On 1/10/20 5:28 PM, Liu, Chunmei wrote:
-----Original Message-----
From: ceph-devel-owner@xxxxxxxxxxxxxxx <ceph-devel-owner@xxxxxxxxxxxxxxx>
On Behalf Of Roman Penyaev
Sent: Friday, January 10, 2020 10:54 AM
To: kefu chai <tchaikov@xxxxxxxxx>
Cc: Radoslaw Zarzynski <rzarzyns@xxxxxxxxxx>; Samuel Just
<sjust@xxxxxxxxxx>; The Esoteric Order of the Squid Cybernetic <ceph-
devel@xxxxxxxxxxxxxxx>
Subject: Re: crimson-osd vs legacy-osd: should the perf difference be already
noticeable?
On 2020-01-10 17:18, kefu chai wrote:
[skip]
First thing that catches my eye is that for small blocks there is no
big difference at all, but as the block increases, crimsons iops
starts to
that's also our findings. and it's expected. as async messenger uses
the same reactor model as seastar does. actually its original
implementation was adapted from seastar's socket stream
implementation.
Hm, regardless of model messenger should not be a bottleneck. Take a look on
the results of fio_ceph_messenger load (runs pure messenger), I can squeeze
IOPS=89.8k, BW=351MiB/s on 4k block size, iodepth=32.
(also good example https://github.com/ceph/ceph/pull/26932 , almost
~200k)
With PG layer (memstore_debug_omit_block_device_write=true option) I can
reach 40k iops max. Without PG layer (immediate completion from the
transport callback, osd_immediate_completions=true) I get almost 60k.
Seems that here starts playing costs on client side and these costs prevail.
decline. Can it be the transport issue? Can be tested as well.
because seastar's socket facility reads from the wire with 4K chunk
size, while classic OSD's async messenger reads the payload with the
size suggested by the header. so when it comes to larger block size,
it takes crimson-osd multiple syscalls and memcpy calls to read the
request from wire, that's why classic OSD wins in this case.
Do you plan to fix that?
have you tried to use multiple fio clients to saturate CPU capacity of
OSD nodes?
Not yet. But regarding CPU I have these numbers:
output of pidstat while rbd.fio is running, 4k block only:
legacy-osd
[roman@dell ~]$ pidstat 1 -p 109930
Linux 5.3.13-arch1-1 (dell) 01/09/2020 _x86_64_ (8 CPU)
03:51:49 PM UID PID %usr %system %guest %wait %CPU
CPU Command
03:51:51 PM 1000 109930 14.00 8.00 0.00 0.00 22.00
1 ceph-osd
03:51:52 PM 1000 109930 40.00 19.00 0.00 0.00 59.00
1 ceph-osd
03:51:53 PM 1000 109930 44.00 17.00 0.00 0.00 61.00
1 ceph-osd
03:51:54 PM 1000 109930 40.00 20.00 0.00 0.00 60.00
1 ceph-osd
03:51:55 PM 1000 109930 39.00 18.00 0.00 0.00 57.00
1 ceph-osd
03:51:56 PM 1000 109930 41.00 20.00 0.00 0.00 61.00
1 ceph-osd
03:51:57 PM 1000 109930 41.00 15.00 0.00 0.00 56.00
1 ceph-osd
03:51:58 PM 1000 109930 42.00 16.00 0.00 0.00 58.00
1 ceph-osd
03:51:59 PM 1000 109930 42.00 15.00 0.00 0.00 57.00
1 ceph-osd
03:52:00 PM 1000 109930 43.00 15.00 0.00 0.00 58.00
1 ceph-osd
03:52:01 PM 1000 109930 24.00 12.00 0.00 0.00 36.00
1 ceph-osd
crimson-osd
[roman@dell ~]$ pidstat 1 -p 108141
Linux 5.3.13-arch1-1 (dell) 01/09/2020 _x86_64_ (8 CPU)
03:47:50 PM UID PID %usr %system %guest %wait %CPU
CPU Command
03:47:55 PM 1000 108141 67.00 11.00 0.00 0.00 78.00
0 crimson-osd
03:47:56 PM 1000 108141 79.00 12.00 0.00 0.00 91.00
0 crimson-osd
03:47:57 PM 1000 108141 81.00 9.00 0.00 0.00 90.00
0 crimson-osd
03:47:58 PM 1000 108141 78.00 12.00 0.00 0.00 90.00
0 crimson-osd
03:47:59 PM 1000 108141 78.00 12.00 0.00 1.00 90.00
0 crimson-osd
03:48:00 PM 1000 108141 78.00 13.00 0.00 0.00 91.00
0 crimson-osd
03:48:01 PM 1000 108141 79.00 13.00 0.00 0.00 92.00
0 crimson-osd
03:48:02 PM 1000 108141 78.00 12.00 0.00 0.00 90.00
0 crimson-osd
03:48:03 PM 1000 108141 77.00 11.00 0.00 0.00 88.00
0 crimson-osd
03:48:04 PM 1000 108141 79.00 12.00 0.00 1.00 91.00
0 crimson-osd
Seems quite saturated, almost twice more than legacy-osd. Did you see
something similar?
Crimson-osd (seastar) use epoll, by default, it will use more cpu capacity,(you can change epoll mode setting to reduce it), add Ma, Jianpeng in the thread since he did more study on it.
BTW, by default crimson-osd is one thread, and legacy ceph-osd (3 threads for async messenger, 2x8 threads for osd (SDD), finisher thread etc,) ,so by default setting, it is 1 thread compare to over 10 threads work, it is expected crimson-osd not show obvious difference. you can change the default thread number for legacy ceph-osd(such as thread=1 for each layer to see more difference.)
BTW, please use release build to do test.
Crimson-osd is aysnc model, if workload is very light, can't take more advantage of it.
--
Roman
FWIW I can drive the classical OSD pretty hard and get around 70-80K
IOPS out of a single OSD, but as Kefu says above it will consume a
larger number of cores. I do think per-OSD throughput is still
important to look at, but the per-OSD efficiency numbers as Radek has
been testing (I gathered some for classical OSD a while back for him)
are probably going to be more important overall.
Mark
