Lustre devel - Dec 2006 - [Bug 10950] mkdir takes 60+seconds

Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
https://bugzilla.lustre.org/show_bug.cgi?id=10950



Created an attachment (id=9180)
Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
 --> (https://bugzilla.lustre.org/attachment.cgi?id=9180&action=view)
ddn stats with 5.24e

We finally skipped an internal xt3 version and bypassed the build problem. Now
we are unning at cray internal 1.5.30 release.
 
I applied Andreas'' patch (attachment 9043) and it seems to have helped
quite a
bit. During the test we couldn''t get mkdir/touch/df operations delayed
to above

3 seconds. We could also keep the load in a fairly constant rate. The only
exception is that at the end when we decided to kill the jobs and quit, we got
some long delays, up to about 1 minute with df/touch. But that may be related
to
other issues. The ddn stats is attached here and syslog/lustre stats will be
attached next.

Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
https://bugzilla.lustre.org/show_bug.cgi?id=10950



Created an attachment (id=9181)
Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
 --> (https://bugzilla.lustre.org/attachment.cgi?id=9181&action=view)
syslog and lustre stats

Please comment on the stats. I do still see some slow journal commits in the
log
and long IOs. 

Andreas, do you think we can commit the patch? It does seem to help and I see
one  "precreate slow", is it good enough? Thanks,

Wally
-----

Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
https://bugzilla.lustre.org/show_bug.cgi?id=10950



It seems we started the stats a little after the tests and all the slow commits
are related to the starting of the tests. I just want to see if there is
anything alarmful bad.

We did a short test with the patch and ddn 5.24e again and the result is still
pretty good. We will test it overnight sometime as well as commit the patch.

Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
https://bugzilla.lustre.org/show_bug.cgi?id=10950

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Status Whiteboard|2006-12-21: updated DDN     |2006-12-31: patch landed on
                   |firmware and modified cache |b1_4 for Lustre 1.4.9;
                   |parameters have reduced     |requires confirmation of
                   |severity of problem; CFS    |landing on b1_5
                   |provided patch further      |
                   |improves results; fix to be |
                   |landed                      |


Comment Tags: Status Change

2006-12-31: patch landed on b1_4 for Lustre 1.4.9; requires confirmation of
landing on b1_5

Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
https://bugzilla.lustre.org/show_bug.cgi?id=10950

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Status Whiteboard|2006-01-04: patch landed on |2006-01-09: patch landed on
                   |b1_4 for Lustre 1.4.9; needs|b1_4 for Lustre 1.4.9;
                   |landing on b1_5             |landed on b1_5


landed on b1_5

Please don''t reply to lustre-devel. Instead, comment in Bugzilla by
using the following link:
https://bugzilla.lustre.org/show_bug.cgi?id=10950

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED


marked as FIXED.