On Jun 21, 2007  10:20 -0300, Balagopal Pillai wrote:
>           I am using Lustre 1.6.0.1 with one OST and 20 clients in an 
> HPC cluster.
> The OST/MDT/MGS has a 16 channel 3ware 9650 using raid6.
> 
> *Now the problem - Lustre performance dips a lot when it comes to small 
> files. Please see the following fileop -f 5 test comparing NFS and Lustre -

We have had reports of read performance problems with 3ware in the past.

> Lustre -
>         
> Fileop:  File size is 1,  Output is in Ops/sec. (A=Avg, B=Best, W=Worst)
> .      mkdir  rmdir create   read  write  close   stat access  chmod 
> readdir link   unlink delete  Total_files
> A    5   1654    691    132  14228    719   4874   1987  32737   1718   
> 2506   1262   1340   1608          125
> 
> 
> NFS -
> Fileop:  File size is 1,  Output is in Ops/sec. (A=Avg, B=Best, W=Worst)
> .      mkdir  rmdir create   read  write  close   stat access  chmod 
> readdir link   unlink delete  Total_files
> A    5    177    594    459 380747 137392   2282   1219 444312    502   
> 1274    306    513    464          125

>        Could you please recommend any tunables to get a bit more 
> performance out of Lustre with lots of small files? Lots of small files 
> was bad in GFS too, but it was better than NFS though.

Try setting the debug on the clients to zero "sysctl -w lnet.debug=0".

> *Also the read performance of Lustre seems to be a little behind NFS. I 
> had /opt which has all the software for users moved to Lustre in the new 
> setup. But
> software like Matlab, Splus etc takes almost a minute to come up. The 
> second time is very fast though, maybe due to caching. So i am thinking 
> of putting /opt
> back to NFS. Is it possible to boost the read performance of Lustre a bit?

Turn off debug is the easiest.

> *One last question. In the older Lustre setup (version 1.4.5), i have 5 
> scsi drives one each as an OST for a single volume. The volume became 
> full. But df still reported that there is 27GB free.  There doesn't seem
> to be an lfs df option in that version of Lustre. So i couldn't see the
> individual utilization of each of the 5 OST. Is this a striping problem?

In the absence of "lfs df" use /proc to find per-OST usage on the client:

grep '[0-9]' /proc/fs/lustre/osc/*/kbytes*

Lustre 1.6 should be better at balancing the OST usage than 1.4.

Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.

_______________________________________________
Lustre-discuss mailing list
[email protected]
https://mail.clusterfs.com/mailman/listinfo/lustre-discuss

Reply via email to