Re: RadosGW (?) slow parallel upload

Mihály Árva-Tóth <mihaly.arva-toth@xxxxxxxxxxxxxxxxxxxxxx> · Mon, 27 Jan 2014 08:30:24 +0100

Hello,

Have you got any idea? HDDs are fast, filesystem is XFS, mounted with nobarrier,noatime... I think everything is right for the great performance.

Thank you,

Mihaly

2014-01-14 Mihály Árva-Tóth <mihaly.arva-toth@xxxxxxxxxxxxxxxxxxxxxx>

Hello,

I have got 3 servers, with 3 HDD-OSD / server (4 TB WD RE). I'm using radosgw primary. Every .rgw.* pool has 3 replica. Every server is rados gateway with apache2+fastcgi (ceph patched version). Servers type: SuperMicro SSG-6047R-E1R36L.

I've got 10 client machines which uploads objects parallel. Size of each object is variable, but between 50-500 KiB typical. haproxy is installed to every client machine, and upload script connects to localhost (haproxy) which round-robin distribute requests between three rgw servers.

Typical avg upload times:

Best: 260 ms
Avg: 775 ms
Worst: 1826 ms

Not rare 10 secs upload time with 74 KiB (or same low) object size.

Sample log entry from radosgw:

2 req 157744:0.000124::PUT /snd1_rec/20140111_human.mp4::initializing
2 req 157744:0.000214:s3:PUT /snd1_rec/20140111_human.mp4::getting op
2 req 157744:0.000230:s3:PUT /snd1_rec/20140111_human.mp4:put_obj:authorizing

2 req 157744:0.000518:s3:PUT /snd1_rec/20140111_human.mp4:put_obj:reading permissions
2 req 157744:0.000662:s3:PUT /snd1_rec/20140111_human.mp4:put_obj:init op
2 req 157744:0.000679:s3:PUT /snd1_rec/20140111_human.mp4:put_obj:verifying op mask

2 req 157744:0.000689:s3:PUT /snd1_rec/20140111_human.mp4:put_obj:verifying op permissions
2 req 157744:0.000699:s3:PUT /snd1_rec/20140111_human.mp4:put_obj:verifying op params
2 req 157744:0.000727:s3:PUT /snd1_rec/20140111_human.mp4:put_obj:executing

2 req 157744:11.025770:s3:PUT /snd1_rec/20140111_human.mp4:put_obj:http status=200

11 seconds between the last two logline. I think not radosgw is the slowest part, but I don't known what is. I'm monitoring lot of resources of servers (vmstat, disk throughput, disk io, cpu...) and there is nothing to run out. 

I don't know how important but there is not rare container with 4 million of objects.

Ubuntu 12.04-x64 with 0.72.1-1precise.

Thank you,

Mihaly

_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com