Re: [PATCH V3] Interleave cfs bandwidth timers for improved single thread performance at low utilization

[Peter Zijlstra]

On Fri, Feb 24, 2023 at 12:21:53AM +0530, Shrikanth Hegde wrote:
> 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
> 
> Signed-off-by: Shrikanth Hegde<sshegde@xxxxxxxxxxxxxxxxxx>
> Suggested-by: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Suggested-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> 
> Initial RFC PATCH, discussions and details on the problem:
> Link1: https://lore.kernel.org/lkml/5ae3cb09-8c9a-11e8-75a7-cc774d9bc283@xxxxxxxxxxxxxxxxxx/
> Link2: https://lore.kernel.org/lkml/9c57c92c-3e0c-b8c5-4be9-8f4df344a347@xxxxxxxxxxxxxxxxxx/

Picked up for tip/sched/core.