytes/sec per client node,
> using 5 (or more) Quadrics Elan 4 in the client node, with appropriate
> OSSs (for example SFS from HP)
In theory, yes -- the Lustre architecture was designed to scale in exactly
this way.

However, this is further than we have tried to go in practice; besides
requiring very serious hardware, it will likely require some specialized
tuning and some efforts to remove bottlenecks in the Lustre client or Linux
kernel.

We did a similar exercise to reach 2.5 GB/s with 1 server and 2 clients last
year; you can read about at http://www.clusterfs.com/pr/2004-11-11.html

Regarding SFS, I don''t know what level of performance HP can provide
with a
single OSS -- you would have to ask them.  But no matter which vendor you
choose, a single client can do parallel I/O to as many servers as you need
to reach your target performance.

Hope this helps--

-Phil

Hi, Phil,

> In theory, yes -- the Lustre architecture was designed to scale in
exactly
> this way.
> However, this is further than we have tried to go in practice; besides
> requiring very serious hardware, it will likely require some
specialized
> tuning and some efforts to remove bottlenecks in the Lustre client or
> Linux
> kernel.
Well, so can we achieve 4GB/s for client (we need to get 4GB/s per
client but not simultaneously for all clients, so OSSs total throughput
can be only 4GB/s total) if we will use for example 6 Quadrics rails per
client (and 6 rails per OSS too, due to Quadrics multirail rules)?
We suppose using fat computing and storage nodes, like HP Integrity
rx8620-32.

And the second question is - do we really must to do kernel and Lustre
tuning for it?

Best regards,
Vasily

Hi,

 

Does anyone achieve or know, if Lustre client can get real file system
read and write bandwidth equal or more than 4GBytes/sec per client node,
using 5 (or more) Quadrics Elan 4 in the client node, with appropriate
OSSs (for example SFS from HP)

 

Best regards,

Vasily