I am also seeing if reducing filestore queue ops limit from 500 to 250. On my graphs I can see the file store ops queue goes from 1 or 2 to 500 for the period of the load. I am looking to see if throttling down helps spread out the load. The normal ops load is not enough to worry the current limit.
Sent from my SAMSUNG Galaxy S7 on the Telstra Mobile Network -------- Original message -------- From: Nick Fisk <n...@fisk.me.uk> Date: 23/09/2016 7:26 PM (GMT+10:00) To: Adrian Saul <adrian.s...@tpgtelecom.com.au>, ceph-users@lists.ceph.com Subject: RE: Snap delete performance impact Looking back through my graphs when this happened to me I can see that the queue on the disks was up as high as 30 during the period when the snapshot was removed, this would explain the high latencies, the disk is literally having fits trying to jump all over the place. I need to test with the higher osd_snap_trim_sleep to see if that helps. What I'm interested in finding out is why so much disk activity is required for deleting an object. It feels to me that the process is async, in that Ceph will quite happily flood the Filestore with delete requests without any feedback to the higher layers. > -----Original Message----- > From: Adrian Saul [mailto:adrian.s...@tpgtelecom.com.au] > Sent: 23 September 2016 10:04 > To: n...@fisk.me.uk; ceph-users@lists.ceph.com > Subject: RE: Snap delete performance impact > > > I did some observation today - with the reduced filestore_op_threads it seems > to ride out the storm better, not ideal but better. > > The main issue is that for the 10 minutes from the moment the rbd snap rm > command is issued, the SATA systems in my configuration > load up massively on disk IO and I think this is what is rolling on to all > other issues (OSDs unresponsive, queue backlogs). The disks all > go 100% busy - the average SATA write latency goes from 14ms to 250ms. I was > observing disks doing 400, 700 and higher service > times. After those few minutes it tapers down and goes back to normal. > > There are all ST6000VN0001 disks - anyone aware of anything that might > explain this sort of behaviour? It seems odd that even if the > disks were hit with high write traffic (average of 50 write IOPS going up to > 270-300 during this activity) that the service times would > blow out that much. > > Cheers, > Adrian > > > > > > > > -----Original Message----- > > From: ceph-users [mailto:ceph-users-boun...@lists.ceph.com] On Behalf > > Of Adrian Saul > > Sent: Thursday, 22 September 2016 7:15 PM > > To: n...@fisk.me.uk; ceph-users@lists.ceph.com > > Subject: Re: [ceph-users] Snap delete performance impact > > > > > > I tried 2 this afternoon and saw the same results. Essentially the > > disks appear to go to 100% busy doing very small but high numbers of IO and > > incur massive > > service times (300-400ms). During that period I get blocked request errors > > continually. > > > > I suspect part of that might be the SATA servers had > > filestore_op_threads set too high and hammering the disks with too > > much concurrent work. As they have inherited a setting targeted for > > SSDs, so I have wound that back to defaults on those machines see if it > > makes a difference. > > > > But I suspect going by the disk activity there is a lot of very small > > FS metadata updates going on and that is what is killing it. > > > > Cheers, > > Adrian > > > > > > > -----Original Message----- > > > From: Nick Fisk [mailto:n...@fisk.me.uk] > > > Sent: Thursday, 22 September 2016 7:06 PM > > > To: Adrian Saul; ceph-users@lists.ceph.com > > > Subject: RE: Snap delete performance impact > > > > > > Hi Adrian, > > > > > > I have also hit this recently and have since increased the > > > osd_snap_trim_sleep to try and stop this from happening again. > > > However, I haven't had an opportunity to actually try and break it > > > again yet, but your mail seems to suggest it might not be the silver > > > bullet I > > was looking for. > > > > > > I'm wondering if the problem is not with the removal of the > > > snapshot, but actually down to the amount of object deletes that > > > happen, as I see similar results when doing fstrim's or deleting > > > RBD's. Either way I agree that a settable throttle to allow it to > > > process more slowly would be a > > good addition. > > > Have you tried that value set to higher than 1, maybe 10? > > > > > > Nick > > > > > > > -----Original Message----- > > > > From: ceph-users [mailto:ceph-users-boun...@lists.ceph.com] On > > > > Behalf Of Adrian Saul > > > > Sent: 22 September 2016 05:19 > > > > To: 'ceph-users@lists.ceph.com' <ceph-users@lists.ceph.com> > > > > Subject: Re: [ceph-users] Snap delete performance impact > > > > > > > > > > > > Any guidance on this? I have osd_snap_trim_sleep set to 1 and it > > > > seems to have tempered some of the issues but its still bad > > > enough > > > > that NFS storage off RBD volumes become unavailable for over 3 > > minutes. > > > > > > > > It seems that the activity which the snapshot deletes are actioned > > > > triggers massive disk load for around 30 minutes. The logs > > > show > > > > OSDs marking each other out, OSDs complaining they are wrongly > > > > marked out and blocked requests errors for around 10 minutes at > > > > the start of this > > > activity. > > > > > > > > Is there any way to throttle snapshot deletes to make them much > > > > more of a background activity? It really should not make the > > > entire > > > > platform unusable for 10 minutes. > > > > > > > > > > > > > > > > > -----Original Message----- > > > > > From: ceph-users [mailto:ceph-users-boun...@lists.ceph.com] On > > > > > Behalf Of Adrian Saul > > > > > Sent: Wednesday, 6 July 2016 3:41 PM > > > > > To: 'ceph-users@lists.ceph.com' > > > > > Subject: [ceph-users] Snap delete performance impact > > > > > > > > > > > > > > > I recently started a process of using rbd snapshots to setup a > > > > > backup regime for a few file systems contained in RBD images. > > > > > While this generally works well at the time of the snapshots > > > > > there is a massive increase in latency (10ms to multiple seconds > > > > > of rbd device > > > > > latency) across the entire cluster. This has flow on effects > > > > > for some cluster timeouts as well as general performance hits to > > applications. > > > > > > > > > > In research I have found some references to osd_snap_trim_sleep > > > > > being > > > the > > > > > way to throttle this activity but no real guidance on values for it. > > > > > I also > > > see > > > > > some other osd_snap_trim tunables (priority and cost). > > > > > > > > > > Is there any recommendations around setting these for a Jewel cluster? > > > > > > > > > > cheers, > > > > > Adrian > > > > > _______________________________________________ > > > > > ceph-users mailing list > > > > > ceph-users@lists.ceph.com > > > > > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > > > > Confidentiality: This email and any attachments are confidential > > > > and may be subject to copyright, legal or some other professional > > > > privilege. They are intended solely for the attention and use of > > > > the named addressee(s). They may only be copied, distributed or > > > > disclosed with the consent of the copyright owner. If you have > > > > received this email by > > > mistake or by breach of the confidentiality clause, please notify > > > the sender immediately by return email and delete or destroy all > > > copies of the > > email. > > > Any confidentiality, privilege or copyright is not waived or lost > > > because this email has been sent to you by mistake. > > > > _______________________________________________ > > > > ceph-users mailing list > > > > ceph-users@lists.ceph.com > > > > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > > > > Confidentiality: This email and any attachments are confidential and > > may be subject to copyright, legal or some other professional > > privilege. They are intended solely for the attention and use of the > > named addressee(s). They may only be copied, distributed or disclosed > > with the consent of the copyright owner. If you have received this > > email by mistake or by breach of the confidentiality clause, please > > notify the sender immediately by return email and delete or destroy > > all copies of the email. Any confidentiality, privilege or copyright > > is not waived or lost because this email has been sent to you by mistake. > > _______________________________________________ > > ceph-users mailing list > > ceph-users@lists.ceph.com > > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > Confidentiality: This email and any attachments are confidential and may be > subject to copyright, legal or some other professional > privilege. They are intended solely for the attention and use of the named > addressee(s). They may only be copied, distributed or > disclosed with the consent of the copyright owner. If you have received this > email by mistake or by breach of the confidentiality > clause, please notify the sender immediately by return email and delete or > destroy all copies of the email. Any confidentiality, > privilege or copyright is not waived or lost because this email has been sent > to you by mistake. Confidentiality: This email and any attachments are confidential and may be subject to copyright, legal or some other professional privilege. They are intended solely for the attention and use of the named addressee(s). They may only be copied, distributed or disclosed with the consent of the copyright owner. If you have received this email by mistake or by breach of the confidentiality clause, please notify the sender immediately by return email and delete or destroy all copies of the email. Any confidentiality, privilege or copyright is not waived or lost because this email has been sent to you by mistake.
_______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com