Closed givemealung closed 5 months ago
Fio issued 7889321 write commands whereas blktrace is provided data on 1549489 operations. I think it's unreasonable to expect the latencies to agree.
Fio issued 7889321 write commands whereas blktrace is provided data on 1549489 operations. I think it's unreasonable to expect the latencies to agree.
Sorry for omitting the reply.
The difference of write commands count is mainly caused by sampling from blktrace is merely 1minutes, whereas fio run for 5 minutes plus 3-minute ramp time. However, average IO's completion time(Q2C, 33689ns) is even less than 1th percentile of clat(34048ns), it should tell on the problem.
Description of the bug: Hi,
In my case, I tried to figure where the latency comes from using blktrace. Somehow I found Q2D(should be the stat from when IO submitted to block layer and completed) from blktrace is kindly different from clat from fio. Tested with both async I/O engine and sync I/O engine, and it's only reproduced with sync I/O engine. In sync I/O case, latency data from fio statistics and blktrace is:
Environment: CentOS Linux release 7.9.2009 (AltArch) 4.18.0-193.28.1.el7.aarch64
fio version: 3.35
Reproduction steps Test with below config file.
Other I found the explantion of ambigious latency in sync I/O case, possibly relative to the issue?#1110