[ovirt-users] Re: Ovirt cluster unstable; gluster to blame (again)

Thank you for your help. After more troubleshooting and host reboots, I accidentally discovered that the backing disk on ovirt2 (host) had suffered a failure. On reboot, the raid card refused to see it at all. It said it had cache waiting to be written to disk, and in the end, as it couldn't (wouldn't) see that disk, I had no choice but to discard that cache and boot up without the physical disk. Since doing so (and running a gluster volume remove for the affected host), things are running like normal, although it appears it corrupted two disks (I've now lost 5 VMs to gluster-induced disk failures during poorly handled failures). I don't understand why one bad disk wasn't simply failed, or if one underlying process was having such a problem, the other hosts didn't take it offline and continue (much like RAID would have done). Instead, everything was broke (including gluster volumes on unaffected disks that are fully functional across all hosts) as well as very poor performance of affected machine AND no diagnostic reports that would allude to a failing hard drive. Is this expected behavior? --Jim On Sun, Jul 8, 2018 at 3:54 AM, Yaniv Kaul <ykaul(a)redhat.com&gt; wrote: On Sat, Jul 7, 2018 at 8:45 AM, Jim Kusznir <jim(a)palousetech.com&gt; wrote: So, I'm still at a loss...It sounds like its either insufficient ram/swap, or insufficient network. It seems to be neither now. At this point, it appears that gluster is just "broke" and killing my systems for no descernable reason. Here's detals, all from the same system (currently running 3 VMs): [root@ovirt3 ~]# w 22:26:53 up 36 days, 4:34, 1 user, load average: 42.78, 55.98, 53.31 USER TTY FROM LOGIN@ IDLE JCPU PCPU WHAT root pts/0 192.168.8.90 22:26 2.00s 0.12s 0.11s w bwm-ng reports the highest data usage was about 6MB/s during this test (and that was combined; I have two different gig networks. One gluster network (primary VM storage) runs on one, the other network handles everything else). [root@ovirt3 ~]# free -m total used free shared buff/cache available Mem: 31996 13236 232 18 18526 18195 Swap: 16383 1475 14908 top - 22:32:56 up 36 days, 4:41, 1 user, load average: 17.99, 39.69, 47.66 That is indeed a high load average. How many CPUs do you have, btw? Tasks: 407 total, 1 running, 405 sleeping, 1 stopped, 0 zombie %Cpu(s): 8.6 us, 2.1 sy, 0.0 ni, 87.6 id, 1.6 wa, 0.0 hi, 0.1 si, 0.0 st KiB Mem : 32764284 total, 228296 free, 13541952 used, 18994036 buff/cache KiB Swap: 16777212 total, 15246200 free, 1531012 used. 18643960 avail Mem Can you check what's swapping here? (a tweak to top output will show that) PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 30036 qemu 20 0 6872324 5.2g 13532 S 144.6 16.5 216:14.55 /usr/libexec/qemu-kvm -name guest=BillingWin,debug-threads=on -S -object secret,id=masterKey0,format=raw,file=/v+ 28501 qemu 20 0 5034968 3.6g 12880 S 16.2 11.7 73:44.99 /usr/libexec/qemu-kvm -name guest=FusionPBX,debug-threads=on -S -object secret,id=masterKey0,format=raw,file=/va+ 2694 root 20 0 2169224 12164 3108 S 5.0 0.0 3290:42 /usr/sbin/glusterfsd -s ovirt3.nwfiber.com --volfile-id data.ovirt3.nwfiber.com.gluster-brick2-data -p /var/run/+ This one's certainly taking quite a bit of your CPU usage overall. 14293 root 15 -5 944700 13356 4436 S 4.0 0.0 16:32.15 /usr/sbin/glusterfs --volfile-server=192.168.8.11 --volfile-server=192.168.8.12 --volfile-server=192.168.8.13 --+ I'm not sure what the sorting order is, but doesn't look like Gluster is taking a lot of memory? 25100 vdsm 0 -20 6747440 107868 12836 S 2.3 0.3 21:35.20 /usr/bin/python2 /usr/share/vdsm/vdsmd 28971 qemu 20 0 2842592 1.5g 13548 S 1.7 4.7 241:46.49 /usr/libexec/qemu-kvm -name guest=unifi.palousetech.com,debug-threads=on -S -object secret,id=masterKey0,format=+ 12095 root 20 0 162276 2836 1868 R 1.3 0.0 0:00.25 top 2708 root 20 0 1906040 12404 3080 S 1.0 0.0 1083:33 /usr/sbin/glusterfsd -s ovirt3.nwfiber.com --volfile-id engine.ovirt3.nwfiber.com.gluster-brick1-engine -p /var/+ 28623 qemu 20 0 4749536 1.7g 12896 S 0.7 5.5 4:30.64 /usr/libexec/qemu-kvm -name guest=billing.nwfiber.com,debug-threads=on -S -object secret,id=masterKey0,format=ra+ The VMs I see here and above together account for most? (5.2+3.6+1.5+1.7 = 12GB) - still plenty of memory left. 10 root 20 0 0 0 0 S 0.3 0.0 215:54.72 [rcu_sched] 1030 sanlock rt 0 773804 27908 2744 S 0.3 0.1 35:55.61 /usr/sbin/sanlock daemon 1890 zabbix 20 0 83904 1696 1612 S 0.3 0.0 24:30.63 /usr/sbin/zabbix_agentd: collector [idle 1 sec] 2722 root 20 0 1298004 6148 2580 S 0.3 0.0 38:10.82 /usr/sbin/glusterfsd -s ovirt3.nwfiber.com --volfile-id iso.ovirt3.nwfiber.com.gluster-brick4-iso -p /var/run/gl+ 6340 root 20 0 0 0 0 S 0.3 0.0 0:04.30 [kworker/7:0] 10652 root 20 0 0 0 0 S 0.3 0.0 0:00.23 [kworker/u64:2] 14724 root 20 0 1076344 17400 3200 S 0.3 0.1 10:04.13 /usr/sbin/glusterfs -s localhost --volfile-id gluster/glustershd -p /var/run/gluster/glustershd/glustershd.pid -+ 22011 root 20 0 0 0 0 S 0.3 0.0 0:05.04 [kworker/10:1] Not sure why the system load dropped other than I was trying to take a picture of it :) In any case, it appears that at this time, I have plenty of swap, ram, and network capacity, and yet things are still running very sluggish; I'm still getting e-mails from servers complaining about loss of communication with something or another; I still get e-mails from the engine about bad engine status, then recovery, etc. 1g isn't good enough for Gluster. It doesn't help that you have SSD, because network is certainly your bottleneck even for regular performance, not to mention when you are healing. Jumbo frames would give you additional 5% or so - nothing to write home about. I've shut down 2/3 of my VMs, too....just trying to keep the critical ones operating. At this point, I don't believe the problem is the memory leak, but it seems to be triggered by the memory leak, as in all my problems started when I got low ram warnings from one of my 3 nodes and began recovery efforts from that. I do really like the idea / concept behind glusterfs, but I really have to figure out why its been so poor performing from day one, and its caused 95% of my outages (including several large ones lately). If I can get it stable, reliable, and well performing, then I'd love to keep it. If I can't, then perhaps NFS is the way to go? I don't like the single point of failure aspect of it, but my other NAS boxes I run for clients (central storage for windows boxes) have been very solid; If I could get that kind of reliability for my ovirt stack, it would be a substantial improvement. Currently, it seems about every other month I have a gluster-induced outage. Sometimes I wonder if its just hyperconverged is the issue, but my infrastructure doesn't justify three servers at the same location...I might be able to do two, but even that seems like its pushing it. We have many happy users running Gluster and hyperconverged. We need to understand where's the failure in your setup. Looks like I can upgrade to 10G for about $900. I can order a dual-Xeon supermicro 12-disk server, loaded with 2TB WD Enterprise disks and a pair of SSDs for the os, 32GB ram, 2.67Ghz CPUs for about $720 delivered. I've got to do something to improve my reliability; I can't keep going the way I have been.... Agreed. Thanks for continuing looking into this, we'll probably need some Gluster logs to understand what's going on. Y. --Jim On Fri, Jul 6, 2018 at 9:13 PM, Johan Bernhardsson <johan(a)kafit.se&gt; wrote: Load like that is mostly io based either the machine is swapping or network is to slow. Check I/o wait in top. And the problem where you get oom killer to kill off gluster. That means that you don't monitor ram usage on the servers? Either it's eating all your ram and swap gets really io intensive and then is killed off. Or you have the wrong swap settings in sysctl.conf (there are tons of broken guides that recommends swappines to 0 but that disables swap on newer kernels. The proper swappines for only swapping when nesseary is 1 or a sufficiently low number like 10 default is 60) Moving to nfs will not improve things. You will get more memory since gluster isn't running and that is good. But you will have a single node that can fail with all your storage and it would still be on 1 gigabit only and your three node cluster would easily saturate that link. On July 7, 2018 04:13:13 Jim Kusznir <jim(a)palousetech.com&gt; wrote: > So far it does not appear to be helping much. I'm still getting VM's > locking up and all kinds of notices from overt engine about non-responsive > hosts. I'm still seeing load averages in the 20-30 range. > > Jim > > On Fri, Jul 6, 2018, 3:13 PM Jim Kusznir <jim(a)palousetech.com&gt; wrote: > Thank you for the advice and help > > I do plan on going 10Gbps networking; haven't quite jumped off that cliff > yet, though. > > I did put my data-hdd (main VM storage volume) onto a dedicated 1Gbps > network, and I've watched throughput on that and never seen more than > 60GB/s achieved (as reported by bwm-ng). I have a separate 1Gbps network > for communication and ovirt migration, but I wanted to break that up > further (separate out VM traffice from migration/mgmt traffic). My three > SSD-backed gluster volumes run the main network too, as I haven't been able > to get them to move to the new network (which I was trying to use as all > gluster). I tried bonding, but that seamed to reduce performance rather > than improve it. > > --Jim > > On Fri, Jul 6, 2018 at 2:52 PM, Jamie Lawrence <jlawrence(a)squaretrade.com&gt; > wrote: > > Hi Jim, > > I don't have any targeted suggestions, because there isn't much to latch on > to. I can say Gluster replica three (no arbiters) on dedicated servers > serving a couple Ovirt VM clusters here have not had these sorts of issues. > > I suspect your long heal times (and the resultant long periods of high > load) are at least partly related to 1G networking. That is just a matter > of IO - heals of VMs involve moving a lot of bits. My cluster uses 10G > bonded NICs on the gluster and ovirt boxes for storage traffic and separate > bonded 1G for ovirtmgmt and communication with other machines/people, and > we're occasionally hitting the bandwidth ceiling on the storage network. > I'm starting to think about 40/100G, different ways of splitting up > intensive systems, and considering iSCSI for specific volumes, although I > really don't want to go there. > > I don't run FreeNAS[1], but I do run FreeBSD as storage servers for their > excellent ZFS implementation, mostly for backups. ZFS will make your `heal` > problem go away, but not your bandwidth problems, which become worse > (because of fewer NICS pushing traffic). 10G hardware is not exactly in the > impulse-buy territory, but if you can, I'd recommend doing some testing > using it. I think at least some of your problems are related. > > If that's not possible, my next stops would be optimizing everything I > could about sharding, healing and optimizing for serving the shard size to > squeeze as much performance out of 1G as I could, but that will only go so far. > > -j > > [1] FreeNAS is just a storage-tuned FreeBSD with a GUI. > > >> On Jul 6, 2018, at 1:19 PM, Jim Kusznir <jim(a)palousetech.com&gt; wrote: >> >> hi all: >> >> Once again my production ovirt cluster is collapsing in on itself. My >> servers are intermittently unavailable or degrading, customers are noticing >> and calling in. This seems to be yet another gluster failure that I >> haven't been able to pin down. >> >> I posted about this a while ago, but didn't get anywhere (no replies that I >> found). The problem started out as a glusterfsd process consuming large >> amounts of ram (up to the point where ram and swap were exhausted and the >> kernel OOM killer killed off the glusterfsd process). For reasons not >> clear to me at this time, that resulted in any VMs running on that host and >> that gluster volume to be paused with I/O error (the glusterfs process is >> usually unharmed; why it didn't continue I/O with other servers is >> confusing to me). >> >> I have 3 servers and a total of 4 gluster volumes (engine, iso, data, and >> data-hdd). The first 3 are replica 2+arb; the 4th (data-hdd) is replica 3. >> The first 3 are backed by an LVM partition (some thin provisioned) on an >> SSD; the 4th is on a seagate hybrid disk (hdd + some internal flash for >> acceleration). data-hdd is the only thing on the disk. Servers are Dell >> R610 with the PERC/6i raid card, with the disks individually passed through >> to the OS (no raid enabled). >> >> The above RAM usage issue came from the data-hdd volume. Yesterday, I >> cought one of the glusterfsd high ram usage before the OOM-Killer had to >> run. I was able to migrate the VMs off the machine and for good measure, >> reboot the entire machine (after taking this opportunity to run the >> software updates that ovirt said were pending). Upon booting back up, the >> necessary volume healing began. However, this time, the healing caused all >> three servers to go to very, very high load averages (I saw just under 200 >> on one server; typically they've been 40-70) with top reporting IO Wait at >> 7-20%. Network for this volume is a dedicated gig network. According to >> bwm-ng, initially the network bandwidth would hit 50MB/s (yes, bytes), but >> tailed off to mostly in the kB/s for a while. All machines' load averages >> were still 40+ and gluster volume heal data-hdd info reported 5 items >> needing healing. Server's were intermittently experiencing IO issues, even >> on the 3 gluster volumes that appeared largely unaffected. Even the OS >> activities on the hosts itself (logging in, running commands) would often >> be very delayed. The ovirt engine was seemingly randomly throwing engine >> down / engine up / engine failed notifications. Responsiveness on ANY VM >> was horrific most of the time, with random VMs being inaccessible. >> >> I let the gluster heal run overnight. By morning, there were still 5 items >> needing healing, all three servers were still experiencing high load, and >> servers were still largely unstable. >> >> I've noticed that all of my ovirt outages (and I've had a lot, way more >> than is acceptable for a production cluster) have come from gluster. I >> still have 3 VMs who's hard disk images have become corrupted by my last >> gluster crash that I haven't had time to repair / rebuild yet (I believe >> this crash was caused by the OOM issue previously mentioned, but I didn't >> know it at the time). >> >> Is gluster really ready for production yet? It seems so unstable to me.... >> I'm looking at replacing gluster with a dedicated NFS server likely >> FreeNAS. Any suggestions? What is the "right" way to do production >> storage on this (3 node cluster)? Can I get this gluster volume stable >> enough to get my VMs to run reliably again until I can deploy another >> storage solution? >> >> --Jim >> _______________________________________________ >> Users mailing list -- users(a)ovirt.org >> To unsubscribe send an email to users-leave(a)ovirt.org >> Privacy Statement: https://www.ovirt.org/site/privacy-policy/ >> oVirt Code of Conduct: >> https://www.ovirt.org/community/about/community-guidelines/ >> List Archives: >> https://lists.ovirt.org/archives/list/users@ovirt.org/message/YQX3LQFQQPW... > > > _______________________________________________ > Users mailing list -- users(a)ovirt.org > To unsubscribe send an email to users-leave(a)ovirt.org > Privacy Statement: https://www.ovirt.org/site/privacy-policy/ > oVirt Code of Conduct: > https://www.ovirt.org/community/about/community-guidelines/ > List Archives: > https://lists.ovirt.org/archives/list/users@ovirt.org/message/O2HIECLFMYG... _______________________________________________ Users mailing list -- users(a)ovirt.org To unsubscribe send an email to users-leave(a)ovirt.org Privacy Statement: https://www.ovirt.org/site/privacy-policy/ oVirt Code of Conduct: https://www.ovirt.org/community/about/community-guidelines/ List Archives: https://lists.ovirt.org/archives/list/users@ovirt.org/message/73F7P66ARAQ...