f876e446: +1.2% fileio.requests_per_sec
Hi Dave,
We noticed the below changes on commit f876e44603ad091c840a5fae5b0753bbb421c037 ("xfs: always do log forces via the workqueue"):
Basically there are +1.2% increased throughput with +29.5% increased context switches.
test box: NHM
test case: fileio rndwr-sync
v3.14-rc1 f876e44603ad091c840a5fae5
--------------- -------------------------
192 ~ 0% +1.2% 194 ~ 0% TOTAL fileio.requests_per_sec
2 ~25% +47.4% 4 ~ 0% TOTAL time.percent_of_cpu_this_job_got
1907 ~16% +29.5% 2471 ~ 0% TOTAL vmstat.system.cs
331131 ~ 0% +10.9% 367298 ~ 0% TOTAL time.voluntary_context_switches
3255 ~ 0% +2330.6% 79126 ~ 0% TOTAL time.involuntary_context_switches
3703982 ~ 0% +1.2% 3747974 ~ 0% TOTAL time.file_system_outputs
35.04 ~20% +9.8% 38.48 ~ 0% TOTAL iostat.sda.wrqm/stime.voluntary_context_switches
380000 ++----------------------------------------------------------------+ 360000 O+ O O O O O O O O O O O O O O O O O O O O O O O
| |
340000 *+.*..*..*.*..*..*..*..*..*..*..* *..*..*..*..*..*..*.* |
320000 ++ : : |
| : : |
300000 ++ : : |
280000 ++ : : |
260000 ++ : : |
| : : |
240000 ++ : : |
220000 ++ :: |
| : |
200000 ++ * |
180000 ++----------------------------------------------------------------+
time.involuntary_context_switches
80000 ++------------------------O--O--O--O-----O--O--O--O-O--O--O--O--O--O
O O O O O O O O O O |
70000 ++ |
60000 ++ |
| |
50000 ++ |
| |
40000 ++ |
| |
30000 ++ |
20000 ++ |
| |
10000 ++ |
*..*..*..*..*..*.*..*..*..*..*..*..*..*..*..*..*..*.*..*..* |
0 ++-----------------------------------------------------------------+
Thanks,
Fengguang
Quoted 7 lines · Fengguang Wu
Hi Dave,
We noticed the below changes on commit
f876e44603ad091c840a5fae5b0753bbb421c037("xfs: always do log forces via the workqueue"):Basically there are +1.2% increased throughput with +29.5% increased context switches.
Yes, the increase in context switches is expected. Every sync IO causes a log force if it changed inode metadata, and the above commit, as it describes, moves them to a work queue. Hence there's more context switches involved in a log force that has to do work.
Cheers,
Dave.
Signature · 2 lines
Dave Chinner [email protected]