* Michael Gerdau <[email protected]> wrote:
> There are 3 scenarios:
> j1 - all 3 tasks run sequentially
> /proc/sys/kernel/sched_granularity_ns=4000000
> /proc/sys/kernel/rr_interval=16
> j3 - all 3 tasks run in parallel
> /proc/sys/kernel/sched_granularity_ns=4000000
> /proc/sys/kernel/rr_interval=16
> j3big - all 3 tasks run in parallel with timeslice extended
> by 2 magnitudes (not run for mainline)
> /proc/sys/kernel/sched_granularity_ns=400000000
> /proc/sys/kernel/rr_interval=400
>
> All 3 tasks are run while the system does nothing else except for
> the "normal" (KDE) daemons. The system had not been used for
> interactive work during the tests.
>
> I'm giving user time as provided by the "time" cmd followed by wallclock time
> (all values in seconds).
>
> LTMM
> j1 j3 j3big
> 2.6.21-cfs-v6 5655.07/ 5682 5437.84/ 5531 5434.04/ 8072
> LTMB
> 2.6.21-cfs-v6 7729.81/ 7755 7470.10/10244 7449.16/10186
> LTBM
> 2.6.21-cfs-v6 7720.70/ 7746 7567.09/10362 7464.17/10335
> LTMM+LTMB+LTBM
> 2.6.21-cfs-v6 21105.58/21183 20475.03/26137 20347.37/28593
> User time apparently is subject to some variance. I'm particularly
> surprised by the wallclock time of scenario j1 and j3 for case LTMM
> with 2.6.21-cfs-v6. I'm not sure what to make of this, i.e. whether I
> had happening something else on my machine during j1 of LTMM -- that's
> always been the first test I ran and it might be that there were still
> some other jobs running after the initial boot.
thanks for the testing!
regarding the fairness of the different schedulers, please note the
different runtimes for each component of the workload:
LTMM: 5655.07/ 5682
LTMB: 7729.81/ 7755
LTBM: 7720.70/ 7746
this means that a fair scheduler would _not_ be the one that finishes
them first on wall-clock time (!). A fair scheduler would run each of
them at 33% capacity until the fastest one (LTMM) reaches ~5650 seconds
runtime and finishes, and _then_ the remaining ~2050 seconds of runtime
would be done at 50%/50% capacity between the remaining two jobs. I.e.
the fair wall-clock results should be around:
LTMM: ~8500 seconds
LTMB: ~10600 seconds
LTBM: ~10600 seconds
(but the IO portion of the workloads and other scheduling effects could
easily shift these numbers by a few minutes.)
regarding the results: it seems the wallclock portion of LTMM/j3 is too
small - even though the 3 tasks ran in parallel, in the CFS test LTMM
finished just as fast as if it were running alone, right? That does not
seem to be logical and indeed suggests some sort of testing artifact.
That makes it hard to judge which scheduler achieved the above 'ideal
fair distribution' of the workloads better - for some of the results it
was SD, for some it was CFS - but the missing LTMM/j3 number makes it
hard to decide it conclusively. They are certainly both close enough and
the noise of the results seems quite high.
Ingo
> regarding the fairness of the different schedulers, please note the
> different runtimes for each component of the workload:
>
> LTMM: 5655.07/ 5682
> LTMB: 7729.81/ 7755
> LTBM: 7720.70/ 7746
>
> this means that a fair scheduler would _not_ be the one that finishes
> them first on wall-clock time (!). A fair scheduler would run each of
> them at 33% capacity until the fastest one (LTMM) reaches ~5650 seconds
> runtime and finishes, and _then_ the remaining ~2050 seconds of runtime
> would be done at 50%/50% capacity between the remaining two jobs. I.e.
> the fair wall-clock results should be around:
>
> LTMM: ~8500 seconds
> LTMB: ~10600 seconds
> LTBM: ~10600 seconds
>
> (but the IO portion of the workloads and other scheduling effects could
> easily shift these numbers by a few minutes.)
Correct. I will try to cook up some statistics for the next round
of results (been redoing cfs-v6, done cfs-v7 and then redone it
because of strange results and hopefully will do cfs-v8 tonight
as well as sd048 during the next days -- not easy to keep up with
you guys rolling out new versions ;-)
> regarding the results: it seems the wallclock portion of LTMM/j3 is too
> small - even though the 3 tasks ran in parallel, in the CFS test LTMM
> finished just as fast as if it were running alone, right?
Right. I'm really very surprised but this remained while redoing both
LTMM/j1 (twice!) and LTMM/j3 with cfs-v6.
I don't really understand that and now think it shows some hidden
"fairness deficiency" that hopefully is corrected in cfs-v8. At least
cfs-v7 did behave differently.
> That does not
> seem to be logical and indeed suggests some sort of testing artifact.
Since it doesn't appear with any other scheduler tested I'd not rule
out a feature of the code.
> That makes it hard to judge which scheduler achieved the above 'ideal
> fair distribution' of the workloads better - for some of the results it
> was SD, for some it was CFS - but the missing LTMM/j3 number makes it
> hard to decide it conclusively. They are certainly both close enough and
> the noise of the results seems quite high.
Oh, I'm confident both are excellent scheduler. On the other hand
I find mainline isn't that bad either, at least for this type of
load.
Anyway, as I wrote above I'm continuing my tests and am collecting
data for the various scheduler. Presumably over the weekend I'll
mail out the next round.
Best,
Michael
--
Technosis GmbH, Gesch?ftsf?hrer: Michael Gerdau, Tobias Dittmar
Sitz Hamburg; HRB 89145 Amtsgericht Hamburg
Vote against SPAM - see http://www.politik-digital.de/spam/
Michael Gerdau email: [email protected]
GPG-keys available on request or at public keyserver