Hi Sebastian,
If you are trying with the latest Ceph master, there are some changes we made that will be increasing your read performance from SSD a factor of ~5X if the ios are hitting the disks. Otherwise, the serving from memory the improvement is even more. The single OSD will be cpu bound with increasing number of clients eventually both reading from disk and memory scenario. Some new config option are introduced and here are those.
osd_op_num_threads_per_shard
osd_op_num_shards
throttler_perf_counter
osd_enable_op_tracker
filestore_fd_cache_size
filestore_fd_cache_shards
The work pool for the io path is now sharded and the above options are for controlling this. Osd_op_threads are no longer in the io path. Also, the filestore FDcache is sharded now.
In my setup(64GB RAM, 40 core CPU with HT enabled) the following config file on a single OSD is giving optimum result for 4k RR read.
[global]
filestore_xattr_use_omap = true
debug_lockdep = 0/0
debug_context = 0/0
debug_crush = 0/0
debug_buffer = 0/0
debug_timer = 0/0
debug_filer = 0/0
debug_objecter = 0/0
debug_rados = 0/0
debug_rbd = 0/0
debug_journaler = 0/0
debug_objectcatcher = 0/0
debug_client = 0/0
debug_osd = 0/0
debug_optracker = 0/0
debug_objclass = 0/0
debug_filestore = 0/0
debug_journal = 0/0
debug_ms = 0/0
debug_monc = 0/0
debug_tp = 0/0
debug_auth = 0/0
debug_finisher = 0/0
debug_heartbeatmap = 0/0
debug_perfcounter = 0/0
debug_asok = 0/0
debug_throttle = 0/0
debug_mon = 0/0
debug_paxos = 0/0
debug_rgw = 0/0
osd_op_threads = 5
osd_op_num_threads_per_shard = 1
osd_op_num_shards = 25
#osd_op_num_sharded_pool_threads = 25
filestore_op_threads = 4
ms_nocrc = true
filestore_fd_cache_size = 64
filestore_fd_cache_shards = 32
cephx sign messages = false
cephx require signatures = false
ms_dispatch_throttle_bytes = 0
throttler_perf_counter = false
[osd]
osd_client_message_size_cap = 0
osd_client_message_cap = 0
osd_enable_op_tracker = false
What I saw optracker is one of the major bottleneck and we are in process of optimizing that. For now, optracker enabled/disabled code introduced. Also, there are several bottlenecks in the filestore level are removed.
Unfortunately, we are yet to optimize the write path. All of these should help the write path as well, but, write path improvement will not be visible till all the lock serialization are removed.
Thanks & Regards
Somnath
-----Original Message-----
From: ceph-users [mailto:ceph-users-bounces@xxxxxxxxxxxxxx] On Behalf Of Sebastien Han
Sent: Thursday, August 28, 2014 9:12 AM
To: ceph-users
Cc: Mark Nelson
Subject: [Single OSD performance on SSD] Can't go over 3, 2K IOPS
Hey all,
It has been a while since the last thread performance related on the ML :p I've been running some experiment to see how much I can get from an SSD on a Ceph cluster.
To achieve that I did something pretty simple:
* Debian wheezy 7.6
* kernel from debian 3.14-0.bpo.2-amd64
* 1 cluster, 3 mons (i'd like to keep this realistic since in a real deployment i'll use 3)
* 1 OSD backed by an SSD (journal and osd data on the same device)
* 1 replica count of 1
* partitions are perfectly aligned
* io scheduler is set to noon but deadline was showing the same results
* no updatedb running
About the box:
* 32GB of RAM
* 12 cores with HT @ 2,4 GHz
* WB cache is enabled on the controller
* 10Gbps network (doesn't help here)
The SSD is a 200G Intel DC S3700 and is capable of delivering around 29K iops with random 4k writes (my fio results) As a benchmark tool I used fio with the rbd engine (thanks deutsche telekom guys!).
O_DIECT and D_SYNC don't seem to be a problem for the SSD:
# dd if=/dev/urandom of=rand.file bs=4k count=65536
65536+0 records in
65536+0 records out
268435456 bytes (268 MB) copied, 29.5477 s, 9.1 MB/s
# du -sh rand.file
256M rand.file
# dd if=rand.file of=/dev/sdo bs=4k count=65536 oflag=dsync,direct
65536+0 records in
65536+0 records out
268435456 bytes (268 MB) copied, 2.73628 s, 98.1 MB/s
See my ceph.conf:
[global]
auth cluster required = cephx
auth service required = cephx
auth client required = cephx
fsid = 857b8609-8c9b-499e-9161-2ea67ba51c97
osd pool default pg num = 4096
osd pool default pgp num = 4096
osd pool default size = 2
osd crush chooseleaf type = 0
debug lockdep = 0/0
debug context = 0/0
debug crush = 0/0
debug buffer = 0/0
debug timer = 0/0
debug journaler = 0/0
debug osd = 0/0
debug optracker = 0/0
debug objclass = 0/0
debug filestore = 0/0
debug journal = 0/0
debug ms = 0/0
debug monc = 0/0
debug tp = 0/0
debug auth = 0/0
debug finisher = 0/0
debug heartbeatmap = 0/0
debug perfcounter = 0/0
debug asok = 0/0
debug throttle = 0/0
[mon]
mon osd down out interval = 600
mon osd min down reporters = 13
[mon.ceph-01]
host = ceph-01
mon addr = 172.20.20.171
[mon.ceph-02]
host = ceph-02
mon addr = 172.20.20.172
[mon.ceph-03]
host = ceph-03
mon addr = 172.20.20.173
debug lockdep = 0/0
debug context = 0/0
debug crush = 0/0
debug buffer = 0/0
debug timer = 0/0
debug journaler = 0/0
debug osd = 0/0
debug optracker = 0/0
debug objclass = 0/0
debug filestore = 0/0
debug journal = 0/0
debug ms = 0/0
debug monc = 0/0
debug tp = 0/0
debug auth = 0/0
debug finisher = 0/0
debug heartbeatmap = 0/0
debug perfcounter = 0/0
debug asok = 0/0
debug throttle = 0/0
[osd]
osd mkfs type = xfs
osd mkfs options xfs = -f -i size=2048
osd mount options xfs = rw,noatime,logbsize=256k,delaylog
osd journal size = 20480
cluster_network = 172.20.20.0/24
public_network = 172.20.20.0/24
osd mon heartbeat interval = 30
# Performance tuning
filestore merge threshold = 40
filestore split multiple = 8
osd op threads = 8
# Recovery tuning
osd recovery max active = 1
osd max backfills = 1
osd recovery op priority = 1
debug lockdep = 0/0
debug context = 0/0
debug crush = 0/0
debug buffer = 0/0
debug timer = 0/0
debug journaler = 0/0
debug osd = 0/0
debug optracker = 0/0
debug objclass = 0/0
debug filestore = 0/0
debug journal = 0/0
debug ms = 0/0
debug monc = 0/0
debug tp = 0/0
debug auth = 0/0
debug finisher = 0/0
debug heartbeatmap = 0/0
debug perfcounter = 0/0
debug asok = 0/0
debug throttle = 0/0
Disabling all debugging made me win 200/300 more IOPS.
See my fio template:
[global]
#logging
#write_iops_log=write_iops_log
#write_bw_log=write_bw_log
#write_lat_log=write_lat_lo
time_based
runtime=60
ioengine=rbd
clientname=admin
pool=test
rbdname=fio
invalidate=0 # mandatory
#rw=randwrite
rw=write
bs=4k
#bs=32m
size=5G
group_reporting
[rbd_iodepth32]
iodepth=32
direct=1
See my rio output:
rbd_iodepth32: (g=0): rw=write, bs=4K-4K/4K-4K/4K-4K, ioengine=rbd, iodepth=32 fio-2.1.11-14-gb74e Starting 1 process rbd engine: RBD version: 0.1.8
Jobs: 1 (f=1): [W(1)] [100.0% done] [0KB/12876KB/0KB /s] [0/3219/0 iops] [eta 00m:00s]
rbd_iodepth32: (groupid=0, jobs=1): err= 0: pid=32116: Thu Aug 28 00:28:26 2014
write: io=771448KB, bw=12855KB/s, iops=3213, runt= 60010msec
slat (usec): min=42, max=1578, avg=66.50, stdev=16.96
clat (msec): min=1, max=28, avg= 9.85, stdev= 1.48
lat (msec): min=1, max=28, avg= 9.92, stdev= 1.47
clat percentiles (usec):
| 1.00th=[ 6368], 5.00th=[ 8256], 10.00th=[ 8640], 20.00th=[ 9152],
| 30.00th=[ 9408], 40.00th=[ 9664], 50.00th=[ 9792], 60.00th=[10048],
| 70.00th=[10176], 80.00th=[10560], 90.00th=[10944], 95.00th=[11456],
| 99.00th=[13120], 99.50th=[16768], 99.90th=[25984], 99.95th=[27008],
| 99.99th=[28032]
bw (KB /s): min=11864, max=13808, per=100.00%, avg=12864.36, stdev=407.35
lat (msec) : 2=0.03%, 4=0.54%, 10=59.79%, 20=39.24%, 50=0.41%
cpu : usr=19.15%, sys=4.69%, ctx=326309, majf=0, minf=426088
IO depths : 1=0.1%, 2=0.1%, 4=0.1%, 8=0.1%, 16=33.9%, 32=66.1%, >=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=99.6%, 8=0.4%, 16=0.1%, 32=0.1%, 64=0.0%, >=64=0.0%
issued : total=r=0/w=192862/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):
WRITE: io=771448KB, aggrb=12855KB/s, minb=12855KB/s, maxb=12855KB/s, mint=60010msec, maxt=60010msec
Disk stats (read/write):
dm-1: ios=0/49, merge=0/0, ticks=0/12, in_queue=12, util=0.01%, aggrios=0/22, aggrmerge=0/27, aggrticks=0/12, aggrin_queue=12, aggrutil=0.01%
sda: ios=0/22, merge=0/27, ticks=0/12, in_queue=12, util=0.01%
I tried to tweak several parameters like:
filestore_wbthrottle_xfs_ios_start_flusher = 10000 filestore_wbthrottle_xfs_ios_hard_limit = 10000 filestore_wbthrottle_btrfs_ios_start_flusher = 10000 filestore_wbthrottle_btrfs_ios_hard_limit = 10000 filestore queue max ops = 2000
But didn't any improvement.
Then I tried other things:
* Increasing the io_depth up to 256 or 512 gave me between 50 to 100 more IOPS but it's not a realistic workload anymore and not that significant.
* adding another SSD for the journal, still getting 3,2K IOPS
* I tried with rbd bench and I also got 3K IOPS
* I ran the test on a client machine and then locally on the server, still getting 3,2K IOPS
* put the journal in memory, still getting 3,2K IOPS
* with 2 clients running the test in parallel I got a total of 3,6K IOPS but I don't seem to be able to go over
* I tried is to add another OSD to that SSD, so I had 2 OSD and 2 journals on 1 SSD, got 4,5K IOPS YAY!
Given the results of the last time it seems that something is limiting the number of IOPS per OSD process.
Running the test on a client or locally didn't show any difference.
So it looks to me that there is some contention within Ceph that might cause this.
I also ran perf and looked at the output, everything looks decent, but someone might want to have a look at it :).
We have been able to reproduce this on 3 distinct platforms with some deviations (because of the hardware) but the behaviour is the same.
Any thoughts will be highly appreciated, only getting 3,2k out of an 29K IOPS SSD is a bit frustrating :).
Cheers.
----
S?bastien Han
Cloud Architect
"Always give 100%. Unless you're giving blood."
Phone: +33 (0)1 49 70 99 72
Mail: sebastien.han at enovance.com
Address : 11 bis, rue Roqu?pine - 75008 Paris Web : www.enovance.com - Twitter : @enovance
________________________________
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).
