Geoffrey Letessier
2015-Aug-06 23:32 UTC
[Gluster-users] Cascading errors and very bad write performance
Hi, No idea to help me fix this issue? (big logs, small write performance (/4), etc.) For comparison, here to volumes: - home: distributed on 4 bricks / 2 nodes (and replicated on 4 other bricks / 2 other nodes): # ddt -t 35g /home Writing to /home/ddt.24172 ... syncing ... done. sleeping 10 seconds ... done. Reading from /home/ddt.24172 ... done. 33792MiB KiB/s CPU% Write 103659 1 Read 391955 3 - workdir: distributed on 4 bricks / 2 nodes (one the same RAID volumes and servers than home): # ddt -t 35g /workdir Writing to /workdir/ddt.24717 ... syncing ... done. sleeping 10 seconds ... done. Reading from /workdir/ddt.24717 ... done. 35840MiB KiB/s CPU% Write 738314 4 Read 536497 4 For information, previously on 3.5.3-2 version, I obtained roughly 1.1GBs for workdir volume and ~550-600MBs for home. All my tests (CP, RSYNC, etc.) provides me the same result (write throughput between 100MBs and 150MBs) Thanks. Geoffrey ------------------------------------------------------ Geoffrey Letessier Responsable informatique & ing?nieur syst?me UPR 9080 - CNRS - Laboratoire de Biochimie Th?orique Institut de Biologie Physico-Chimique 13, rue Pierre et Marie Curie - 75005 Paris Tel: 01 58 41 50 93 - eMail: geoffrey.letessier at ibpc.fr Le 5 ao?t 2015 ? 10:40, Geoffrey Letessier <geoffrey.letessier at cnrs.fr> a ?crit :> Hello, > > In addition, knowing I have reactivated the log (brick-log-level = INFO not CRITICAL) only for the file creation duration (i.e. a few minutes), do you have noticed the log sizes and the number of lines inside: > # ls -lh storage* > -rw------- 1 letessier staff 18M 5 ao? 00:54 storage1__export-brick_home-brick1-data.log > -rw------- 1 letessier staff 2,1K 5 ao? 00:54 storage1__export-brick_home-brick2-data.log > -rw------- 1 letessier staff 15M 5 ao? 00:56 storage2__export-brick_home-brick1-data.log > -rw------- 1 letessier staff 2,1K 5 ao? 00:54 storage2__export-brick_home-brick2-data.log > -rw------- 1 letessier staff 47M 5 ao? 00:55 storage3__export-brick_home-brick1-data.log > -rw------- 1 letessier staff 2,1K 5 ao? 00:54 storage3__export-brick_home-brick2-data.log > -rw------- 1 letessier staff 47M 5 ao? 00:55 storage4__export-brick_home-brick1-data.log > -rw------- 1 letessier staff 2,1K 5 ao? 00:55 storage4__export-brick_home-brick2-data.log > > # wc -l storage* > 55381 storage1__export-brick_home-brick1-data.log > 17 storage1__export-brick_home-brick2-data.log > 41636 storage2__export-brick_home-brick1-data.log > 17 storage2__export-brick_home-brick2-data.log > 270360 storage3__export-brick_home-brick1-data.log > 17 storage3__export-brick_home-brick2-data.log > 270358 storage4__export-brick_home-brick1-data.log > 17 storage4__export-brick_home-brick2-data.log > 637803 total > > If the let brick-log-level to INFO, the brick log files in each server will consume all my /var partition capacity within only a few hours/days? > > Thanks in advance, > Geoffrey > ------------------------------------------------------ > Geoffrey Letessier > Responsable informatique & ing?nieur syst?me > UPR 9080 - CNRS - Laboratoire de Biochimie Th?orique > Institut de Biologie Physico-Chimique > 13, rue Pierre et Marie Curie - 75005 Paris > Tel: 01 58 41 50 93 - eMail: geoffrey.letessier at ibpc.fr > > Le 5 ao?t 2015 ? 01:12, Geoffrey Letessier <geoffrey.letessier at cnrs.fr> a ?crit : > >> Hello, >> >> Since the problem motioned previously (all errors noticed in brick log files), i notice a very very bad performance: i can note my write performance divided by 4 than previously -knowing it was not so good before. >> Now, a write of a 33GB file, my write throughput is around 150MBs (with Infiniband), before it was around 550-600MBs; and this, both with RDMA and TCP protocol. >> >> During this test, more than 40 000 error lines (as the following) were added to the brick log files. >> [2015-08-04 22:34:27.337622] E [dict.c:1418:dict_copy_with_ref] (-->/usr/lib64/glusterfs/3.7.3/xlator/protocol/server.so(server_resolve_inode+0x60) [0x7f021c6f7410] -->/usr/lib64/glusterfs/3.7.3/xlator/protocol/server.so(resolve_gfid+0x88) [0x7f021c6f7188] -->/usr/lib64/libglusterfs.so.0(dict_copy_with_ref+0xa4) [0x7f0229cba674] ) 0-dict: invalid argument: dict [Argument invalide] >> >> >> All brick log files are in attachments. >> >> Thanks in advance for all your help and fix, >> Best, >> Geoffrey >> >> PS: question: is it possible to easily downgrade GlusterFS to a previous version from 3.7 (for example: v3.5)? >> >> ------------------------------------------------------ >> Geoffrey Letessier >> Responsable informatique & ing?nieur syst?me >> UPR 9080 - CNRS - Laboratoire de Biochimie Th?orique >> Institut de Biologie Physico-Chimique >> 13, rue Pierre et Marie Curie - 75005 Paris >> Tel: 01 58 41 50 93 - eMail: geoffrey.letessier at ibpc.fr >> <bricks-logs.tgz> >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20150807/d330524a/attachment.html>
Geoffrey Letessier
2015-Aug-06 23:48 UTC
[Gluster-users] Cascading errors and very bad write performance
For additional information, if I [re]set the brick-log-level volume parameter to INFO, where what i can notice (thousands of thousands?) in my brick logs when I?m writing something: [2015-08-06 23:44:22.805293] W [marker-quota.c:3379:_mq_initiate_quota_txn] 0-vol_home-marker: parent is NULL for <gfid:621f5366-18fa-4d46-8796-4532643d94cd>, aborting updation txn [2015-08-06 23:44:22.805772] W [marker-quota.c:3379:_mq_initiate_quota_txn] 0-vol_home-marker: parent is NULL for <gfid:621f5366-18fa-4d46-8796-4532643d94cd>, aborting updation txn [2015-08-06 23:44:22.806081] W [marker-quota.c:3379:_mq_initiate_quota_txn] 0-vol_home-marker: parent is NULL for <gfid:621f5366-18fa-4d46-8796-4532643d94cd>, aborting updation txn [2015-08-06 23:44:22.806107] W [marker-quota.c:3379:_mq_initiate_quota_txn] 0-vol_home-marker: parent is NULL for <gfid:621f5366-18fa-4d46-8796-4532643d94cd>, aborting updation txn An idea? Thanks, Geoffrey -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20150807/cb5c3661/attachment.html>
Vijaikumar M
2015-Aug-07 04:28 UTC
[Gluster-users] Cascading errors and very bad write performance
Hi Geoffrey, Some performance improvements has been done in quota in glusterfs-3.7.3. Could you upgrade to glusterfs-3.7.3 and see if this helps Thanks, Vijay On Friday 07 August 2015 05:02 AM, Geoffrey Letessier wrote:> Hi, > > No idea to help me fix this issue? (big logs, small write performance > (/4), etc.) > > For comparison, here to volumes: > - home: distributed on 4 bricks / 2 nodes (and replicated on 4 other > bricks / 2 other nodes): > # ddt -t 35g /home > Writing to /home/ddt.24172 ... syncing ... done. > sleeping 10 seconds ... done. > Reading from /home/ddt.24172 ... done. > 33792MiB KiB/s CPU% > Write 103659 1 > Read 391955 3 > > - workdir: distributed on 4 bricks / 2 nodes (one the same RAID > volumes and servers than home): > # ddt -t 35g /workdir > Writing to /workdir/ddt.24717 ... syncing ... done. > sleeping 10 seconds ... done. > Reading from /workdir/ddt.24717 ... done. > 35840MiB KiB/s CPU% > Write 738314 4 > Read 536497 4 > > For information, previously on 3.5.3-2 version, I obtained roughly > 1.1GBs for workdir volume and ~550-600MBs for home. > > All my tests (CP, RSYNC, etc.) provides me the same result (write > throughput between 100MBs and 150MBs) > > Thanks. > Geoffrey > ------------------------------------------------------ > Geoffrey Letessier > Responsable informatique & ing?nieur syst?me > UPR 9080 - CNRS - Laboratoire de Biochimie Th?orique > Institut de Biologie Physico-Chimique > 13, rue Pierre et Marie Curie - 75005 Paris > Tel: 01 58 41 50 93 - eMail: geoffrey.letessier at ibpc.fr > <mailto:geoffrey.letessier at ibpc.fr> > > Le 5 ao?t 2015 ? 10:40, Geoffrey Letessier <geoffrey.letessier at cnrs.fr > <mailto:geoffrey.letessier at cnrs.fr>> a ?crit : > >> Hello, >> >> In addition, knowing I have reactivated the log (brick-log-level = >> INFO not CRITICAL) only for the file creation duration (i.e. a few >> minutes), do you have noticed the log sizes and the number of lines >> inside: >> # ls -lh storage* >> -rw------- 1 letessier staff 18M 5 ao? 00:54 >> storage1__export-brick_home-brick1-data.log >> -rw------- 1 letessier staff 2,1K 5 ao? 00:54 >> storage1__export-brick_home-brick2-data.log >> -rw------- 1 letessier staff 15M 5 ao? 00:56 >> storage2__export-brick_home-brick1-data.log >> -rw------- 1 letessier staff 2,1K 5 ao? 00:54 >> storage2__export-brick_home-brick2-data.log >> -rw------- 1 letessier staff 47M 5 ao? 00:55 >> storage3__export-brick_home-brick1-data.log >> -rw------- 1 letessier staff 2,1K 5 ao? 00:54 >> storage3__export-brick_home-brick2-data.log >> -rw------- 1 letessier staff 47M 5 ao? 00:55 >> storage4__export-brick_home-brick1-data.log >> -rw------- 1 letessier staff 2,1K 5 ao? 00:55 >> storage4__export-brick_home-brick2-data.log >> >> # wc -l storage* >> 55381 storage1__export-brick_home-brick1-data.log >> 17 storage1__export-brick_home-brick2-data.log >> 41636 storage2__export-brick_home-brick1-data.log >> 17 storage2__export-brick_home-brick2-data.log >> 270360 storage3__export-brick_home-brick1-data.log >> 17 storage3__export-brick_home-brick2-data.log >> 270358 storage4__export-brick_home-brick1-data.log >> 17 storage4__export-brick_home-brick2-data.log >> 637803 total >> >> If the let brick-log-level to INFO, the brick log files in each >> server will consume all my /var partition capacity within only a few >> hours/days? >> >> Thanks in advance, >> Geoffrey >> ------------------------------------------------------ >> Geoffrey Letessier >> Responsable informatique & ing?nieur syst?me >> UPR 9080 - CNRS - Laboratoire de Biochimie Th?orique >> Institut de Biologie Physico-Chimique >> 13, rue Pierre et Marie Curie - 75005 Paris >> Tel: 01 58 41 50 93 - eMail: geoffrey.letessier at ibpc.fr >> <mailto:geoffrey.letessier at ibpc.fr> >> >> Le 5 ao?t 2015 ? 01:12, Geoffrey Letessier >> <geoffrey.letessier at cnrs.fr <mailto:geoffrey.letessier at cnrs.fr>> a >> ?crit : >> >>> Hello, >>> >>> Since the problem motioned previously (all errors noticed in brick >>> log files), i notice a very very bad performance: i can note my >>> write performance divided by 4 than previously -knowing it was not >>> so good before. >>> Now, a write of a 33GB file, my write throughput is around 150MBs >>> (with Infiniband), before it was around 550-600MBs; and this, both >>> with RDMA and TCP protocol. >>> >>> During this test, more than 40 000 error lines (as the following) >>> were added to the brick log files. >>> [2015-08-04 22:34:27.337622] E [dict.c:1418:dict_copy_with_ref] >>> (-->/usr/lib64/glusterfs/3.7.3/xlator/protocol/server.so(server_resolve_inode+0x60) >>> [0x7f021c6f7410] >>> -->/usr/lib64/glusterfs/3.7.3/xlator/protocol/server.so(resolve_gfid+0x88) >>> [0x7f021c6f7188] >>> -->/usr/lib64/libglusterfs.so.0(dict_copy_with_ref+0xa4) >>> [0x7f0229cba674] ) 0-dict: invalid argument: dict [Argument invalide] >>> >>> >>> All brick log files are in attachments. >>> >>> Thanks in advance for all your help and fix, >>> Best, >>> Geoffrey >>> >>> PS: question: is it possible to easily downgrade GlusterFS to a >>> previous version from 3.7 (for example: v3.5)? >>> >>> ------------------------------------------------------ >>> Geoffrey Letessier >>> Responsable informatique & ing?nieur syst?me >>> UPR 9080 - CNRS - Laboratoire de Biochimie Th?orique >>> Institut de Biologie Physico-Chimique >>> 13, rue Pierre et Marie Curie - 75005 Paris >>> Tel: 01 58 41 50 93 - eMail: geoffrey.letessier at ibpc.fr >>> <mailto:geoffrey.letessier at ibpc.fr> >>> <bricks-logs.tgz> >> >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://www.gluster.org/pipermail/gluster-users/attachments/20150807/5454c964/attachment.html>