Thank you very much, Jason. Our cluster’s target workload is something like monitoring system data center, we need save a lot of video stream into cluster. I have to reconsider test case. Besides, a lot tests to do about the config parameters as you mentioned. Help me a lot, thanks.
On Thu, Nov 15, 2018 at 2:30 PM 赵赵贺东 <zhaohedong@xxxxxxxxx> wrote:
I test in 12 osds cluster, change objecter_inflight_op_bytes from 100MB to 300MB, performance seems not change obviously.
But at the beginning , librbd works in better performance in 12 osds cluster. So it seems meaning less for me.
In a small cluster(12 osds), 4m seq write performance for Librbd VS KRBD is about 0.89 : 1 (177MB/s : 198MB/s ).
In a big cluster (72 osds), 4m seq write performance for Librbd VS KRBD is about 0.38: 1 (420MB/s : 1080MB/s).
Our problem is librbd bad performance in big cluster (72 osds)
But I can not test in 72 osds right now, some other tests are running .
I will test in 72 osds when our cluster is ready.
It is a little hard to understand that objecter_inflight_op_bytes=100MB works well in 12 osds cluster, but works poor in 72 osd clusters.
Dose objecter_inflight_op_bytes not have an effect on krbd, only effect librbd?
Correct -- the "ceph.conf" config settings are for user-space toolingonly. Given the fact that you are writing full 4MiB objects in yourtest, any user-space performance degradation is probably going to bein the librados layer and below. That 100 MiB limit setting will blockthe IO path while it waits for in-flight IO to complete. You alsomight be just hitting the default throughput of the lower-levelmessenger code, so perhaps you need to throw more threads at it(ms_async_op_threads / ms_async_max_op_threads) or change itsthrottles (ms_dispatch_throttle_bytes). Also, depending on yourcluster and krbd versions, perhaps the OSDs are telling your clientsto back-off but only librados is responding to it. You should alsotake into account the validity of your test case -- does it reallymatch your expected workload that you are trying to optimize against?Thanks.
在 2018年11月15日，下午3:50，赵赵贺东 <zhaohedong@xxxxxxxxx> 写道：
Thanks you for your suggestion.
It really give me a lot of inspirations.
I will test as your suggestion, and browse through src/common/config_opts.h to see if I can find some configs performance related.
But, our osd nodes hardware itself is very poor, that is the truth…we have to face it.
Two osds in an arm board, two gb memory and 2*10T hdd disk on board, so one osd has 1gb memory to support 10TB hdd disk, we must try to make cluster works better as we can.
Thanks.
在 2018年11月15日，下午2:08，Jason Dillaman <jdillama@xxxxxxxxxx> 写道：
Attempting to send 256 concurrent 4MiB writes via librbd will pretty
quickly hit the default "objecter_inflight_op_bytes = 100 MiB" limit,
which will drastically slow (stall) librados. I would recommend
re-testing librbd w/ a much higher throttle override.
On Thu, Nov 15, 2018 at 11:34 AM 赵赵贺东 <zhaohedong@xxxxxxxxx> wrote:
Thank you for your attention.
Our test are in run in physical machine environments.
Fio for KRBD:
[seq-write]
description="seq-write"
direct=1
ioengine=libaio
filename=/dev/rbd0
numjobs=1
iodepth=256
group_reporting
rw=write
bs=4M
size=10T
runtime=180
*/dev/rbd0 mapped by rbd_pool/image2, so KRBD & librbd fio test use the same image.
Fio for librbd:
[global]
direct=1
numjobs=1
ioengine=rbd
clientname=admin
pool=rbd_pool
rbdname=image2
invalidate=0 # mandatory
rw=write
bs=4M
size=10T
runtime=180
[rbd_iodepth32]
iodepth=256
Image info:
rbd image 'image2':
size 50TiB in 13107200 objects
order 22 (4MiB objects)
data_pool: ec_rbd_pool
block_name_prefix: rbd_data.8.148bb6b8b4567
format: 2
features: layering, data-pool
flags:
create_timestamp: Wed Nov 14 09:21:18 2018
* data_pool is a EC pool
Pool info:
pool 8 'rbd_pool' replicated size 2 min_size 1 crush_rule 0 object_hash rjenkins pg_num 256 pgp_num 256 last_change 82627 flags hashpspool stripe_width 0 application rbd
pool 9 'ec_rbd_pool' erasure size 6 min_size 5 crush_rule 4 object_hash rjenkins pg_num 256 pgp_num 256 last_change 82649 flags hashpspool,ec_overwrites stripe_width 16384 application rbd
Rbd cache: Off (Because I think in tcmu , rbd cache will mandatory off, and our cluster will export disk by iscsi in furture.)
Thanks!
在 2018年11月15日，下午1:22，Gregory Farnum <gfarnum@xxxxxxxxxx> 写道：
You'll need to provide more data about how your test is configured and run for us to have a good idea. IIRC librbd is often faster than krbd because it can support newer features and things, but krbd may have less overhead and is not dependent on the VM's driver configuration in QEMU...
On Thu, Nov 15, 2018 at 8:22 AM 赵赵贺东 <zhaohedong@xxxxxxxxx> wrote:
Hi cephers,
All our cluster osds are deployed in armhf.
Could someone say something about what is the rational performance rates for librbd VS KRBD ?
Or rational performance loss range when we use librbd compare to KRBD.
I googled a lot, but I could not find a solid criterion.
In fact , it confused me for a long time.
About our tests:
In a small cluster(12 osds), 4m seq write performance for Librbd VS KRBD is about 0.89 : 1 (177MB/s : 198MB/s ).
In a big cluster (72 osds), 4m seq write performance for Librbd VS KRBD is about 0.38: 1 (420MB/s : 1080MB/s).
We expect even increase osd numbers, Librbd performance can keep being close to KRBD.
PS: Librbd performance are tested both in fio rbd engine & iscsi (tcmu+librbd).
Thanks.
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
--
Jason
-- Jason
|
