>>Is there anything changed from Hammer to Jewel that might be affecting the qemu-img convert performance? maybe object map for exclusive lock ? (I think it could be a little bit slower when objects are created first) you could test it, create the target rbd volume, disable exclusive lock,objet map, and try qemu-img convert. ----- Mail original ----- De: "Mahesh Jambhulkar" <mahesh.jambhulkar@xxxxxxxxx> À: "aderumier" <aderumier@xxxxxxxxx> Cc: "dillaman" <dillaman@xxxxxxxxxx>, "ceph-users" <ceph-users@xxxxxxxxxxxxxx> Envoyé: Vendredi 21 Juillet 2017 14:38:20 Objet: Re: qemu-img convert vs rbd import performance Thanks Alexandre! We were using ceph - Hammer before and we never had these performance issues with qemu-img convert. Is there anything changed from Hammer to Jewel that might be affecting the qemu-img convert performance? On Fri, Jul 21, 2017 at 2:24 PM, Alexandre DERUMIER < [ mailto:aderumier@xxxxxxxxx | aderumier@xxxxxxxxx ] > wrote: It's already in qemu 2.9 [ http://git.qemu.org/?p=qemu.git;a=commit;h=2d9187bc65727d9dd63e2c410b5500add3db0b0d | http://git.qemu.org/?p=qemu.git;a=commit;h=2d9187bc65727d9dd63e2c410b5500add3db0b0d ] " This patches introduces 2 new cmdline parameters. The -m parameter to specify the number of coroutines running in parallel (defaults to 8). And the -W parameter to allow qemu-img to write to the target out of order rather than sequential. This improves performance as the writes do not have to wait for each other to complete. " ----- Mail original ----- De: "aderumier" < [ mailto:aderumier@xxxxxxxxx | aderumier@xxxxxxxxx ] > À: "dillaman" < [ mailto:dillaman@xxxxxxxxxx | dillaman@xxxxxxxxxx ] > Cc: "Mahesh Jambhulkar" < [ mailto:mahesh.jambhulkar@xxxxxxxxx | mahesh.jambhulkar@xxxxxxxxx ] >, "ceph-users" < [ mailto:ceph-users@xxxxxxxxxxxxxx | ceph-users@xxxxxxxxxxxxxx ] > Envoyé: Vendredi 21 Juillet 2017 10:51:21 Objet: Re: qemu-img convert vs rbd import performance Hi, they are an RFC here: "[RFC] qemu-img: make convert async" [ https://patchwork.kernel.org/patch/9552415/ | https://patchwork.kernel.org/patch/9552415/ ] maybe it could help ----- Mail original ----- De: "Jason Dillaman" < [ mailto:jdillama@xxxxxxxxxx | jdillama@xxxxxxxxxx ] > À: "Mahesh Jambhulkar" < [ mailto:mahesh.jambhulkar@xxxxxxxxx | mahesh.jambhulkar@xxxxxxxxx ] > Cc: "ceph-users" < [ mailto:ceph-users@xxxxxxxxxxxxxx | ceph-users@xxxxxxxxxxxxxx ] > Envoyé: Jeudi 20 Juillet 2017 15:20:32 Objet: Re: qemu-img convert vs rbd import performance Running a similar 20G import test within a single OSD VM-based cluster, I see the following: $ time qemu-img convert -p -O raw -f raw ~/image rbd:rbd/image (100.00/100%) real 3m20.722s user 0m18.859s sys 0m20.628s $ time rbd import ~/image Importing image: 100% complete...done. real 2m11.907s user 0m12.236s sys 0m20.971s Examining the IO patterns from qemu-img, I can see that it is effectively using synchronous IO (i.e. only a single write is in-flight at a time), whereas "rbd import" will send up to 10 (by default) IO requests concurrently. Therefore, the higher the latencies to your cluster, the worse qemu-img will perform as compared to "rbd import". On Thu, Jul 20, 2017 at 5:07 AM, Mahesh Jambhulkar < [ mailto: [ mailto:mahesh.jambhulkar@xxxxxxxxx | mahesh.jambhulkar@xxxxxxxxx ] | [ mailto:mahesh.jambhulkar@xxxxxxxxx | mahesh.jambhulkar@xxxxxxxxx ] ] > wrote: Adding rbd readahead disable after bytes = 0 did not help. [root@cephlarge mnt]# time qemu-img convert -p -O raw /mnt/data/workload_326e8a43-a90a-4fe9-8aab-6d33bcdf5a05/snapshot_9f0cee13-8200-4562-82ec-1fb9f234bcd8/vm_id_05e9534e-5c84-4487-9613-1e0e227e4c1a/vm_res_id_24291e4b-93d2-47ad-80a8-bf3c395319b9_vdb/66582225-6539-4e5e-9b7a-59aa16739df1 rbd:volumes/24291e4b-93d2-47ad-80a8-bf3c395319b9 (100.00/100%) real 4858m13.822s user 73m39.656s sys 32m11.891s It took 80 hours to complete. Also, its not feasible to test this with huge 465GB file every time. So I tested qemu-img convert with a 20GB file. Parameters Time taken -t writeback 38mins -t none 38 mins -S 4k 38 mins With client options mentions by Irek Fasikhov 40 mins The time taken is almost the same. On Thu, Jul 13, 2017 at 6:40 PM, Jason Dillaman < [ mailto: [ mailto:jdillama@xxxxxxxxxx | jdillama@xxxxxxxxxx ] | [ mailto:jdillama@xxxxxxxxxx | jdillama@xxxxxxxxxx ] ] > wrote: On Thu, Jul 13, 2017 at 8:57 AM, Irek Fasikhov < [ mailto: [ mailto:malmyzh@xxxxxxxxx | malmyzh@xxxxxxxxx ] | [ mailto:malmyzh@xxxxxxxxx | malmyzh@xxxxxxxxx ] ] > wrote: > rbd readahead disable after bytes = 0 There isn't any reading from an RBD image in this example -- plus readahead disables itself automatically after the first 50MBs of IO (i.e. after the OS should have had enough time to start its own readahead logic). -- Jason -- Regards, mahesh j -- Jason _______________________________________________ ceph-users mailing list [ mailto:ceph-users@xxxxxxxxxxxxxx | ceph-users@xxxxxxxxxxxxxx ] [ http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com | http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com ] -- Regards, mahesh j _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |