sched: Interleave cfs bandwidth timers for improved single thread performance at...

sched: Interleave cfs bandwidth timers for improved single thread performance at low utilization

CPU cfs bandwidth controller uses hrtimer. Currently there is no initial
value set. Hence all period timers would align at expiry.
This happens when there are multiple CPU cgroup's.

There is a performance gain that can be achieved here if the timers are
interleaved when the utilization of each CPU cgroup is low and total
utilization of all the CPU cgroup's is less than 50%. If the timers are
interleaved, then the unthrottled cgroup can run freely without many
context switches and can also benefit from SMT Folding. This effect will
be further amplified in SPLPAR environment.

This commit adds a random offset after initializing each hrtimer. This
would result in interleaving the timers at expiry, which helps in achieving
the said performance gain.

This was tested on powerpc platform with 8 core SMT=8. Socket power was
measured when the workload. Benchmarked the stress-ng with power
information. Throughput oriented benchmarks show significant gain up to
25% while power consumption increases up to 15%.

Workload: stress-ng --cpu=32 --cpu-ops=50000.
1CG - 1 cgroup is running.
2CG - 2 cgroups are running together.
Time taken to complete stress-ng in seconds and power is in watts.
each cgroup is throttled at 25% with 100ms as the period value.
           6.2-rc6                     |   with patch
8 core   1CG    power   2CG     power  |  1CG    power  2 CG    power
        27.5    80.6    40      90     |  27.3    82    32.3    104
        27.5    81      40.2    91     |  27.5    81    38.7     96
        27.7    80      40.1    89     |  27.6    80    29.7    106
        27.7    80.1    40.3    94     |  27.6    80    31.5    105

Latency might be affected by this change. That could happen if the CPU was
in a deep idle state which is possible if we interleave the timers. Used
schbench for measuring the latency. Each cgroup is throttled at 25% with
period value is set to 100ms. Numbers are when both the cgroups are
running simultaneously. Latency values don't degrade much. Some
improvement is seen in tail latencies.

		6.2-rc6        with patch
Groups: 16
50.0th:          39.5            42.5
75.0th:         924.0           922.0
90.0th:         972.0           968.0
95.0th:        1005.5           994.0
99.0th:        4166.0          2287.0
99.5th:        7314.0          7448.0
99.9th:       15024.0         13600.0

Groups: 32
50.0th:         819.0           463.0
75.0th:        1596.0           918.0
90.0th:        5992.0          1281.5
95.0th:       13184.0          2765.0
99.0th:       21792.0         14240.0
99.5th:       25696.0         18920.0
99.9th:       33280.0         35776.0

Groups: 64
50.0th:        4806.0          3440.0
75.0th:       31136.0         33664.0
90.0th:       54144.0         58752.0
95.0th:       66176.0         67200.0
99.0th:       84736.0         91520.0
99.5th:       97408.0        114048.0
99.9th:      136448.0        140032.0

Initial RFC PATCH, discussions and details on the problem:

Link1: https://lore.kernel.org/lkml/5ae3cb09-8c9a-11e8-75a7-cc774d9bc283@linux.vnet.ibm.com/
Link2: https://lore.kernel.org/lkml/9c57c92c-3e0c-b8c5-4be9-8f4df344a347@linux.vnet.ibm.com/
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Suggested-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Shrikanth Hegde<sshegde@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Ben Segall <bsegall@google.com>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20230223185153.1499710-1-sshegde@linux.vnet.ibm.com