Re: [patch v3 0/8] sched: use runnable avg in load balance

[Mike Galbraith]

On Tue, 2013-04-02 at 15:23 +0800, Michael Wang wrote: 
> On 04/02/2013 11:23 AM, Alex Shi wrote:
> [snip]
> > 
> > [patch v3 1/8] Revert "sched: Introduce temporary FAIR_GROUP_SCHED
> > [patch v3 2/8] sched: set initial value of runnable avg for new
> > [patch v3 3/8] sched: only count runnable avg on cfs_rq's nr_running
> > [patch v3 4/8] sched: update cpu load after task_tick.
> > [patch v3 5/8] sched: compute runnable load avg in cpu_load and
> > [patch v3 6/8] sched: consider runnable load average in move_tasks
> > [patch v3 7/8] sched: consider runnable load average in
> > [patch v3 8/8] sched: use instant load for burst wake up
> 
> I've tested the patch set on 12 cpu X86 box with 3.9.0-rc2, and pgbench
> show regression on high-end this time.
> 
> | db_size | clients |  tps  |   |  tps  |
> +---------+---------+-------+   +-------+
> | 22 MB   |       1 | 10662 |   | 10446 |
> | 22 MB   |       2 | 21483 |   | 20887 |
> | 22 MB   |       4 | 42046 |   | 41266 |
> | 22 MB   |       8 | 55807 |   | 51987 |
> | 22 MB   |      12 | 50768 |   | 50974 |
> | 22 MB   |      16 | 49880 |   | 49510 |
> | 22 MB   |      24 | 45904 |   | 42398 |
> | 22 MB   |      32 | 43420 |   | 40995 |
> | 7484 MB |       1 |  7965 |   |  7376 |
> | 7484 MB |       2 | 19354 |   | 19149 |
> | 7484 MB |       4 | 37552 |   | 37458 |
> | 7484 MB |       8 | 48655 |   | 46618 |
> | 7484 MB |      12 | 45778 |   | 45756 |
> | 7484 MB |      16 | 45659 |   | 44911 |
> | 7484 MB |      24 | 42192 |   | 37185 |	-11.87%
> | 7484 MB |      32 | 36385 |   | 34447 |
> | 15 GB   |       1 |  7677 |   |  7359 |
> | 15 GB   |       2 | 19227 |   | 19049 |
> | 15 GB   |       4 | 37335 |   | 36947 |
> | 15 GB   |       8 | 48130 |   | 46898 |
> | 15 GB   |      12 | 45393 |   | 43986 |
> | 15 GB   |      16 | 45110 |   | 45719 |
> | 15 GB   |      24 | 41415 |   | 36813 |	-11.11%
> | 15 GB   |      32 | 35988 |   | 34025 |
> 
> The reason may caused by wake_affine()'s higher overhead, and pgbench is
> really sensitive to this stuff...

For grins, you could try running the whole thing SCHED_BATCH.  (/me sees
singing/dancing red herring whenever wake_affine() and pgbench appear in
the same sentence;)

-Mike

--
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/