[PATCH 1/1] intel_pstate: Increase hold-off time before busyness is scaled

From: Mel Gorman
Date: Thu Feb 18 2016 - 06:11:32 EST

(cc'ing pm and scheduler people as the problem could be blamed on either
subsystem depending on your point of view)

The PID relies on samples of equal time but this does not apply for
deferrable timers when the CPU is idle. intel_pstate checks if the actual
duration between samples is large and if so, the "busyness" of the CPU
is scaled.

This assumes the delay was a deferred timer but a workload may simply have
been idle for a short time if it's context switching between a server and
client or waiting very briefly on IO. It's compounded by the problem that
server/clients migrate between CPUs due to wake-affine trying to maximise
hot cache usage. In such cases, the cores are not considered busy and the
frequency is dropped prematurely.

This patch increases the hold-off value before the busyness is scaled. It
was selected based simply on testing until the desired result was found.
Tests were conducted with workloads that are either client/server based
or short-lived IO.

dbench4
4.5.0-rc2 4.5.0-rc2
vanilla sample-v1r1
Hmean mb/sec-1 309.82 ( 0.00%) 327.01 ( 5.55%)
Hmean mb/sec-2 594.92 ( 0.00%) 613.02 ( 3.04%)
Hmean mb/sec-4 669.17 ( 0.00%) 712.27 ( 6.44%)
Hmean mb/sec-8 700.82 ( 0.00%) 724.04 ( 3.31%)
Hmean mb/sec-64 425.38 ( 0.00%) 448.02 ( 5.32%)

4.5.0-rc2 4.5.0-rc2
vanilla sample-v1r1
Mean %Busy 27.28 26.81
Mean CPU%c1 42.50 44.29
Mean CPU%c3 7.16 7.14
Mean CPU%c6 23.05 21.76
Mean CPU%c7 0.00 0.00
Mean CorWatt 4.60 5.08
Mean PkgWatt 6.83 7.32

There is fairly sizable performance boost from the modification and while
the percentage of time spent in C1 is increased, it is not by a substantial
amount and the power usage increase is tiny.

iozone for small files and varying block sizes. Format is IOOperation-filessize-recordsize

4.5.0-rc2 4.5.0-rc2
vanilla sample-v1r1
Hmean SeqWrite-200704-1 740152.30 ( 0.00%) 748432.35 ( 1.12%)
Hmean SeqWrite-200704-2 1052506.25 ( 0.00%) 1169065.30 ( 11.07%)
Hmean SeqWrite-200704-4 1450716.41 ( 0.00%) 1725335.69 ( 18.93%)
Hmean SeqWrite-200704-8 1523917.72 ( 0.00%) 1881610.25 ( 23.47%)
Hmean SeqWrite-200704-16 1572519.89 ( 0.00%) 1750277.07 ( 11.30%)
Hmean SeqWrite-200704-32 1611078.69 ( 0.00%) 1923796.62 ( 19.41%)
Hmean SeqWrite-200704-64 1656755.37 ( 0.00%) 1892766.99 ( 14.25%)
Hmean SeqWrite-200704-128 1641739.24 ( 0.00%) 1952081.27 ( 18.90%)
Hmean SeqWrite-200704-256 1660046.05 ( 0.00%) 1931237.50 ( 16.34%)
Hmean SeqWrite-200704-512 1634394.86 ( 0.00%) 1860369.95 ( 13.83%)
Hmean SeqWrite-200704-1024 1629526.38 ( 0.00%) 1810320.92 ( 11.09%)
Hmean SeqWrite-401408-1 828943.43 ( 0.00%) 876152.50 ( 5.70%)
Hmean SeqWrite-401408-2 1231519.20 ( 0.00%) 1368986.18 ( 11.16%)
Hmean SeqWrite-401408-4 1724109.56 ( 0.00%) 1838265.22 ( 6.62%)
Hmean SeqWrite-401408-8 1806615.84 ( 0.00%) 1969611.74 ( 9.02%)
Hmean SeqWrite-401408-16 1859268.96 ( 0.00%) 2003005.51 ( 7.73%)
Hmean SeqWrite-401408-32 1887759.67 ( 0.00%) 2415913.37 ( 27.98%)
Hmean SeqWrite-401408-64 1941717.11 ( 0.00%) 1971929.24 ( 1.56%)
Hmean SeqWrite-401408-128 1919515.58 ( 0.00%) 2127647.53 ( 10.84%)
Hmean SeqWrite-401408-256 1908766.57 ( 0.00%) 2067473.02 ( 8.31%)
Hmean SeqWrite-401408-512 1908999.37 ( 0.00%) 2195587.56 ( 15.01%)
Hmean SeqWrite-401408-1024 1912232.98 ( 0.00%) 2150068.56 ( 12.44%)
Hmean Rewrite-200704-1 1151067.57 ( 0.00%) 1155309.64 ( 0.37%)
Hmean Rewrite-200704-2 1786824.53 ( 0.00%) 1837093.18 ( 2.81%)
Hmean Rewrite-200704-4 2539338.19 ( 0.00%) 2649019.78 ( 4.32%)
Hmean Rewrite-200704-8 2687411.53 ( 0.00%) 2785202.26 ( 3.64%)
Hmean Rewrite-200704-16 2709445.97 ( 0.00%) 2805580.76 ( 3.55%)
Hmean Rewrite-200704-32 2735718.43 ( 0.00%) 2807532.87 ( 2.63%)
Hmean Rewrite-200704-64 2782754.97 ( 0.00%) 2952024.38 ( 6.08%)
Hmean Rewrite-200704-128 2791889.73 ( 0.00%) 2805048.02 ( 0.47%)
Hmean Rewrite-200704-256 2711596.34 ( 0.00%) 2828896.54 ( 4.33%)
Hmean Rewrite-200704-512 2665066.25 ( 0.00%) 2868058.05 ( 7.62%)
Hmean Rewrite-200704-1024 2675375.89 ( 0.00%) 2685664.19 ( 0.38%)
Hmean Rewrite-401408-1 1350713.78 ( 0.00%) 1358762.21 ( 0.60%)
Hmean Rewrite-401408-2 2079420.61 ( 0.00%) 2097399.02 ( 0.86%)
Hmean Rewrite-401408-4 2889535.90 ( 0.00%) 2912795.03 ( 0.80%)
Hmean Rewrite-401408-8 3068155.32 ( 0.00%) 3090915.84 ( 0.74%)
Hmean Rewrite-401408-16 3103789.43 ( 0.00%) 3162486.65 ( 1.89%)
Hmean Rewrite-401408-32 3112447.72 ( 0.00%) 3243067.63 ( 4.20%)
Hmean Rewrite-401408-64 3232651.39 ( 0.00%) 3227701.02 ( -0.15%)
Hmean Rewrite-401408-128 3149556.47 ( 0.00%) 3165694.24 ( 0.51%)
Hmean Rewrite-401408-256 3093348.93 ( 0.00%) 3104229.97 ( 0.35%)
Hmean Rewrite-401408-512 3026305.45 ( 0.00%) 3121151.02 ( 3.13%)
Hmean Rewrite-401408-1024 3005431.18 ( 0.00%) 3046910.32 ( 1.38%)

4.5.0-rc2 4.5.0-rc2
vanilla sample-v1r1
Mean %Busy 3.10 3.09
Mean CPU%c1 6.16 5.55
Mean CPU%c3 0.08 0.10
Mean CPU%c6 90.65 91.26
Mean CPU%c7 0.00 0.00
Mean CorWatt 1.71 1.74
Mean PkgWatt 3.88 3.91
Max %Busy 16.51 16.22
Max CPU%c1 17.03 21.99
Max CPU%c3 2.57 2.15
Max CPU%c6 96.39 96.31
Max CPU%c7 0.00 0.00
Max CorWatt 5.40 5.42
Max PkgWatt 7.53 7.56

The other operations are omitted as they showed no performance difference.
For sequential writes and rewrites there is a massive gain in throughput
for very small files. The increase in power consumption is negligible.
It is known that the increase is not universal. Larger core machines see
a much smaller benefit so the rate of CPU migrations are a factor.

netperf-UDP_STREAM

4.5.0-rc2 4.5.0-rc2
vanilla sample-v1r1
Hmean send-64 233.96 ( 0.00%) 244.76 ( 4.61%)
Hmean send-128 466.74 ( 0.00%) 479.16 ( 2.66%)
Hmean send-256 929.12 ( 0.00%) 964.00 ( 3.75%)
Hmean send-1024 3631.36 ( 0.00%) 3781.89 ( 4.15%)
Hmean send-2048 6984.60 ( 0.00%) 7169.60 ( 2.65%)
Hmean send-3312 10792.94 ( 0.00%) 11103.42 ( 2.88%)
Hmean send-4096 12895.57 ( 0.00%) 13112.58 ( 1.68%)
Hmean send-8192 23057.34 ( 0.00%) 23443.80 ( 1.68%)
Hmean send-16384 37871.11 ( 0.00%) 38292.60 ( 1.11%)
Hmean recv-64 233.89 ( 0.00%) 244.71 ( 4.63%)
Hmean recv-128 466.63 ( 0.00%) 479.09 ( 2.67%)
Hmean recv-256 928.88 ( 0.00%) 963.74 ( 3.75%)
Hmean recv-1024 3630.54 ( 0.00%) 3780.96 ( 4.14%)
Hmean recv-2048 6983.20 ( 0.00%) 7167.55 ( 2.64%)
Hmean recv-3312 10790.92 ( 0.00%) 11100.63 ( 2.87%)
Hmean recv-4096 12891.37 ( 0.00%) 13110.35 ( 1.70%)
Hmean recv-8192 23054.79 ( 0.00%) 23438.27 ( 1.66%)
Hmean recv-16384 37866.79 ( 0.00%) 38283.73 ( 1.10%)

4.5.0-rc2 4.5.0-rc2
vanilla sample-v1r1
Mean %Busy 37.30 37.10
Mean CPU%c1 37.52 37.30
Mean CPU%c3 0.10 0.10
Mean CPU%c6 25.08 25.49
Mean CPU%c7 0.00 0.00
Mean CorWatt 11.20 11.18
Mean PkgWatt 13.30 13.28
Max %Busy 50.64 51.73
Max CPU%c1 49.80 50.53
Max CPU%c3 9.14 8.95
Max CPU%c6 62.46 63.48
Max CPU%c7 0.00 0.00
Max CorWatt 16.46 16.44
Max PkgWatt 18.58 18.55

In this test, the client/server are pinned to cores so the scheduler
decisions are not a factor. There is still a mild performance boost
with no impact on power consumption.

cyclictest-pinned
4.5.0-rc2 4.5.0-rc2
vanilla sample-v1r1
Amean LatAvg 3.00 ( 0.00%) 2.64 ( 11.94%)
Amean LatMax 156.93 ( 0.00%) 106.89 ( 31.89%)

4.5.0-rc2 4.5.0-rc2
vanilla sample-v1r1
Mean %Busy 99.74 99.73
Mean CPU%c1 0.02 0.02
Mean CPU%c3 0.00 0.01
Mean CPU%c6 0.23 0.24
Mean CPU%c7 0.00 0.00
Mean CorWatt 5.06 5.92
Mean PkgWatt 7.12 7.99
Max %Busy 100.00 100.00
Max CPU%c1 3.88 3.50
Max CPU%c3 0.71 0.99
Max CPU%c6 41.79 43.17
Max CPU%c7 0.00 0.00
Max CorWatt 6.80 8.66
Max PkgWatt 8.85 10.71

This test measures how quickly a task wakes up after a timeout. The test
could be defeated by selecting a different timeout value that is outside
the new hold-off value. Furthermore, a workload that is very sensitive to
wakeup latencies should use the performance governor. Nevertheless it's
interesting to note the impact of increasing the hold-off value. There is
an increase in power usage because the CPU remains active during sleep times.

In all cases, there are some CPU migrations because wakers pull wakees to
nearby CPUs. It could be argued that the workload should be pinned but this
puts a burden on the user that may not even be possible in all cases. The
scheduler could try keeping processes on the same CPUs but that would impact
cache hotness and cause a different class of issues. It is inevitable that
there will be some conflict between power management and scheduling decisions
but there is some gains from delaying idling slightly without a severe impact
on power consumption.

Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
---
drivers/cpufreq/intel_pstate.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
index cd83d477e32d..54250084174a 100644
--- a/drivers/cpufreq/intel_pstate.c
+++ b/drivers/cpufreq/intel_pstate.c
@@ -999,7 +999,7 @@ static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
sample_time = pid_params.sample_rate_ms * USEC_PER_MSEC;
duration_us = ktime_us_delta(cpu->sample.time,
cpu->last_sample_time);
- if (duration_us > sample_time * 3) {
+ if (duration_us > sample_time * 12) {
sample_ratio = div_fp(int_tofp(sample_time),
int_tofp(duration_us));
core_busy = mul_fp(core_busy, sample_ratio);
--
2.6.4