Thanks!
The first result was indeed with the client on the node where the
virtual instance hosted the MDS. When I switched to put the client on
one of the machines where the virtual machine hosts one of the OSS
performance look a lot more what I would have expected (i.e., decent
scaling when going from one to two files). Guess my "hypothesis" was wrong!
/jon
On 02/16/2017 04:55 PM, Mohr Jr, Richard Frank (Rick Mohr) wrote:
On Feb 16, 2017, at 9:56 AM, Jon Tegner <[email protected]> wrote:
I have three (physical) machines, and each one have a virtual machine on them
(KVM). On one of the virtual machines there is an MDS and on two of them there
are OSS:es installed.
All system use CentOS-7.3 and Lustre 2.9.0, and I mount the file system on one
of the physical machines.
Which machine of you mount the file system on? Is it an OSS node or the MDS
node?
Anyway, when evaluating the performance (using FIO) I expected to see an increased
transfer rate when going from one to two simultaneous files in the tests, and I do see
this, but only for very small file sizes (around 0.2 GB). On the "non virtual"
file systems I have tested the effect of increased transfer rate as the number of files
are increased is on the other hand easily verified.
Have you tried mounting the file system on different nodes? This could help
determine if the problem is always the same or if it might be affected by the
type of node (MDS vs OSS) that is being used for the client.
--
Rick Mohr
Senior HPC System Administrator
National Institute for Computational Sciences
http://www.nics.tennessee.edu
_______________________________________________
lustre-discuss mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
