Steffen Weiberle
2006-Oct-23 16:46 UTC
[zfs-discuss] ?: cyclical kernel/system processing (approx every 4 minutes)
Customer benchmarked an X4600 using UFS on top of VxVM a while back and got consistent performance under heavy load. Now they have put the system into system test, but in the process moved from UFS/VxVM to ZFS. This is 6/06 The are running at approximately 40% idle most of the time, with 10+% system time, and 40+% user time. However, every 4 minutes or so, the system time goes up to 60%, leaving no idle time. This is under steady load, and the customer is concerned that it will not be able to ride out peaks higher than this. Is there any known processing that would be done every 4 minutes last 10 to 15 seconds that might be introduced by ZFS? I don''t have sufficient data yet to verify that the I/O is going to ZFS volumes, but I do see a pair of disk with extended device statistics device r/s w/s kr/s kw/s wait actv svc_t %w %b sd21 0.1 129.3 8.5 9396.0 0.0 0.9 6.7 0 7 sd22 0.1 138.7 8.5 9872.0 0.0 0.9 6.7 0 7 going to: extended device statistics device r/s w/s kr/s kw/s wait actv svc_t %w %b sd21 0.0 59.0 0.0 1383.4 0.0 0.0 0.8 0 1 sd22 0.0 61.5 0.0 1455.2 0.0 0.1 0.9 0 1 or to: extended device statistics device r/s w/s kr/s kw/s wait actv svc_t %w %b sd21 0.0 180.2 2.1 16513.5 0.0 2.5 14.0 0 11 sd22 0.0 195.2 2.1 17485.1 0.0 3.0 15.5 0 11 Don''t know yet which is ''normal'' and which is not. Thanks for any insights! Steffen
Matthew Ahrens
2006-Oct-24 01:40 UTC
[zfs-discuss] ?: cyclical kernel/system processing (approx every 4 minutes)
Steffen Weiberle wrote:> Is there any known processing that would be done every 4 minutes last 10 > to 15 seconds that might be introduced by ZFS?No. But you can use tools like dtrace or lockstat -kgIW to determine where that additional CPU time is being spent, and then see if ZFS is to blame... --matt