hackbench.c interactivity results: vanilla versus SD/RSDL
Quoted 4 lines · Ingo Molnar
Quoted 2 lines · Con Kolivas
I'm cautiously optimistic that we're at the thin edge of the bugfix wedge now.
[...]
Quoted 6 lines
and the numbers he posted:
http://marc.info/?l=linux-kernel&m=117448900626028&w=2
his test conclusion was that under CPU load, RSDL (SD) generally does not hold up to mainline's interactivity.
Today i've done some hackbench.c (which is similar to VolanoMark) interactivity testing on SD-latest (v0.37), doing "hackbench 10", "hackbench 50" and "hackbench 100" runs, comparing it to vanilla+Mike's-task-promotion-patch.
Performance
-----------
performance was largely comparable, within noise: the vanilla scheduler was slightly (~5%) better at "hackbench 10", SD and vanilla was within noise on "hackbench 50" and "hackbench 100" runs. (I think vanilla's better results might be due to the longer timeslices vanilla can give out - SD has to operate via short timeslices, by design.)
Shell interactivity
-------------------
The vanilla scheduler kept the shell completely usable during all tests:
'ls' output was instantaneous and characters could be typed without delay.
The SD/RSDL scheduler kept the shell barely usable during the hackbench 10 test, and it was completely unusable during the hackbench 50 and 100 tests. A simple 'ls' took 2-3 seconds to complete (up to 6 seconds at times) and the shell printed characters in a very laggy way. 'vi' was totally unusable, etc. etc.
[ i've also re-tested this with RSDL 0.34, and there interactivity got
better although it still didnt match that of the vanilla scheduler's interactivity. So this is a definitive SD regression. ]
[ A quick guess: could SD's substandard interactivity in this test be
due to the SMP migration logic inconsistencies Mike noticed? This is an SMP system and the hackbench workload is very scheduling intense and tasks are frequently queued from one CPU to another. ]
Conclusion
----------
i consider this a must-fix item as SD badly misbehaves under this workload.
Test environment details
------------------------
the test system is a 2GHz Athlon64 dual-core box. All tests were running on default nice 0 levels. All tests were done 10 times on a totally idle test-system. Mike's patch is the one that improves vanilla scheduler's anti-starvation code:
http://marc.info/?l=linux-kernel&m=117507110922009&w=2
( Mike's patch does not make a measurable performance difference in
hackbench, nor does it make a visible interactivity difference for this workload, but since i think Mike's patch improves the vanilla scheduler i included it in the test for completeness, so that both variants of interactivity code are at the 'bleeding edge'. )
Ingo
Quoted 4 lines · Ingo Molnar
[ A quick guess: could SD's substandard interactivity in this test be
due to the SMP migration logic inconsistencies Mike noticed? This is an SMP system and the hackbench workload is very scheduling intense and tasks are frequently queued from one CPU to another. ]
I assume you put it on and endless loop since hackbench 10 runs for .5 second on my machine. Doubtful it's an SMP issue. update_if_moved should maintain cross cpu scheduling decisions. The same slowdown would happen on UP and is almost certainly due to the fact that hackbench 10 induces a load of _160_ on the machine.
Signature · 1 line
-ck