It seems that there's some bottleneck is blocking the I/O, when the bottleneck is reached, I/O is blocked and curve goes down, when it is released, I/O resumes and the curve gose up. -----邮件原件----- 发件人: ceph-users [mailto:ceph-users-bounces@xxxxxxxxxxxxxx] 代表 Patrick Dinnen 发送时间: 2017年5月12日 3:47 收件人: ceph-users@xxxxxxxxxxxxxx 主题: Odd cyclical cluster performance Seeing some odd behaviour while testing using rados bench. This is on a pre-split pool, two node cluster with 12 OSDs total. ceph osd pool create newerpoolofhopes 2048 2048 replicated "" replicated_ruleset 500000000 rados -p newerpoolofhopes bench -t 32 -b 20000 30000000 write --no-cleanup Using Prometheus/Grafana to watch what's going on, we see oddly regular peaks and dips in writer performance. The frequency changes gradually but it's on the order of hours (not the seconds that might seem easier to explain by system phenomena). It starts off at roughly one cycle per hour and we've seen it for multiple days of constant bench running with nothing else happening on the cluster. A bunch of graphs showing the pattern: https://ibb.co/djXUVk https://ibb.co/gMNk35 https://ibb.co/iKViqk https://ibb.co/jOXJO5 https://ibb.co/isUMbQ sdg and sdi are SSD journal disks. The activity on the OSDs and SSDs seems anti-correlated. SSDs peak in activity as OSDs reach the bottom of the trough. Then the reverse. Repeat. Does anyone have any suggestions as to what could possibly be causing a regular pattern like this at such a low frequency? Thanks, Patrick Dinnen _______________________________________________ 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
 |