In our case it was the co-scheduling with lots QEMUs that made it run bad. If you have a dedicated CEPH OSD server it would be beneficial only if the scheduler was moving the OSDs between different NUMA zones. (Which our ancient 2.6.32 EL6 kernel AFAIK does but newer kernels do not). I’m not seeing our OSDs having a problem with IO speed, Dumpling is just so slow and CPU-bound it would probably run at the same speed even with spindles. It is hard to design a NUMA system where all the resources are local - you’d have to have dedicated separate bonds for all zones etc. - easier to just use smaller 1 socket machines in greater numbers. When you add QEMU to the mix it’s practically impossible to have everything local to all the zones in one box. (CEPH needs HBAs and NICs, QEMU needs NICs but shouldn’t ideally share cores with CEPH… depends very much on the workload and scale - with greater scale it probably doesn’t make sense to have a hyperconverged solution, unless it’s easier to just throw more hardware at the problem and only scaling horizontaly). migratepages is a one-shot operation - memory placement after that will depends on the kernel you are running and scheduler and other settings. Having zone_reclaim_mode=1 should prevent memory from “leaking” to the other node, but could prevent effective filesystem caching depending on how much free memory you have. Surprisingly, I had ~150GB free (about 50% in each of two NUMA zones) that wasn’t used for cache without turning it off, not yet sure why. Scripts will be coming tomorrow, I’d love to see if it makes a change for someone else, maybe I’m just un-breaking something in my setup. Jan > On 24 Jun 2015, at 20:05, Somnath Roy <Somnath.Roy@xxxxxxxxxxx> wrote: > > Jan, > This is interesting as I tried to pin OSDs (though I didn't pin all the threads) as part of our tuning and didn't see much difference. I thought this could be primarily because of the following. > > The NICs and HBAs could be always remote to some OSDs , unless you dedicate NICs to the OSDs running on the same NUMA node. > > I never tried 'migratepages' though. But, I guess 'migratepages' we need to do one time after pinning, right ? > I would love to see your scripts and try it out in my environment. > > Thanks & Regards > Somnath > > > > -----Original Message----- > From: ceph-users [mailto:ceph-users-bounces@xxxxxxxxxxxxxx] On Behalf Of Jan Schermer > Sent: Wednesday, June 24, 2015 10:54 AM > To: Ben Hines > Cc: ceph-users@xxxxxxxxxxxxxx > Subject: Re: Switching from tcmalloc > > We did, but I don’t have the numbers. I have lots of graphs, though. We were mainly trying to solve the CPU usage, since our nodes are converged QEMU+CEPH OSDs, so this made a difference. We were also seeing the performance capped on CPUs when deleting snapshots of backfilling, all this should be solved with this. > > We graph latency, outstanding operations, you name it - I can share a few graphs with you tomorrow if I get the permission from my boss :-) Makes for a nice comparison with real workload to have one node tcmalloc-free and the others running vanilla ceph-osd. > > I guess I can share the final script once that’s finished - right now it uses taskset and then migratepages to the correct NUMA node and is not that nice, the cgroup one will be completely different. > > You can try migratepages for yourself to test if it makes a difference - pin an OSD to a specific node (don’t forget to pin all threads) and then run “migratepages $pid old_node new_node”. > You can confirm the memory moving with “numastat -p $pid”. If it doesn’t seem to move then it is probably pagecache allocated on the wrong node, not sure if that can be migrated but you can use /proc/sys/vm/zone_reclaim_mode (=1) which should drop it. I advise setting it to 0 in the end, though as cache is always faster than disks. > YMMV depending on bottlenecks your system has. > > Jan > > >> On 24 Jun 2015, at 19:36, Ben Hines <bhines@xxxxxxxxx> wrote: >> >> Did you do before/after Ceph performance benchmarks? I dont care if my >> systems are using 80% cpu, if Ceph performance is better than when >> it's using 20% cpu. >> >> Can you share any scripts you have to automate these things? (NUMA >> pinning, migratepages) >> >> thanks, >> >> -Ben >> >> On Wed, Jun 24, 2015 at 10:25 AM, Jan Schermer <jan@xxxxxxxxxxx> wrote: >>> There were essentialy three things we had to do for such a drastic >>> drop >>> >>> 1) recompile CEPH —without-tcmalloc >>> 2) pin the OSDs to a set of a specific NUMA zone - we had this for a >>> long time and it really helped >>> 3) migrate the OSD memory to the correct CPU with migratepages >>> - we will use cgroups in the future for this, should make life easier >>> and is the only correct solution >>> >>> It is similiar to the effect of just restarting the OSD, but much >>> better - since we immediately see hundreds of connections on a >>> freshly restarted OSD (and in the benchmark the tcmalloc issue >>> manifested with just two clients in >>> parallel) I’d say we never saw the raw performance with tcmalloc >>> (undegraded), but it was never this good - consistently low >>> latencies, much smaller spikes when something happens and much lower >>> CPU usage (about 50% savings but we’re also backfilling a lot on the >>> background). Workloads are faster as well - like reweighting OSDs on >>> that same node was much (hundreds of percent) faster. >>> >>> So far the effect has been drastic. I wonder why tcmalloc was even >>> used when people are having problems with it? The glibc malloc seems >>> to work just fine for us. >>> >>> The only concerning thing is the virtual memory usage - we are over >>> 400GB VSS with a few OSDs. That doesn’t hurt anything, though. >>> >>> Jan >>> >>> >>> On 24 Jun 2015, at 18:46, Robert LeBlanc <robert@xxxxxxxxxxxxx> wrote: >>> >>> -----BEGIN PGP SIGNED MESSAGE----- >>> Hash: SHA256 >>> >>> Did you see what the effect of just restarting the OSDs before using >>> tcmalloc? I've noticed that there is usually a good drop for us just >>> by restarting them. I don't think it is usually this drastic. >>> >>> - ---------------- >>> Robert LeBlanc >>> GPG Fingerprint 79A2 9CA4 6CC4 45DD A904 C70E E654 3BB2 FA62 B9F1 >>> >>> On Wed, Jun 24, 2015 at 2:08 AM, Jan Schermer wrote: >>> Can you guess when we did that? >>> Still on dumpling, btw... >>> >>> http://www.zviratko.net/link/notcmalloc.png >>> >>> Jan >>> >>> _______________________________________________ >>> ceph-users mailing list >>> ceph-users@xxxxxxxxxxxxxx >>> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >>> >>> >>> -----BEGIN PGP SIGNATURE----- >>> Version: Mailvelope v0.13.1 >>> Comment: https://www.mailvelope.com >>> >>> wsFcBAEBCAAQBQJVit75CRDmVDuy+mK58QAAmjcP/jU+wyohdwKDP+FHDAgJ >>> DcqdB5aPG2AM79iLcYUub5bQjdNJpcWN/hyZcNdF3aSzEV3aY6jIqu9OpOIB >>> c2fIzfGOoczzW/FEf7qKRVGpxaQL21Sw1LpwMEscNe0ETz9HMHoaAnBO9IFn >>> nUEOCdEpRBO5W1rWwNAx9EVnOUPklb7vVEpY23sgtHhQSprb9oeO8D99AMRz >>> /RhdHKlRDgHBjun/stCiR6lFuvBUx0GBmyaMuO5rfsLGRIkySLv++3CLQI6X >>> NCt/MjYwTTNNfO/y/MjkiV/j+Cm1G1lcjlgbDjilf7bgf8/7W2vJa1sMtaA4 >>> xJL+PpZxiKcGSdC96B+EBYxLhLcwsNpbfq7uxQOkIspa66mkIMAVzJgt4DFL >>> Ca+UY3ODA26VtWF5U/hkdupgld+YSxXTyJakeShrBSFAX0a4cygV9Ll7SIhO >>> IDS+0Mbur0IGzIWRgtCQhRXsc7wn3IoIovqe8Nfk4xupeoK2P5UHO1rW9pWy >>> Jwj5PXieDqxgx8RKlulN1bCbSgTaEdveTiqqVxlnM9L0MhgesuB8vkpHbsqn >>> mYJHNzU7ghU89xLnRuia9rBlpjw4OzagfowAJTH3UnaO67kxES+IWO8onQbN >>> RhY0QR5cB5rVSjYkzzlsuLM17fQPcT8++yMarKdsrr6WIGppXUFFdATAqIaY >>> DHD1 >>> =goL4 >>> -----END PGP SIGNATURE----- >>> >>> >>> >>> _______________________________________________ >>> 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 > > ________________________________ > > PLEASE NOTE: The information contained in this electronic mail message is intended only for the use of the designated recipient(s) named above. If the reader of this message is not the intended recipient, you are hereby notified that you have received this message in error and that any review, dissemination, distribution, or copying of this message is strictly prohibited. If you have received this communication in error, please notify the sender by telephone or e-mail (as shown above) immediately and destroy any and all copies of this message in your possession (whether hard copies or electronically stored copies). > _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |