[gpfsug-discuss] Performance Issues with SMB/NFS to GPFS Backend

[Howard, Stewart Jameson]

Hi All,


I have a two-site replicate GPFS cluster running GPFS v3.5.0-26. We have recently run into a performane problem while exporting an SMB mount to one of our client labs.


Specifically, this lab is attempting to run a MatLab SPM job in the SMB share and seeing sharply degraded performance versus running it over NFS to their own NFS service. The job does time-slice correction on MRI image volumes that result in roughly 15,000 file creates, plus at lease one read and at least one write to each file. Here is a list that briefly describes the time-to-completion for this job, as run under various conditions:


1) Backed by their local fileserver, running over NFS - 5 min

2) Backed by our GPFS, running over SMB - 30 min

3) Backed by our GPFS, running over NFS - 20 min

4) Backed by local disk on our exporting protocol node, over SMB - 6 min

5) Backed by local disk on our exporting protocol node, over NFS - 6 min

6) Back by GPFS, running over GPFS native client on our supercomputer - 2 min


>From this list, it seems that the performance problems arise when combining either SMB or NFS with the GPFS backend. It is our conclusion that neither SMB nor NFS per se create the problem, exporting a local disk share over either of these protocols yields decent performance.


Do you have any insight as to why the combination of the GPFS back-end with either NFS or SMB yields such anemic performance? Can you offer any tuning recommendations that may improve the performance when running over SMB to the GPFS back-end (our preferred method of deployment)?


Thank you so much for your help as always!


Stewart Howard

Indiana University

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gpfsug.org/pipermail/gpfsug-discuss_gpfsug.org/attachments/20160721/7195ed93/attachment-0001.htm>