Can this be related to numa issues? I have also dual processor nodes, and was wondering if there is some guide on how to optimize for numa. -----Original Message----- From: Tyler Bishop [mailto:tyler.bishop@xxxxxxxxxxxxxxxxx] Sent: vrijdag 24 augustus 2018 3:11 To: Andras Pataki Cc: ceph-users@xxxxxxxxxxxxxx Subject: Re: Stability Issue with 52 OSD hosts Thanks for the info. I was investigating bluestore as well. My host dont go unresponsive but I do see parallel io slow down. On Thu, Aug 23, 2018, 8:02 PM Andras Pataki <apataki@xxxxxxxxxxxxxxxxxxxxx> wrote: We are also running some fairly dense nodes with CentOS 7.4 and ran into similar problems. The nodes ran filestore OSDs (Jewel, then Luminous). Sometimes a node would be so unresponsive that one couldn't even ssh to it (even though the root disk was a physically separate drive on a separate controller from the OSD drives). Often these would coincide with kernel stack traces about hung tasks. Initially we did blame high load, etc. from all the OSDs. But then we benchmarked the nodes independently of ceph (with iozone and such) and noticed problems there too. When we started a few dozen iozone processes on separate JBOD drives with xfs, some didn't even start and write a single byte for minutes. The conclusion we came to was that there is some interference among a lot of mounted xfs file systems in the Red Hat 3.10 kernels. Some kind of central lock that prevents dozens of xfs file systems from running in parallel. When we do I/O directly to raw devices in parallel, we saw no problems (no high loads, etc.). So we built a newer kernel, and the situation got better. 4.4 is already much better, nowadays we are testing moving to 4.14. Also, migrating to bluestore significantly reduced the load on these nodes too. At busy times, the filestore host loads were 20-30, even higher (on a 28 core node), while the bluestore nodes hummed along at a lot of perhaps 6 or 8. This also confirms that somehow lots of xfs mounts don't work in parallel. Andras On 08/23/2018 03:24 PM, Tyler Bishop wrote: > Yes I've reviewed all the logs from monitor and host. I am not > getting useful errors (or any) in dmesg or general messages. > > I have 2 ceph clusters, the other cluster is 300 SSD and i never have > issues like this. That's why Im looking for help. > > On Thu, Aug 23, 2018 at 3:22 PM Alex Gorbachev <ag@xxxxxxxxxxxxxxxxxxx> wrote: >> On Wed, Aug 22, 2018 at 11:39 PM Tyler Bishop >> <tyler.bishop@xxxxxxxxxxxxxxxxx> wrote: >>> During high load testing I'm only seeing user and sys cpu load around 60%... my load doesn't seem to be anything crazy on the host and iowait stays between 6 and 10%. I have very good `ceph osd perf` numbers too. >>> >>> I am using 10.2.11 Jewel. >>> >>> >>> On Wed, Aug 22, 2018 at 11:30 PM Christian Balzer <chibi@xxxxxxx> wrote: >>>> Hello, >>>> >>>> On Wed, 22 Aug 2018 23:00:24 -0400 Tyler Bishop wrote: >>>> >>>>> Hi, I've been fighting to get good stability on my cluster for about >>>>> 3 weeks now. I am running into intermittent issues with OSD flapping >>>>> marking other OSD down then going back to a stable state for hours and >>>>> days. >>>>> >>>>> The cluster is 4x Cisco UCS S3260 with dual E5-2660, 256GB ram, 40G >>>>> Network to 40G Brocade VDX Switches. The OSD are 6TB HGST SAS drives >>>>> with 400GB HGST SAS 12G SSDs. My configuration is 4 journals per >>>>> host with 12 disk per journal for a total of 56 disk per system and 52 >>>>> OSD. >>>>> >>>> Any denser and you'd have a storage black hole. >>>> >>>> You already pointed your finger in the (or at least one) right direction >>>> and everybody will agree that this setup is woefully underpowered in the >>>> CPU department. >>>> >>>>> I am using CentOS 7 with kernel 3.10 and the redhat tuned-adm profile >>>>> for throughput-performance enabled. >>>>> >>>> Ceph version would be interesting as well... >>>> >>>>> I have these sysctls set: >>>>> >>>>> kernel.pid_max = 4194303 >>>>> fs.file-max = 6553600 >>>>> vm.swappiness = 0 >>>>> vm.vfs_cache_pressure = 50 >>>>> vm.min_free_kbytes = 3145728 >>>>> >>>>> I feel like my issue is directly related to the high number of OSD per >>>>> host but I'm not sure what issue I'm really running into. I believe >>>>> that I have ruled out network issues, i am able to get 38Gbit >>>>> consistently via iperf testing and mtu for jump pings successfully >>>>> with no fragment set and 8972 packet size. >>>>> >>>> The fact that it all works for days at a time suggests this as well, but >>>> you need to verify these things when they're happening. >>>> >>>>> From FIO testing I seem to be able to get 150-200k iops write from my >>>>> rbd clients on 1gbit networking... This is about what I expected due >>>>> to the write penalty and my underpowered CPU for the number of OSD. >>>>> >>>>> I get these messages which I believe are normal? >>>>> 2018-08-22 10:33:12.754722 7f7d009f5700 0 -- 10.20.136.8:6894/718902 >>>>>>> 10.20.136.10:6876/490574 pipe(0x55aed77fd400 sd=192 :40502 s=2 >>>>> pgs=1084 cs=53 l=0 c=0x55aed805bc80).fault with nothing to send, going >>>>> to standby >>>>> >>>> Ignore. >>>> >>>>> Then randomly I'll get a storm of this every few days for 20 minutes or so: >>>>> 2018-08-22 15:48:32.631186 7f44b7514700 -1 osd.127 37333 >>>>> heartbeat_check: no reply from 10.20.142.11:6861 osd.198 since back >>>>> 2018-08-22 15:48:08.052762 front 2018-08-22 15:48:31.282890 (cutoff >>>>> 2018-08-22 15:48:12.630773) >>>>> >>>> Randomly is unlikely. >>>> Again, catch it in the act, atop in huge terminal windows (showing all >>>> CPUs and disks) for all nodes should be very telling, collecting and >>>> graphing this data might work, too. >>>> >>>> My suspects would be deep scrubs and/or high IOPS spikes when this is >>>> happening, starving out OSD processes (CPU wise, RAM should be fine one >>>> supposes). >>>> >>>> Christian >>>> >>>>> Please help!!! >> Have you looked at the OSD logs on the OSD nodes by chance? I found >> that correlating the messages in those logs with your master ceph log >> and also correlating with any messages in syslog or kern.log can >> elucidate the cause of the problem pretty well. >> -- >> Alex Gorbachev >> Storcium >> >> >>>>> _______________________________________________ >>>>> ceph-users mailing list >>>>> ceph-users@xxxxxxxxxxxxxx >>>>> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >>>>> >>>> >>>> -- >>>> Christian Balzer Network/Systems Engineer >>>> chibi@xxxxxxx Rakuten Communications >>> _______________________________________________ >>> ceph-users mailing list >>> ceph-users@xxxxxxxxxxxxxx >>> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > _______________________________________________ > ceph-users mailing list > ceph-users@xxxxxxxxxxxxxx > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |