-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA256 I also saw a similar performance increase by using alternative memory allocators. What I found was that Ceph OSDs performed well with either tcmalloc or jemalloc (except when RocksDB was built with jemalloc instead of tcmalloc, I'm still working to dig into why that might be the case). However, I found that tcmalloc with QEMU/KVM was very detrimental to small I/O, but provided huge gains in I/O >=1MB. Jemalloc was much better for QEMU/KVM in the tests that we ran. [1] I'm currently looking into I/O bottlenecks around the 16KB range and I'm seeing a lot of time in thread creation and destruction, the memory allocators are quite a bit down the list (both fio with ioengine rbd and on the OSDs). I wonder what the difference can be. I've tried using the async messenger but there wasn't a huge difference. [2] Further down the rabbit hole.... [1] https://www.mail-archive.com/ceph-users@xxxxxxxxxxxxxx/msg20197.html [2] https://www.mail-archive.com/ceph-devel@xxxxxxxxxxxxxxx/msg23982.html -----BEGIN PGP SIGNATURE----- Version: Mailvelope v0.13.1 Comment: https://www.mailvelope.com wsFcBAEBCAAQBQJVdw2ZCRDmVDuy+mK58QAA4MwP/1vt65cvTyyVGGSGRrE8 unuWjafMHzl486XH+EaVrDVTXFVFOoncJ6kugSpD7yavtCpZNdhsIaTRZguU YpfAppNAJU5biSwNv9QPI7kPP2q2+I7Z8ZkvhcVnkjIythoeNnSjV7zJrw87 afq46GhPHqEXdjp3rOB4RRPniOMnub5oU6QRnKn3HPW8Dx9ZqTeCofRDnCY2 S695Dt1gzt0ERUOgrUUkt0FQJdkkV6EURcUschngjtEd5727VTLp02HivVl3 vDYWxQHPK8oS6Xe8GOW0JjulwiqlYotSlrqSU5FMU5gozbk9zMFPIUW1e+51 9ART8Ta2ItMhPWtAhRwwvxgy51exCy9kBc+m+ptKW5XRUXOImGcOQxszPGOO qIIOG1vVG/GBmo/0i6tliqBFYdXmw1qFV7tFiIbisZRH7Q/1NahjYTHqHhu3 Dv61T6WrerD+9N6S1Lrz1QYe2Fqa56BHhHSXM82NE86SVxEvUkoGegQU+c7b 6rY1JvuJHJzva7+M2XHApYCchCs4a1Yyd1qWB7yThJD57RIyX1TOg0+siV13 R+v6wxhQU0vBovH+5oAWmCZaPNT+F0Uvs3xWAxxaIR9r83wMj9qQeBZTKVzQ 1aFIi15KqAwOp12yWCmrqKTeXhjwYQNd8viCQCGN7AQyPglmzfbuEHalVjz4 oSJX =k281 -----END PGP SIGNATURE----- ---------------- Robert LeBlanc GPG Fingerprint 79A2 9CA4 6CC4 45DD A904 C70E E654 3BB2 FA62 B9F1 On Tue, Jun 9, 2015 at 6:02 AM, Alexandre DERUMIER <aderumier@xxxxxxxxx> wrote: >>>Frankly, I'm a little impressed that without RBD cache we can hit 80K >>>IOPS from 1 VM! > > Note that theses result are not in a vm (fio-rbd on host), so in a vm we'll have overhead. > (I'm planning to send results in qemu soon) > >>>How fast are the SSDs in those 3 OSDs? > > Theses results are with datas in buffer memory of osd nodes. > > When reading fulling on ssd (intel s3500), > > For 1 client, > > I'm around 33k iops without cache and 32k iops with cache, with 1 osd. > I'm around 55k iops without cache and 38k iops with cache, with 3 osd. > > with multiple clients jobs, I can reach around 70kiops by osd , and 250k iops by osd when datas are in buffer. > > (cpus servers/clients are 2x 10 cores 3,1ghz e5 xeon) > > > > small tip : > I'm using tcmalloc for fio-rbd or rados bench to improve latencies by around 20% > > LD_PRELOAD=/usr/lib/libtcmalloc_minimal.so.4 fio ... > LD_PRELOAD=/usr/lib/libtcmalloc_minimal.so.4 rados bench ... > > as a lot of time is spent in malloc/free > > > (qemu support also tcmalloc since some months , I'll bench it too > https://lists.gnu.org/archive/html/qemu-devel/2015-03/msg05372.html) > > > > I'll try to send full bench results soon, from 1 to 18 ssd osd. > > > > > ----- Mail original ----- > De: "Mark Nelson" <mnelson@xxxxxxxxxx> > À: "aderumier" <aderumier@xxxxxxxxx>, "pushpesh sharma" <pushpesh.eck@xxxxxxxxx> > Cc: "ceph-devel" <ceph-devel@xxxxxxxxxxxxxxx>, "ceph-users" <ceph-users@xxxxxxxxxxxxxx> > Envoyé: Mardi 9 Juin 2015 13:36:31 > Objet: Re: [ceph-users] rbd_cache, limiting read on high iops around 40k > > Hi All, > > In the past we've hit some performance issues with RBD cache that we've > fixed, but we've never really tried pushing a single VM beyond 40+K read > IOPS in testing (or at least I never have). I suspect there's a couple > of possibilities as to why it might be slower, but perhaps joshd can > chime in as he's more familiar with what that code looks like. > > Frankly, I'm a little impressed that without RBD cache we can hit 80K > IOPS from 1 VM! How fast are the SSDs in those 3 OSDs? > > Mark > > On 06/09/2015 03:36 AM, Alexandre DERUMIER wrote: >> It's seem that the limit is mainly going in high queue depth (+- > 16) >> >> Here the result in iops with 1client- 4krandread- 3osd - with differents queue depth size. >> rbd_cache is almost the same than without cache with queue depth <16 >> >> >> cache >> ----- >> qd1: 1651 >> qd2: 3482 >> qd4: 7958 >> qd8: 17912 >> qd16: 36020 >> qd32: 42765 >> qd64: 46169 >> >> no cache >> -------- >> qd1: 1748 >> qd2: 3570 >> qd4: 8356 >> qd8: 17732 >> qd16: 41396 >> qd32: 78633 >> qd64: 79063 >> qd128: 79550 >> >> >> ----- Mail original ----- >> De: "aderumier" <aderumier@xxxxxxxxx> >> À: "pushpesh sharma" <pushpesh.eck@xxxxxxxxx> >> Cc: "ceph-devel" <ceph-devel@xxxxxxxxxxxxxxx>, "ceph-users" <ceph-users@xxxxxxxxxxxxxx> >> Envoyé: Mardi 9 Juin 2015 09:28:21 >> Objet: Re: [ceph-users] rbd_cache, limiting read on high iops around 40k >> >> Hi, >> >>>> We tried adding more RBDs to single VM, but no luck. >> >> If you want to scale with more disks in a single qemu vm, you need to use iothread feature from qemu and assign 1 iothread by disk (works with virtio-blk). >> It's working for me, I can scale with adding more disks. >> >> >> My bench here are done with fio-rbd on host. >> I can scale up to 400k iops with 10clients-rbd_cache=off on a single host and around 250kiops 10clients-rbdcache=on. >> >> >> I just wonder why I don't have performance decrease around 30k iops with 1osd. >> >> I'm going to see if this tracker >> http://tracker.ceph.com/issues/11056 >> >> could be the cause. >> >> (My master build was done some week ago) >> >> >> >> ----- Mail original ----- >> De: "pushpesh sharma" <pushpesh.eck@xxxxxxxxx> >> À: "aderumier" <aderumier@xxxxxxxxx> >> Cc: "ceph-devel" <ceph-devel@xxxxxxxxxxxxxxx>, "ceph-users" <ceph-users@xxxxxxxxxxxxxx> >> Envoyé: Mardi 9 Juin 2015 09:21:04 >> Objet: Re: rbd_cache, limiting read on high iops around 40k >> >> Hi Alexandre, >> >> We have also seen something very similar on Hammer(0.94-1). We were doing some benchmarking for VMs hosted on hypervisor (QEMU-KVM, openstack-juno). Each Ubuntu-VM has a RBD as root disk, and 1 RBD as additional storage. For some strange reason it was not able to scale 4K- RR iops on each VM beyond 35-40k. We tried adding more RBDs to single VM, but no luck. However increasing number of VMs to 4 on a single hypervisor did scale to some extent. After this there was no much benefit we got from adding more VMs. >> >> Here is the trend we have seen, x-axis is number of hypervisor, each hypervisor has 4 VM, each VM has 1 RBD:- >> >> >> >> >> VDbench is used as benchmarking tool. We were not saturating network and CPUs at OSD nodes. We were not able to saturate CPUs at hypervisors, and that is where we were suspecting of some throttling effect. However we haven't setted any such limits from nova or kvm end. We tried some CPU pinning and other KVM related tuning as well, but no luck. >> >> We tried the same experiment on a bare metal. It was 4K RR IOPs were scaling from 40K(1 RBD) to 180K(4 RBDs). But after that rather than scaling beyond that point the numbers were actually degrading. (Single pipe more congestion effect) >> >> We never suspected that rbd cache enable could be detrimental to performance. It would nice to route cause the problem if that is the case. >> >> On Tue, Jun 9, 2015 at 11:21 AM, Alexandre DERUMIER < aderumier@xxxxxxxxx > wrote: >> >> >> Hi, >> >> I'm doing benchmark (ceph master branch), with randread 4k qdepth=32, >> and rbd_cache=true seem to limit the iops around 40k >> >> >> no cache >> -------- >> 1 client - rbd_cache=false - 1osd : 38300 iops >> 1 client - rbd_cache=false - 2osd : 69073 iops >> 1 client - rbd_cache=false - 3osd : 78292 iops >> >> >> cache >> ----- >> 1 client - rbd_cache=true - 1osd : 38100 iops >> 1 client - rbd_cache=true - 2osd : 42457 iops >> 1 client - rbd_cache=true - 3osd : 45823 iops >> >> >> >> Is it expected ? >> >> >> >> fio result rbd_cache=false 3 osd >> -------------------------------- >> rbd_iodepth32-test: (g=0): rw=randread, bs=4K-4K/4K-4K/4K-4K, ioengine=rbd, iodepth=32 >> fio-2.1.11 >> Starting 1 process >> rbd engine: RBD version: 0.1.9 >> Jobs: 1 (f=1): [r(1)] [100.0% done] [307.5MB/0KB/0KB /s] [78.8K/0/0 iops] [eta 00m:00s] >> rbd_iodepth32-test: (groupid=0, jobs=1): err= 0: pid=113548: Tue Jun 9 07:48:42 2015 >> read : io=10000MB, bw=313169KB/s, iops=78292, runt= 32698msec >> slat (usec): min=5, max=530, avg=11.77, stdev= 6.77 >> clat (usec): min=70, max=2240, avg=336.08, stdev=94.82 >> lat (usec): min=101, max=2247, avg=347.84, stdev=95.49 >> clat percentiles (usec): >> | 1.00th=[ 173], 5.00th=[ 209], 10.00th=[ 231], 20.00th=[ 262], >> | 30.00th=[ 282], 40.00th=[ 302], 50.00th=[ 322], 60.00th=[ 346], >> | 70.00th=[ 370], 80.00th=[ 402], 90.00th=[ 454], 95.00th=[ 506], >> | 99.00th=[ 628], 99.50th=[ 692], 99.90th=[ 860], 99.95th=[ 948], >> | 99.99th=[ 1176] >> bw (KB /s): min=238856, max=360448, per=100.00%, avg=313402.34, stdev=25196.21 >> lat (usec) : 100=0.01%, 250=15.94%, 500=78.60%, 750=5.19%, 1000=0.23% >> lat (msec) : 2=0.03%, 4=0.01% >> cpu : usr=74.48%, sys=13.25%, ctx=703225, majf=0, minf=12452 >> IO depths : 1=0.1%, 2=0.1%, 4=0.1%, 8=0.8%, 16=87.0%, 32=12.1%, >=64=0.0% >> submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0% >> complete : 0=0.0%, 4=91.6%, 8=3.4%, 16=4.5%, 32=0.4%, 64=0.0%, >=64=0.0% >> issued : total=r=2560000/w=0/d=0, short=r=0/w=0/d=0 >> latency : target=0, window=0, percentile=100.00%, depth=32 >> >> Run status group 0 (all jobs): >> READ: io=10000MB, aggrb=313169KB/s, minb=313169KB/s, maxb=313169KB/s, mint=32698msec, maxt=32698msec >> >> Disk stats (read/write): >> dm-0: ios=0/45, merge=0/0, ticks=0/0, in_queue=0, util=0.00%, aggrios=0/24, aggrmerge=0/21, aggrticks=0/0, aggrin_queue=0, aggrutil=0.00% >> sda: ios=0/24, merge=0/21, ticks=0/0, in_queue=0, util=0.00% >> >> >> >> >> fio result rbd_cache=true 3osd >> ------------------------------ >> >> rbd_iodepth32-test: (g=0): rw=randread, bs=4K-4K/4K-4K/4K-4K, ioengine=rbd, iodepth=32 >> fio-2.1.11 >> Starting 1 process >> rbd engine: RBD version: 0.1.9 >> Jobs: 1 (f=1): [r(1)] [100.0% done] [171.6MB/0KB/0KB /s] [43.1K/0/0 iops] [eta 00m:00s] >> rbd_iodepth32-test: (groupid=0, jobs=1): err= 0: pid=113389: Tue Jun 9 07:47:30 2015 >> read : io=10000MB, bw=183296KB/s, iops=45823, runt= 55866msec >> slat (usec): min=7, max=805, avg=21.26, stdev=15.84 >> clat (usec): min=101, max=4602, avg=478.55, stdev=143.73 >> lat (usec): min=123, max=4669, avg=499.80, stdev=146.03 >> clat percentiles (usec): >> | 1.00th=[ 227], 5.00th=[ 274], 10.00th=[ 306], 20.00th=[ 350], >> | 30.00th=[ 390], 40.00th=[ 430], 50.00th=[ 470], 60.00th=[ 506], >> | 70.00th=[ 548], 80.00th=[ 596], 90.00th=[ 660], 95.00th=[ 724], >> | 99.00th=[ 844], 99.50th=[ 908], 99.90th=[ 1112], 99.95th=[ 1288], >> | 99.99th=[ 2192] >> bw (KB /s): min=115280, max=204416, per=100.00%, avg=183315.10, stdev=15079.93 >> lat (usec) : 250=2.42%, 500=55.61%, 750=38.48%, 1000=3.28% >> lat (msec) : 2=0.19%, 4=0.01%, 10=0.01% >> cpu : usr=60.27%, sys=12.01%, ctx=2995393, majf=0, minf=14100 >> IO depths : 1=0.1%, 2=0.1%, 4=0.2%, 8=13.5%, 16=81.0%, 32=5.3%, >=64=0.0% >> submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0% >> complete : 0=0.0%, 4=95.0%, 8=0.1%, 16=1.0%, 32=4.0%, 64=0.0%, >=64=0.0% >> issued : total=r=2560000/w=0/d=0, short=r=0/w=0/d=0 >> latency : target=0, window=0, percentile=100.00%, depth=32 >> >> Run status group 0 (all jobs): >> READ: io=10000MB, aggrb=183295KB/s, minb=183295KB/s, maxb=183295KB/s, mint=55866msec, maxt=55866msec >> >> Disk stats (read/write): >> dm-0: ios=0/61, merge=0/0, ticks=0/8, in_queue=8, util=0.01%, aggrios=0/29, aggrmerge=0/32, aggrticks=0/8, aggrin_queue=8, aggrutil=0.01% >> sda: ios=0/29, merge=0/32, ticks=0/8, in_queue=8, util=0.01% >> > _______________________________________________ > ceph-users mailing list > ceph-users@xxxxxxxxxxxxxx > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
