Hi Cephers,
We're in the middle of trying to triage the issue with ceph cluster
running 0.80.9 which was reported by Songbo and seeking for you
experts' advices.
In fact, per our testing the process of stopping an working OSD and
starting it again will lead to a huge performance downgrade. In
other words, this issue can be reproduced quite easily, and we
cannot lower the impact of the state of OSD by tuning the settings
like
osd_max_backfills/osd_recovery_max_chunk/osd_recovery_max_active.
Through looking into the source code, we notice that the requests
issued by clients will be queued firstly when the corresponding PGs
are in some certain states (like recovering and backfill) and then
processed. During this period, the IOPS outputted by fio drops
significantly(from 2000 to 60). What we can think of this is to
guarantee the data consistency, are we correct? If that's the
design, we're wondering how Ceph can support the applications that
are performance-sensitive? Is there any other parameters we can
tuning to reduce the impact?
Thanks,
Jevon
On 26/10/15 13:27, wangsongbo wrote:
Hi
all,
When an osd is started, I will get a lot of slow requests from the
corresponding osd log, as follows:
2015-10-26 03:42:51.593961 osd.4 [WRN] slow request 3.967808
seconds old, received at 2015-10-26 03:42:47.625968:
osd_repop(client.2682003.0:2686048 43.fcf
d1ddfcf/rbd_data.196483222ac2db.0000000000000010/head//43 v
9744'347845) currently commit_sent
2015-10-26 03:42:51.593964 osd.4 [WRN] slow request 3.964537
seconds old, received at 2015-10-26 03:42:47.629239:
osd_repop(client.2682003.0:2686049 43.b4b
cbcbbb4b/rbd_data.196483222ac2db.000000000000020b/head//43 v
9744'193029) currently commit_sent
2015-10-26 03:42:52.594166 osd.4 [WRN] 40 slow requests, 17
included below; oldest blocked for > 53.692556 secs
2015-10-26 03:42:52.594172 osd.4 [WRN] slow request 2.272928
seconds old, received at 2015-10-26 03:42:50.321151:
osd_repop(client.3684690.0:191908 43.540
f1858540/rbd_data.1fc5ca7429fc17.0000000000000280/head//43 v
9744'63645) currently commit_sent
2015-10-26 03:42:52.594175 osd.4 [WRN] slow request 2.270618
seconds old, received at 2015-10-26 03:42:50.323461:
osd_op(client.3684690.0:191911
rbd_data.1fc5ca7429fc17.0000000000000209 [write 2633728~4096]
43.72b9f039 ack+ondisk+write e9744) currently commit_sent
2015-10-26 03:42:52.594264 osd.4 [WRN] slow request 4.968252
seconds old, received at 2015-10-26 03:42:47.625828:
osd_repop(client.2682003.0:2686047 43.b4b
cbcbbb4b/rbd_data.196483222ac2db.000000000000020b/head//43 v
9744'193028) currently commit_sent
2015-10-26 03:42:52.594266 osd.4 [WRN] slow request 4.968111
seconds old, received at 2015-10-26 03:42:47.625968:
osd_repop(client.2682003.0:2686048 43.fcf
d1ddfcf/rbd_data.196483222ac2db.0000000000000010/head//43 v
9744'347845) currently commit_sent
2015-10-26 03:42:52.594318 osd.4 [WRN] slow request 4.964841
seconds old, received at 2015-10-26 03:42:47.629239:
osd_repop(client.2682003.0:2686049 43.b4b
cbcbbb4b/rbd_data.196483222ac2db.000000000000020b/head//43 v
9744'193029) currently commit_sent
2015-10-26 03:42:53.594527 osd.4 [WRN] 40 slow requests, 16
included below; oldest blocked for > 54.692945 secs
2015-10-26 03:42:53.594533 osd.4 [WRN] slow request 16.004669
seconds old, received at 2015-10-26 03:42:37.589800:
osd_repop(client.2682003.0:2686041 43.b4b
cbcbbb4b/rbd_data.196483222ac2db.000000000000020b/head//43 v
9744'193024) currently commit_sent
2015-10-26 03:42:53.594536 osd.4 [WRN] slow request 16.003889
seconds old, received at 2015-10-26 03:42:37.590580:
osd_repop(client.2682003.0:2686040 43.fcf
d1ddfcf/rbd_data.196483222ac2db.0000000000000010/head//43 v
9744'347842) currently commit_sent
2015-10-26 03:42:53.594538 osd.4 [WRN] slow request 16.000954
seconds old, received at 2015-10-26 03:42:37.593515:
osd_repop(client.2682003.0:2686042 43.b4b
cbcbbb4b/rbd_data.196483222ac2db.000000000000020b/head//43 v
9744'193025) currently commit_sent
2015-10-26 03:42:53.594541 osd.4 [WRN] slow request 29.138828
seconds old, received at 2015-10-26 03:42:24.455641:
osd_repop(client.4764855.0:65121 43.dbe
169a9dbe/rbd_data.49a7a4633ac0b1.0000000000000021/head//43 v
9744'12509) currently commit_sent
2015-10-26 03:42:53.594543 osd.4 [WRN] slow request 15.998814
seconds old, received at 2015-10-26 03:42:37.595656:
osd_repop(client.1800547.0:1205399 43.cc5
9285ecc5/rbd_data.1b794560c6e2ea.00000000000000d0/head//43 v
9744'36732) currently commit_sent
2015-10-26 03:42:54.594892 osd.4 [WRN] 39 slow requests, 17
included below; oldest blocked for > 55.693227 secs
2015-10-26 03:42:54.594908 osd.4 [WRN] slow request 4.273600
seconds old, received at 2015-10-26 03:42:50.321151:
osd_repop(client.3684690.0:191908 43.540
f1858540/rbd_data.1fc5ca7429fc17.0000000000000280/head//43 v
9744'63645) currently commit_sent
2015-10-26 03:42:54.594911 osd.4 [WRN] slow request 4.271290
seconds old, received at 2015-10-26 03:42:50.323461:
osd_op(client.3684690.0:191911
rbd_data.1fc5ca7429fc17.0000000000000209 [write 2633728~4096]
43.72b9f039 ack+ondisk+write e9744) currently commit_sent
Meanwhile, I run fio process with the rbd ioengine.
The iops of read and write were too small to response any IO from
the fio process,
In other words, when an osd is started, the IO of the whole
cluster will be blocked.
Is there some parameter to adjust ?
How to explain this problem?
The results of running fio process were as fllows:
ebs: (g=0): rw=randrw, bs=8K-8K/8K-8K/8K-8K, ioengine=rbd,
iodepth=64
fio-2.2.9-20-g1520
Starting 1 thread
rbd engine: RBD version: 0.1.9
Jobs: 1 (f=1): [m(1)] [0.3% done] [0KB/0KB/0KB /s] [0/0/0 iops]
[eta 05h:10m:14s]
ebs: (groupid=0, jobs=1): err= 0: pid=40323: Mon Oct 26 04:02:00
2015
read : io=10904KB, bw=175183B/s, *iops=21*, runt= 63737msec
slat (usec): min=0, max=61, avg= 1.11, stdev= 3.16
clat (msec): min=1, max=63452, avg=1190.04, stdev=6046.28
lat (msec): min=1, max=63452, avg=1190.04, stdev=6046.28
clat percentiles (msec):
| 1.00th=[ 3], 5.00th=[ 4], 10.00th=[ 5],
20.00th=[ 6],
| 30.00th=[ 6], 40.00th=[ 7], 50.00th=[ 8],
60.00th=[ 9],
| 70.00th=[ 10], 80.00th=[ 12], 90.00th=[ 14],
95.00th=[ 709],
| 99.00th=[16712], 99.50th=[16712], 99.90th=[16712],
99.95th=[16712],
| 99.99th=[16712]
bw (KB /s): min= 129, max= 2038, per=100.00%, avg=774.00,
stdev=1094.73
write: io=24976KB, bw=401264B/s,*iops=48,* runt= 63737msec
slat (usec): min=0, max=40, avg= 2.48, stdev= 3.30
clat (msec): min=2, max=31379, avg=786.91, stdev=4829.02
lat (msec): min=2, max=31379, avg=786.92, stdev=4829.02
clat percentiles (msec):
| 1.00th=[ 4], 5.00th=[ 6], 10.00th=[ 6],
20.00th=[ 8],
| 30.00th=[ 9], 40.00th=[ 9], 50.00th=[ 11],
60.00th=[ 12],
| 70.00th=[ 13], 80.00th=[ 15], 90.00th=[ 19],
95.00th=[ 29],
| 99.00th=[16712], 99.50th=[16712], 99.90th=[16712],
99.95th=[16712],
| 99.99th=[16712]
bw (KB /s): min= 317, max= 5228, per=100.00%, avg=1957.33,
stdev=2832.48
lat (msec) : 2=0.07%, 4=3.23%, 10=52.29%, 20=35.76%, 50=4.41%
lat (msec) : 750=1.36%, 1000=0.02%, 2000=0.02%,
>=2000=2.83%
cpu : usr=0.03%, sys=0.00%, ctx=228, majf=0, minf=19
IO depths : 1=0.1%, 2=0.1%, 4=0.4%, 8=3.9%, 16=18.9%,
32=73.1%, >=64=3.5%
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=97.5%, 8=0.0%, 16=0.2%, 32=0.9%,
64=1.5%, >=64=0.0%
issued : total=r=1363/w=3122/d=0, short=r=0/w=0/d=0,
drop=r=0/w=0/d=0
latency : target=0, window=0, percentile=100.00%, depth=64
Run status group 0 (all jobs):
READ: io=10904KB, aggrb=171KB/s, minb=171KB/s, maxb=171KB/s,
mint=63737msec, maxt=63737msec
WRITE: io=24976KB, aggrb=391KB/s, minb=391KB/s, maxb=391KB/s,
mint=63737msec, maxt=63737msec
Disk stats (read/write):
dm-0: ios=81/34, merge=0/0, ticks=472/26, in_queue=498,
util=0.30%, aggrios=143/102, aggrmerge=106/122,
aggrticks=1209/134, aggrin_queue=1343, aggrutil=0.60%
sdd: ios=143/102, merge=106/122, ticks=1209/134, in_queue=1343,
util=0.60%
Thanks and Regards,
WangSongbo
On 15/10/22 下午10:43, wangsongbo wrote:
Hi all,
When an osd is started, relative IO will be blocked.
According to the test result,the larger iops the clients send ,
the longer it will take to elapse.
Adjustment on all the parameters associate with recovery
operations was also found useless.
How to reduce the impact of this process on the IO ?
Thanks and Regards,
WangSongbo
--
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
|
