sched: Increase SCHED_LOAD_SCALE resolution

Introduce SCHED_LOAD_RESOLUTION, which scales is added to
SCHED_LOAD_SHIFT and increases the resolution of
SCHED_LOAD_SCALE. This patch sets the value of
SCHED_LOAD_RESOLUTION to 10, scaling up the weights for all
sched entities by a factor of 1024. With this extra resolution,
we can handle deeper cgroup hiearchies and the scheduler can do
better shares distribution and load load balancing on larger
systems (especially for low weight task groups).

This does not change the existing user interface, the scaled
weights are only used internally. We do not modify
prio_to_weight values or inverses, but use the original weights
when calculating the inverse which is used to scale execution
time delta in calc_delta_mine(). This ensures we do not lose
accuracy when accounting time to the sched entities. Thanks to
Nikunj Dadhania for fixing an bug in c_d_m() that broken fairness.

Below is some analysis of the performance costs/improvements of
this patch.

1. Micro-arch performance costs:

Experiment was to run Ingo's pipe_test_100k 200 times with the
task pinned to one cpu. I measured instruction, cycles and
stalled-cycles for the runs. See:

   http://thread.gmane.org/gmane.linux.kernel/1129232/focus=1129389

for more info.

-tip (baseline):

 Performance counter stats for '/root/load-scale/pipe-test-100k' (200 runs):

       964,991,769 instructions             #    0.82  insns per cycle
                                            #    0.33  stalled cycles per insn
                                            #    ( +-  0.05% )
     1,171,186,635 cycles                   #    0.000 GHz                      ( +-  0.08% )
       306,373,664 stalled-cycles-backend   #   26.16% backend  cycles idle     ( +-  0.28% )
       314,933,621 stalled-cycles-frontend  #   26.89% frontend cycles idle     ( +-  0.34% )

        1.122405684  seconds time elapsed  ( +-  0.05% )

-tip+patches:

 Performance counter stats for './load-scale/pipe-test-100k' (200 runs):

       963,624,821 instructions             #    0.82  insns per cycle
                                            #    0.33  stalled cycles per insn
                                            #    ( +-  0.04% )
     1,175,215,649 cycles                   #    0.000 GHz                      ( +-  0.08% )
       315,321,126 stalled-cycles-backend   #   26.83% backend  cycles idle     ( +-  0.28% )
       316,835,873 stalled-cycles-frontend  #   26.96% frontend cycles idle     ( +-  0.29% )

        1.122238659  seconds time elapsed  ( +-  0.06% )

With this patch, instructions decrease by ~0.10% and cycles
increase by 0.27%. This doesn't look statistically significant.
The number of stalled cycles in the backend increased from
26.16% to 26.83%. This can be attributed to the shifts we do in
c_d_m() and other places. The fraction of stalled cycles in the
frontend remains about the same, at 26.96% compared to 26.89% in -tip.

2. Balancing low-weight task groups

Test setup: run 50 tasks with random sleep/busy times (biased
around 100ms) in a low weight container (with cpu.shares = 2).
Measure %idle as reported by mpstat over a 10s window.

-tip (baseline):

06:47:48 PM  CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle    intr/s
06:47:49 PM  all   94.32    0.00    0.06    0.00    0.00    0.00    0.00    0.00    5.62  15888.00
06:47:50 PM  all   94.57    0.00    0.62    0.00    0.00    0.00    0.00    0.00    4.81  16180.00
06:47:51 PM  all   94.69    0.00    0.06    0.00    0.00    0.00    0.00    0.00    5.25  15966.00
06:47:52 PM  all   95.81    0.00    0.00    0.00    0.00    0.00    0.00    0.00    4.19  16053.00
06:47:53 PM  all   94.88    0.06    0.00    0.00    0.00    0.00    0.00    0.00    5.06  15984.00
06:47:54 PM  all   93.31    0.00    0.00    0.00    0.00    0.00    0.00    0.00    6.69  15806.00
06:47:55 PM  all   94.19    0.00    0.06    0.00    0.00    0.00    0.00    0.00    5.75  15896.00
06:47:56 PM  all   92.87    0.00    0.00    0.00    0.00    0.00    0.00    0.00    7.13  15716.00
06:47:57 PM  all   94.88    0.00    0.00    0.00    0.00    0.00    0.00    0.00    5.12  15982.00
06:47:58 PM  all   95.44    0.00    0.00    0.00    0.00    0.00    0.00    0.00    4.56  16075.00
Average:     all   94.49    0.01    0.08    0.00    0.00    0.00    0.00    0.00    5.42  15954.60

-tip+patches:

06:47:03 PM  CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle    intr/s
06:47:04 PM  all  100.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  16630.00
06:47:05 PM  all   99.69    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.31  16580.20
06:47:06 PM  all   99.69    0.00    0.06    0.00    0.00    0.00    0.00    0.00    0.25  16596.00
06:47:07 PM  all   99.20    0.00    0.74    0.00    0.00    0.06    0.00    0.00    0.00  17838.61
06:47:08 PM  all  100.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  16540.00
06:47:09 PM  all  100.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  16575.00
06:47:10 PM  all  100.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00  16614.00
06:47:11 PM  all   99.94    0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.06  16588.00
06:47:12 PM  all   99.94    0.00    0.06    0.00    0.00    0.00    0.00    0.00    0.00  16593.00
06:47:13 PM  all   99.94    0.00    0.06    0.00    0.00    0.00    0.00    0.00    0.00  16551.00
Average:     all   99.84    0.00    0.09    0.00    0.00    0.01    0.00    0.00    0.06  16711.58

We see an improvement in idle% on the system (drops from 5.42% on -tip to 0.06%
with the patches).

We see an improvement in idle% on the system (drops from 5.42%
on -tip to 0.06% with the patches).

Signed-off-by: Nikhil Rao <ncrao@google.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Nikunj A. Dadhania <nikunj@linux.vnet.ibm.com>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Stephan Barwolf <stephan.baerwolf@tu-ilmenau.de>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Link: http://lkml.kernel.org/r/1305754668-18792-1-git-send-email-ncrao@google.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>

1 files changed, 21 insertions, 2 deletions

diff --git a/include/linux/sched.h b/include/linux/sched.h
index f2f440221b70..c34a718e20dd 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -788,9 +788,28 @@ enum cpu_idle_type {
 };
 
 /*
- * Increase resolution of nice-level calculations:
+ * Increase resolution of nice-level calculations for 64-bit architectures.
+ * The extra resolution improves shares distribution and load balancing of
+ * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
+ * hierarchies, especially on larger systems. This is not a user-visible change
+ * and does not change the user-interface for setting shares/weights.
+ *
+ * We increase resolution only if we have enough bits to allow this increased
+ * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
+ * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
+ * increased costs.
  */
-#define SCHED_LOAD_SHIFT	10
+#if BITS_PER_LONG > 32
+# define SCHED_LOAD_RESOLUTION	10
+# define scale_load(w)		((w) << SCHED_LOAD_RESOLUTION)
+# define scale_load_down(w)	((w) >> SCHED_LOAD_RESOLUTION)
+#else
+# define SCHED_LOAD_RESOLUTION	0
+# define scale_load(w)		(w)
+# define scale_load_down(w)	(w)
+#endif
+
+#define SCHED_LOAD_SHIFT	(10 + SCHED_LOAD_RESOLUTION)
 #define SCHED_LOAD_SCALE	(1L << SCHED_LOAD_SHIFT)
 
 /*