Lustre devel - Feb 2007 - [Bug 11268] wide striping assessment

Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
https://bugzilla.lustre.org/show_bug.cgi?id=11268



WIDE STRIPING STATUS, 2007-01-31
~~~~~~~~~~~~~~~~~~~~

GENERAL STATUS AND NOTES:

The latest versions of all spreadsheets have been sent to
lustre-discuss.  These (hopefully) contain enough information to
reproduce the results.

In future, it should be possible to use plot-sgpdd to plot results
instead of Excel.  It seems to work for sgpdd-survey results, and is
being extended to work for obdfilter-survey results.


DDN S2A 9550 TESTING:

Several sgpdd-survey runs have been completed.  With the right
settings, including 2-tier LUNs and 4 MB IOs, reasonable performance
has been measured.

Work with obdfilter-survey was started but didn''t get very far -
kernel patches are required to enable 4 MB IOs, and the patches in bug
9945 do not seem to be a complete set.

DDN suggests trying other cache segment sizes (all runs were performed
with 1024 KiB) - 512 is apparently good on SATA and 2048 is apparently
good on FC.

We also need to test more than one port at once.  I suggest testing
all 8 ports on a couplet at once.  Hopefully, performance will still
be stable but if not, try smaler sets of ports to discover where the
problem is introduced.

Braam suggests testing with "many, many more" regions or objects also.


THUMPER TESTING:

sgpdd-survey has been run.  Results are not great, but match the best
we have seen elsewhere.

When running obdfilter-survey, there were several problems.  I''ve sent
statistics and data to Alex for analysis (/proc/mdstat, vmstat 1,
/proc/fs/ldiskfs/*/mb_history, brw_stats for every involved OST.)

Note that next time OSTs are up on Thumper would be a good time to
look into bug 11484 (see comment 4.)