Lustre discuss - May 2006 - Questions, Metadata performance :)

On Feb 09, 2006  09:13 +0100, Christoph Biardzki wrote:
> we''ve been running lustre at our site for a year and we are quite
pleased
> with the results (see http://www.lrz-muenchen.de/~biardzki/lustre.html
> for some details).
Interesting.
> 1) Under high write loads we see messages like this - what do they mean?
> 
> Wed Feb  8 19:24:54 2006: Feb  8 19:23:02 lx32s235 kernel:  
> LustreError: 5957:0:(filter_io.c:123:filter_grant_incoming())  
> OST_lx32s235: cli bc70f_lov1_c9c00c21a2/f77c3580 reports 16384  
> dropped > fed_grant 0
These are non-critical messages, and have been quieted in the upcoming
1.4.6 release.  This can either be a symptom of the OST having been
rebooted at some time in the past (without client shutdown) or RPCs
being handled on the server out-of-order.
> 2) Which options can we tune to optimize metadata performance with  
> small files (for a benchmark run)?
Smaller stripe counts generally improve metadata performance.
Increasing the journal size (though you likely already have the
maximum 400MB internal journal size).
Having clients do operations in separate directories.
> 3) I remember there were some plans to integrate patches into linux  
> to minimize locking when many parallel metadata operations are  
> performed on a single directory. Are these patches already in Lustre?
This is called "pdirops" (parallel directory operations), and has been
prototyped.  However, there are currently no plans to integrate this
into a production release in the near future.

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

Hello,


we''ve been running lustre at our site for a year and we are quite  
pleased with the results (see http://www.lrz-muenchen.de/~biardzki/ 
lustre.html for some details). But of course there are still some  
open questions:


1) Under high write loads we see messages like this - what do they mean?

Wed Feb  8 19:24:54 2006: Feb  8 19:23:02 lx32s235 kernel:  
LustreError: 5957:0:(filter_io.c:123:filter_grant_incoming())  
OST_lx32s235: cli bc70f_lov1_c9c00c21a2/f77c3580 reports 16384  
dropped > fed_grant 0

2) Which options can we tune to optimize metadata performance with  
small files (for a benchmark run)?

3) I remember there were some plans to integrate patches into linux  
to minimize locking when many parallel metadata operations are  
performed on a single directory. Are these patches already in Lustre?


Cheers,

  Christoph



-- 
Leibniz Rechenzentrum M?nchen (LRZ)
http://www.lrz.de
Abteilung Hochleistungssysteme
Barer Str. 21 - 80333 M?nchen - Germany
Tel. ++49-(0)89 / 289-28853, Raum S1527