Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: [RFC][PATCH 13/16] sched: Add core wide task selection and
 scheduling.
To:     Julien Desfossez <jdesfossez@digitalocean.com>,
        Peter Zijlstra <peterz@infradead.org>, mingo@kernel.org,
        tglx@linutronix.de, pjt@google.com, tim.c.chen@linux.intel.com,
        torvalds@linux-foundation.org
Cc:     linux-kernel@vger.kernel.org, fweisbec@gmail.com,
        keescook@chromium.org, kerrnel@google.com,
        Vineeth Pillai <vpillai@digitalocean.com>,
        Nishanth Aravamudan <naravamudan@digitalocean.com>,
        Aaron Lu <aaron.lu@linux.alibaba.com>
References: <20190218173514.667598558@infradead.org>
 <1554835135-11814-1-git-send-email-jdesfossez@digitalocean.com>
From:   Subhra Mazumdar <subhra.mazumdar@oracle.com>
Message-ID: <3514bb97-33bc-3a14-3f15-767682b2f035@oracle.com>
Date:   Wed, 10 Apr 2019 17:11:59 -0700
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101
 Thunderbird/60.5.1
MIME-Version: 1.0
In-Reply-To: <1554835135-11814-1-git-send-email-jdesfossez@digitalocean.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Transfer-Encoding: 8bit
Content-Language: en-US
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk


On 4/9/19 11:38 AM, Julien Desfossez wrote:
> We found the source of the major performance regression we discussed
> previously. It turns out there was a pattern where a task (a kworker in this
> case) could be woken up, but the core could still end up idle before that
> task had a chance to run.
>
> Example sequence, cpu0 and cpu1 and siblings on the same core, task1 and
> task2 are in the same cgroup with the tag enabled (each following line
> happens in the increasing order of time):
> - task1 running on cpu0, task2 running on cpu1
> - sched_waking(kworker/0, target_cpu=cpu0)
> - task1 scheduled out of cpu0
> - kworker/0 cannot run on cpu0 because of task2 is still running on cpu1
>    cpu0 is idle
> - task2 scheduled out of cpu1
> - cpu1 doesn’t select kworker/0 for cpu0, because the optimization path ends
>    the task selection if core_cookie is NULL for currently selected process
>    and the cpu1’s runqueue.
> - cpu1 is idle
> --> both siblings are idle but kworker/0 is still in the run queue of cpu0.
>      Cpu0 may stay idle for longer if it goes deep idle.
>
> With the fix below, we ensure to send an IPI to the sibling if it is idle
> and has tasks waiting in its runqueue.
> This fixes the performance issue we were seeing.
>
> Now here is what we can measure with a disk write-intensive benchmark:
> - no performance impact with enabling core scheduling without any tagged
>    task,
> - 5% overhead if one tagged task is competing with an untagged task,
> - 10% overhead if 2 tasks tagged with a different tag are competing
>    against each other.
>
> We are starting more scaling tests, but this is very encouraging !
>
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index e1fa10561279..02c862a5e973 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -3779,7 +3779,22 @@ pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
>   
>   				trace_printk("unconstrained pick: %s/%d %lx\n",
>   						next->comm, next->pid, next->core_cookie);
> +				rq->core_pick = NULL;
>   
> +				/*
> +				 * If the sibling is idling, we might want to wake it
> +				 * so that it can check for any runnable but blocked tasks
> +				 * due to previous task matching.
> +				 */
> +				for_each_cpu(j, smt_mask) {
> +					struct rq *rq_j = cpu_rq(j);
> +					rq_j->core_pick = NULL;
> +					if (j != cpu && is_idle_task(rq_j->curr) && rq_j->nr_running) {
> +						resched_curr(rq_j);
> +						trace_printk("IPI(%d->%d[%d]) idle preempt\n",
> +							     cpu, j, rq_j->nr_running);
> +					}
> +				}
>   				goto done;
>   			}
>   
I see similar improvement with this patch as removing the condition I
earlier mentioned. So that's not needed. I also included the patch for the
priority fix. For 2 DB instances, HT disabling stands at -22% for 32 users
(from earlier emails).


1 DB instance

users  baseline   %idle    core_sched %idle
16     1          84       -4.9% 84
24     1          76       -6.7% 75
32     1          69       -2.4% 69

2 DB instance

users  baseline   %idle    core_sched %idle
16     1          66       -19.5% 69
24     1          54       -9.8% 57
32     1          42       -27.2%        48