Hi Somnath and Igor,I was able to successfully bisect to hit the commit where the regression occurs. It's https://github.com/ceph/ceph/commit/0e8294c9a. This probably explains why Somnath isn't seeing it since he has csums disabled. It appears that we previously to set the csum_order to the block_size_order, but now set it to the MAX of the block size order or the "preferred" csum order which is based on the trailing zeros of the "expected write size" in the onode. I am guessing this means that since the data was filled to the disk using 4M sequential write, the onode csum order is much higher than it was prior to the patch and that is greatly hurting 4K random reads of those objects.
I am going to try applying a patch to revert this change and see how things go.
Mark On 07/14/2016 07:42 PM, Somnath Roy wrote:
Mark, In fact, I was wrong saying it is way below filestore. I found out my client cpu was saturating ~160K 4K RR iops. I have added another client (and another 4TB image) and it is scaling up well. I am getting ~320K iops (4K RR) saturating my 2 osd node cpu almost. So, pretty similar behavior like filestore. I have reduced bluestore_cache_size = 100MB and memory consumption is also controlled for my 10 min run at least. Thanks & Regards Somnath -----Original Message----- From: Somnath Roy Sent: Thursday, July 14, 2016 10:36 AM To: 'Mark Nelson'; Igor Fedotov Cc: ceph-devel (ceph-devel@xxxxxxxxxxxxxxx) Subject: RE: Bluestore read performance Thanks Igor ! I was not aware of cache shards.. I am running with 25 shards (Generally, we need more shards for the parallelism.), so, it will take ~12G per OSD for the cash only. That is probably clarifies why we are seeing memory spikes.. Regards Somnath -----Original Message----- From: Mark Nelson [mailto:mnelson@xxxxxxxxxx] Sent: Thursday, July 14, 2016 10:28 AM To: Igor Fedotov; Somnath Roy Cc: ceph-devel (ceph-devel@xxxxxxxxxxxxxxx) Subject: Re: Bluestore read performance We are leaking or at least spiking memory much higher than than in some cases. In my tests I can get them up to about 9GB RSS per OSD. I only have 4 nodes per OSD and 64GB of RAM though so I'm not hitting swap (in fact these nodes don't have swap). Mark On 07/14/2016 12:17 PM, Igor Fedotov wrote:Somnath, Mark I have a question and some comments w.r.t. memory swapping. What's amount of RAM do you have at your nodes? How many of it is taken by OSDs? I can see that each BlueStore OSD may occupy bluestore_buffer_cache_size * osd_op_num_shards = 512M * 5 = 2.5G (by default) for buffer cache. Hence in Somnath's environment one might expect up to 20G taken for the cache. Does that estimation correlate with the real life? Thanks, Igor On 14.07.2016 19:50, Somnath Roy wrote:Mark, As we discussed in today's meeting , I ran 100% RR with the following fio profile on a single image of 4TB. Did precondition the entire image with 1M seq write. I have total of 16 OSDs over 2 nodes. [global] ioengine=rbd clientname=admin pool=recovery_test rbdname=recovery_image invalidate=0 # mandatory rw=randread bs=4k direct=1 time_based runtime=30m numjobs=8 group_reporting [rbd_iodepth32] iodepth=128 Here is the ceph.conf option I used for Bluestore. osd_op_num_threads_per_shard = 2 osd_op_num_shards = 25 bluestore_rocksdb_options = "max_write_buffer_number=16,min_write_buffer_number_to_merge=16,recyc le_log_file_num=16,compaction_threads=32,flusher_threads=4, max_background_compactions=32,max_background_flushes=8,max_bytes_for_level_base=5368709120,write_buffer_size=83886080,level0_file_num_compaction_trigger=4,level0_slowdown_writes_trigger=400,level0_stop_writes_trigger=800" rocksdb_cache_size = 4294967296 #bluestore_min_alloc_size = 16384 bluestore_min_alloc_size = 4096 bluestore_csum = false bluestore_csum_type = none bluestore_bluefs_buffered_io = false bluestore_max_ops = 30000 bluestore_max_bytes = 629145600 Here is the output I got. rbd_iodepth32: (g=0): rw=randread, bs=4K-4K/4K-4K/4K-4K, ioengine=rbd, iodepth=128 ... fio-2.1.11 Starting 8 processes rbd engine: RBD version: 0.1.10 rbd engine: RBD version: 0.1.10 rbd engine: RBD version: 0.1.10 rbd engine: RBD version: 0.1.10 rbd engine: RBD version: 0.1.10 rbd engine: RBD version: 0.1.10 rbd engine: RBD version: 0.1.10 rbd engine: RBD version: 0.1.10 ^Cbs: 8 (f=8): [r(8)] [9.4% done] [179.5MB/0KB/0KB /s] [45.1K/0/0 iops] [eta 27m:12s] fio: terminating on signal 2 rbd_iodepth32: (groupid=0, jobs=8): err= 0: pid=1266211: Thu Jul 14 09:42:28 2016 read : io=95898MB, bw=583425KB/s, iops=145856, runt=168316msec slat (usec): min=0, max=13967, avg= 4.56, stdev=38.79 clat (usec): min=15, max=1949.3K, avg=6941.73, stdev=16018.84 lat (usec): min=225, max=1949.3K, avg=6946.30, stdev=16018.92 clat percentiles (usec): | 1.00th=[ 876], 5.00th=[ 2024], 10.00th=[ 2672], 20.00th=[ 3312], | 30.00th=[ 3824], 40.00th=[ 4320], 50.00th=[ 5024], 60.00th=[ 5920], | 70.00th=[ 7072], 80.00th=[ 8768], 90.00th=[11840], 95.00th=[15040], | 99.00th=[22400], 99.50th=[27264], 99.90th=[248832], 99.95th=[366592], | 99.99th=[602112] I was getting > 600MB/s before memory started swapping for me and the fio output came down. I never tested Bluestore read before, but, it is definitely lower than Filestore for me. But, it is far better than you are getting it seems (?). Do you mind trying with the above ceph.conf option as well ? My ceph version : ceph version 11.0.0-536-g8df0c5b (8df0c5bcd90d80e9b309b2a9007b778f7b829edf) Thanks & Regards Somnath 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). -- 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-- 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.htmlPLEASE 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). -- 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
-- 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
