Bug #9217
wrong value of osd::op_w_latency
0%
Description
when i use fio to test rbd performance (4kb random write) , i found that osd::op_w_latency which i got from perf counter is much larger than the latency reported by fio, after i digged into it and found that, with XFS , after data is written to journal , osd sends a reply to client to imply that the write request has completed , it updates the osd::op_w_latency perf counter in ReplicatedPG::eval_repop function , when data is written to filestore, the osd::op_w_latency perf counter updates again, so op_w_latency perf counter records the apply latency , not the commit latency , is this the defined behavior or a bug?
Associated revisions
fix wrong value of op_w_latency perf counter
Fixes: #9217
Signed-off-by: xinxin shu <xinxin.shu@intel.com>
History
#1 Updated by Xinxin Shu over 9 years ago
- Category set to OSD
#2 Updated by Sage Weil over 9 years ago
- Source changed from other to Community (dev)
I think it's a bug:
osd_plb.add_time_avg(l_osd_op_w_rlat, "op_w_rlat"); // client write readable/applied latency osd_plb.add_time_avg(l_osd_op_w_lat, "op_w_latency"); // client write latency
#3 Updated by Xinxin Shu over 9 years ago
- Assignee set to Xinxin Shu
#4 Updated by Xinxin Shu over 9 years ago
- Status changed from New to Fix Under Review
#5 Updated by Xinxin Shu over 9 years ago
#6 Updated by Sage Weil over 9 years ago
- Status changed from Fix Under Review to Resolved