Re: [RFC PATCH 00/14] sched: entity load-tracking re-work

From: Vincent Guittot
Date: Mon Mar 12 2012 - 06:57:37 EST


Hi Paul,

On 2 February 2012 02:38, Paul Turner <pjt@xxxxxxxxxx> wrote:
> Hi all,
>
> The attached series is an RFC on implementing load tracking at the entity
> instead of cfs_rq level. This results in a bottom-up load-computation in which
> entities contribute to their parents load, as opposed to the current top-down
> where the parent averages its children.  In particular this allows us to
> correctly migrate load with their accompanying entities and provides the
> necessary inputs for intelligent load-balancing and power-management.
>
> It was previously well tested and stable, but that was on v3.1-; there's been
> some fairly extensive changes in the wake-up path since so apologies if anything
> was broken in the rebase.Note also, since this is also an RFC on the approach I
> have not yet de-linted the various CONFIG combinations for introduced compiler
> errors.
>
> Background:
> ----------
> We currently track the load average at the parenting cfs_rq level. We divide
> execution into a series of consecutive "windows, w_i".  Within each we track:
>  \Sum load_i * t_i where \Sum t_i = w and each load_i a disjoint load level.
>
> The load average is then \Sum w_j / 2^n.
>
> This this works reasonably well but there's a few problems
> 1) Since decay occurs at boundary window boundary there are 'skews':
> At a window boundary the 'foreground' time has a bias against the
> time immediately preceding it (as a result of the folding division)
> e.g.  __xx_|_yyyy___ vs __xx_yyyy_|___  (where '|' is a window boundary).
>
> The accounting here is 2x + 4y/2 or 2x + 4y, depending on which side of the
> window your load lands on.
>
> 2) Everything within a window 'w' is accounted equally, we only fold at
> the boundaries.  This actually means you can't set 'w' large enough to
> really have meaningful coverage of the sched period without throwing
> decay out the window.  But then with w/2 << sched_period (currently
> w/2=5ms) the average ends up having fairly drastic swings even with
> stable loads.
>
> (Note:  Before we even looked at migrating to per-entity tracking we evaluating
> foreground window into the considered average until it was "complete", this
> represented a measurable improvement in stability under predictable load.)
>
> 3) Since the load sum average is per-cfs_rq and not per-entity when a task
> entity migrates we lose its contribution to load-sum and effectively double
> count it while it former sum decays.
>
> New approach:
> -------------
> Instead of tracking load on a per-cfs_rq basis we do it on a per-sched_entity
> basis.  The load sum for a cfs_rq is then just the sum of its childrens' load
> averages.  The obvious immediately nice property is that when we migrate thread
> A from cpu1-->cpu2, we carry its load with it; leaving the global load sum
> unmodified.
>
> The 'windows' above are replaced with more fine-grained tracking, that is (for
> an entity j):
>
> runnable_j = u_i * y^i , load_avg_j = runnable_j * weight_j [*]
>
> Where: u_i is the usage in the last i`th ~ms and y is chosen such that
> y^~sched_period = 1/2 (we currently choose k=32).This means that load tracked 1
> sched_period ago contributes about ~50% as current execution.

We have a sched_period of 30ms for a 16 cores system but it's only 12
ms on a dual core. Do you think that it's important to keep this rule
(y^~sched_period = 1/2) whatever the number of core is ? The
sched_period also increases with a large number of running thread

>
> Now, the real challenge in tracking at an entity basis is handling blocked
> entities.  Obviously runnable entities will be updated naturally as we iterate
> through them but there could be O(many) blocked entities so we can't afford to
> iterate through them and update their tracking.
>

I have run sysbench of my 32bits ARM platform. The results are available below:

config 1 : v3.3-rc3 and CONFIG_FAIR_GROUP_SCHED not set
config 2 : v3.3-rc3 and CONFIG_FAIR_GROUP_SCHED set
config 3 : v3.3-rc3 with your patches and CONFIG_FAIR_GROUP_SCHED not set
config 4 : v3.3-rc3 with your patches and CONFIG_FAIR_GROUP_SCHED set

sysbench --test=cpu --num-threads=12 --max-time=10 run
total number of events
test1 test2 test3
config1 336 336 338
config2 336 337 338
config3 336 338 336
config4 336 338 337

sysbench --test=threads --thread-locks=9 --num-threads=12 --max-time=10 run
total number of events
test1 test2 test3
config1 5218 5228 5203
config2 5204 5217 5251
config3 5221 5219 5217
config4 4859 4796 4836


Whereas we have no difference for the cpu test (all the threads on in
the run queue waiting for a core), we can see a decrease for the
threads test. I'm wondering if it's linked to the fact that I have a
32bits machine doing 64bits division. Have you seen such kind of
difference on a 64bits system ?

Regards,
Vincent

> That our decay for a unit period is exponential introduces a particularly nice
> property here:
> We can separate the contributed load on a cfs_rq into blocked and runnable.
> Runnable load is just the sum of all load_avg_j above, maintained by the
> entities themselves as they run and self update (when they update their
> load_avg they update the cumulative sum also).
>
> Blocked load then looks like:
>  load_avg_j = weight_k * \Sum u[k]_n * y^n
>
> To account an entirely idle period we then only need to multiply by y.
>
> This ends up being orders of magnitude more accurate than the current
> tracking schema, even with the old shares_window massively dilated to
> better capture a stable load-state.  It's also obviously stable under
> migration.
>
> As referenced above this also allows us to potentially improve decisions within
> the load-balancer, for both distribution and power-management.
>
> Exmaple: consider 1x80% task  and 2x40% tasks on a 2-core machine.  It's
> currently a bit of a gamble as to whether you get an {AB, B} or {A,
> BB} split since they have equal weight (assume 1024).  With per-task
> tracking we can actually consider them at their contributed weight and
> see a stable ~{800,{400, 400}} load-split.  Likewise within balance_tasks we can
> consider the load migrated to be that actually contributed.
>
> Thanks,
>
> - Paul
>
>
> ---
>
> Ben Segall (1):
>      sched: maintain per-rq runnable averages
>
> Paul Turner (13):
>      sched: track the runnable average on a per-task entitiy basis
>      sched: aggregate load contributed by task entities on parenting cfs_rq
>      sched: maintain the load contribution of blocked entities
>      sched: account for blocked load waking back up
>      sched: aggregate total task_group load
>      sched: compute load contribution by a group entity
>      sched: normalize tg load contributions against runnable time
>      sched: maintain runnable averages across throttled periods
>      sched: replace update_shares weight distribution with per-entity computation
>      sched: refactor update_shares_cpu() -> update_blocked_avgs()
>      sched: update_cfs_shares at period edge
>      sched: make __update_entity_runnable_avg() fast
>      sched: implement usage tracking
>
>
>  include/linux/sched.h |   11 +
>  kernel/sched/core.c   |    3
>  kernel/sched/debug.c  |   37 ++-
>  kernel/sched/fair.c   |  714 ++++++++++++++++++++++++++++++++++++++-----------
>  kernel/sched/sched.h  |   29 +-
>  5 files changed, 611 insertions(+), 183 deletions(-)
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/