lkml.org 
[lkml]   [2007]   [May]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [ckrm-tech] [RFC] [PATCH 0/3] Add group fairness to CFS
Srivatsa Vaddagiri wrote:
>> Ingo/Peter, any thoughts here? CFS and smpnice probably is "broken"
>> with respect to such example as above albeit for nice-based tasks.

On Sat, May 26, 2007 at 10:17:42AM +1000, Peter Williams wrote:
> See above. I think that faced with cpu affinity use by the system
> administrator that smpnice will tend towards a task to cpu allocation
> that is (close to) the best that can be achieved without violating the
> cpu affinity assignments. (It may take a little longer than normal but
> it should get there eventually.)
> You have to assume that the system administrator knows what (s)he's
> doing and is willing to accept the impact of their policy decision on
> the overall system performance.
> Having said that, if it was deemed necessary you could probably increase
> the speed at which the load balancer converged on a good result in the
> face of cpu affinity by keeping a "pinned weighted load" value for each
> run queue and using that to modify find_busiest_group() and
> find_busiest_queue() to be a bit smarter. But I'm not sure that it
> would be worth the added complexity.

Just in case anyone was looking for algorithms...

Lag should be considered in lieu of load because lag is what the
scheduler is trying to minimize; load is not directly relevant, but
appears to have some sort of relationship. Also, instead of pinned,
unpinned should be considered. It's unpinned that load balancing can
actually migrate. Using the signed minimax pseudonorm (i.e. the highest
signed lag, where positive is higher than all negative regardless of
magnitude) on unpinned lags yields a rather natural load balancing
algorithm consisting of migrating from highest to lowest signed lag,
with progressively longer periods for periodic balancing across
progressively higher levels of hierarchy in sched_domains etc. as usual.
Basically skip over pinned tasks as far as lag goes.

The trick with all that comes when tasks are pinned within a set of
cpus (especially crossing sched_domains) instead of to a single cpu.
There one can just consider a cpu to enter a periodic load balance
cycle, and then consider pushing and pulling, perhaps what could be
called the "exchange lags" for the pair of cpus. That would be the
minimax lag pseudonorms for the tasks migratable to both cpus of the
pair. That makes the notion of moving things from highest to lowest
lag (where load is now considered) unambiguous apart from whether all
this converges, but not when to actually try to load balance vs. when
not to, or when it's urgent vs. when it should be done periodically.

To clarify that, an O(cpus**2) notion appears to be necessary, namely
the largest exchange lag differential between any pair of cpus. There
is also the open question of whether moving tasks between cpus with the
highest exchange lag differential will actually reduce it or whether it
runs the risk of increasing it by creating a larger exchange lag
differential between different pairs of cpus. A similar open question
is raised by localizing balancing decisions to sched_domains. What
remains clear is that any such movement reduces the worst-case lag in
the whole system. Because of that, the worst-case lag in the whole
system monotonically decreases as balancing decisions are made, and
that much is subject to an infinite descent argument. Unfortunately,
determining the largest exchange lag differential appears to be more
complex than merely finding the highest and lowest lags. Bipartite
forms of the problem also arise from sched_domains.
I doubt anyone's really paying any sort of attention, so I'll not
really bother working out much more in the way of details with respect
to load balancing. It may be that there are better ways to communicate
algorithmic notions than prose descriptions. However, it's doubtful I'll
produce anything in a timely enough fashion to attract or hold interest.

The smpnice affair is better phrased in terms of task weighting. It's
simple to honor nice in such an arrangement. First unravel the
grouping hierarchy, then weight by nice. This looks like

task nice hier1 hier2 ... hierN
t_1 w_n1 w_h11 w_h21 ... w_hN1
t_2 w_n2 w_h12 w_h22 ... w_hN2
...

For the example of nice 0 vs. nice 10 as distinct users with 10%
steppings between nice levels, one would have

task nice hier1
t_1 1 1
t_2 0.3855 1

w_1, the weight of t_1, would be
(w_h11*w_n1/(w_h11*w_n1 + w_h12*w_n2))
= (1*1/(1 + 1*0.3855..))
= 0.7217..
w_2, the weight of t_2, would be
(w_h12*w_n2/(w_h11*w_n1 + w_h12*w_n2))
= (1*0.3855../(1 + 1*0.3855..))
= 0.27826..
This just so happens to work out to being the same as if t_1 and t_2
had their respective nice numbers without the scheduler grouping, which
is basically what everyone wants to happen.

It's more obvious how to extend it to more tasks than levels of
hierarchy. An example of that follows:

task nice hier1 hier2 ... hierN
t_1 0.3 0.6 * ... *
t_2 0.7 0.4 * ... *

hier2 through hierN are ignorable since t_1 and t_2 are both the only
members at those levels of hierarchy. We then get something just like
the above example, w_1 = 0.3*0.6/(0.3*0.6+0.7*0.4) = 0.3913.. and
w2 = 0.7*0.4/(0.3*0.6+0.7*0.4) = 0.6087..

It's more interesting with enough tasks to have more meaningful levels
of hierarchy.

task nice hier1 hier2
t_1 0.7 0.6 0.6
t_2 0.3 0.6 0.4
t_3 0.7 0.4 0.6
t_4 0.3 0.4 0.4

where t_1 and t_2 share a hier1 grouping and t_3 and t_4 also share
a hier1 grouping, but the hier1 grouping for t_1 and t_2 is distinct
from the hier1 grouping for t_3 and t_4. All hier2 groupings are
distinct. So t_1 would have pre-nice weight 0.6*0.6, t_2 0.6*0.4,
t_3 0.6*0.4, and t_4 0.4*0.4 (the numbers were chosen so denominators
conveniently collapse to 1). Now that the hierarchy is flattened,
nice numbers can be factored in for t_1's final weight being
0.7*0.36/(0.7*0.36+0.3*0.24+0.7*0.24+0.3*0.16) = 0.252/0.54 = 0.467..
and the others being 0.133.. (t_2), 0.311.. (t_3), and 0.0889.. (t_4).

In such a manner nice numbers obey the principle of least surprise.


-- wli
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2007-05-26 17:47    [from the cache]
©2003-2011 Jasper Spaans