On Thu, Apr 1, 2010 at 4:18 PM, Olivier Le Cam < Olivier.LeCam at crdp.ac-versailles.fr> wrote: > Hi - > > Thank you Raghavendra for detailed reply and information. > > > Raghavendra G wrote: > >> I am seeing write-behind aggregating large number of writes into a single >> large write on Ubuntu client, but not on lenny client. This is the cause >> of >> large number of frames on network. This can also be the cause of slower >> performance on lenny. >> >> Can you set "option enable-trickling-writes false" in write-behind on >> lenny >> and rerun tests. That will set write-behind to do more aggressive >> aggregation. We are eagerly waiting for your results :). >> > > This is just a bit better with this option disabled: > > * with "option enable-trickling-writes false": > > ~# dd if=/dev/zero of=/mnt/test_file count=262144 bs=1024 > 262144+0 enregistrements lus > 262144+0 enregistrements ?crits > 268435456 bytes (268 MB) copied, 72,7446 s, 3,7 MB/s > > Can you send glusterfs client logs with trace loaded above and below write-behind? > * and without it: > > ~# dd if=/dev/zero of=/mnt/test_file count=262144 bs=1024 > 262144+0 enregistrements lus > 262144+0 enregistrements ?crits > 268435456 bytes (268 MB) copied, 87,8484 s, 3,1 MB/s > > (each test repeated several times with very little variances: +/- 0.1 > MB/s). > > BTW, did you see the comparative results I have posted, using Deban/Lenny > on the same hardware but with either a 2.6.18 and 2.6.26 vernel version? It > seems that kernel version is very critical in the way write-behind is able > to do large aggregation. > > > However I am still investigating what is the reason behind write-behind on >> Ubuntu doing large aggregation but not on lenny. >> > > Thank's a lot! I guess many people will appreciate! > > Kind regards, > -- > Olivier > -- Raghavendra G
