Re: [RFC PATCH] hrtimer: interleave timers for improved single thread performance at low utilization

[Ingo Molnar]

* shrikanth hegde <sshegde@xxxxxxxxxxxxxxxxxx> wrote:

> As per current design of hrtimer, it uses the _softexpires to trigger the
> timer function.  _softexpires is set as multiple of the period/interval value.
> This will benefit the power saving by less wakeups. Due to this, different
> timers of the same period/interval values align and the callbacks functions
> will be called at the same time.
> 
> CPU bandwidth controller (CPU cgroup) uses these hrtimers to implement period
> and quota.  Period timer refills the quota and allow the throttled cgroups to
> start running again.  When there are multiple such cgroup's, if their period
> values are same, then these period timers will be aligned.  Hence multiple
> cgroup's timer fire at the same time and ends up unthrottling each cgroups
> runqueues. Since all cgroups start, they would compete for CPU and use all SMT
> threads likely.
> 
> There is 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%. This is likely true when
> using containers. If the timers aren't rounded-off, then the unthrottled
> cgroup can run freely without many context switches and can also benefit of SMT
> Folding[1]. This effect will be further amplified in SPLPAR environment[2] as
> this would cause less hypervisor preemptions. There can be benefit due to less
> IPI storm as well. Docker provides a config option of period timer value,
> whereas the kubernetes only provides millicore option. Hence with typical
> deployment period values will be set to 100ms as kubernetes millicore will
> set the quota accordingly without altering period values.
> 
> [1] SMT folding is a mechanism were processor core reconfigured to lower SMT
> mode to improve performance when some sibling threads are idle. In a SMT8 core,
> when only one or two threads are running on a core, we get the best throughput
> compared to running all 8 threads.
> 
> [2] SPLPAR is an Shared Processor Logical PARtition. There can be many SPLPARs
> running on the same physical machine sharing the CPU resources.  One SPLPAR can
> consume all CPU resource it can, if the other SPLPARs are idle. Processors
> within the SPLPAR are called vCPU. vCPU can be higher than CPU.  Hence at an
> instance of time if there are more requested vCPU than CPU, then vCPU can be
> preempted. When the timers align, there will be spike in requested vCPU when
> the timers expire. This can lead to preemption when the other SPLPARs are not
> idle.
> 
> Came up with a naive patch, more of hack. Other alternative is to use a
> slightly modified API for cgroups, so that all other timers align and wakeups
> remain reduced. New hrtimer api is likely better, i can send out the patch
> quickly.  Here i am trying to misalign by setting the softexpire at multiple of
> interval/10 instead of interval. Ran the stress-ng with two cgroups. The
> numbers are with patch and without patch on Power10 machine with SMT=8. Below
> table shows time taken by each group to complete. In the last column, both
> cgroup's are run together and data shows average time taken by cgroups to
> complete. Here each cgroup is assigned 25% runtime.
> 
> workload: stress-ng --cpu=4 --cpu-ops=100000 data shows time it took to
> complete in seconds for each run.
> 
> Without Patch:
> period/quota    cgroup1 runs    cgroup2 runs    cgroup1 &cgroup2
>                    alone           alone         run together
> 100ms/200ms         120s            120s            155s
>                     120s            120s            155s
>                     120s            120s            155s
> With Patch:
> period/quota    cgroup1 runs    cgroup2 runs    cgroup1 & cgroup2
>                    alone           alone         run together
> 100ms/200ms         120s            120s            131s
>                     120s            120s            155s
>                     120s            120s            121s
> 
> There is no benefit at higher utilization of 50% or more. There is no
> degradation also.
> 
> Signed-off by: Shrikanth Hegde <sshegde@xxxxxxxxxxxxxxxxxx>
> ---
>  kernel/time/hrtimer.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> diff --git a/kernel/time/hrtimer.c b/kernel/time/hrtimer.c
> index 3ae661ab6260..d160f49f0cce 100644
> --- a/kernel/time/hrtimer.c
> +++ b/kernel/time/hrtimer.c
> @@ -1055,6 +1055,17 @@ u64 hrtimer_forward(struct hrtimer *timer, ktime_t now, ktime_t interval)
> 
>  		orun = ktime_divns(delta, incr);
>  		hrtimer_add_expires_ns(timer, incr * orun);
> +		/*
> +		 * Avoid timer round-off, so that all cfs bandwidth timers
> +		 * don't start at the same time
> +		 */
> +		if (incr >= 100000000ULL) {
> +			s64 interleave = 0;
> +			interleave = ktime_sub_ns(delta,  incr * orun);
> +			interleave = interleave - (ktime_to_ns(delta) % (incr/10));
> +			if (interleave > 0)
> +				hrtimer_add_expires_ns(timer, interleave);
> +		}

Any reason why you did this in the hrtimer code, instead of the 
(sched_cfs_period_timer?) hrtimer handler itself?

Thanks,

	Ingo