[RFC PATCH] drm/scheduler: use idle time to do better loadbalance

[Christian König]

Am 12.01.20 um 02:25 schrieb Nirmoy Das:
> This patch adds required fields to drm_sched_job and drm_gpu_scheduler
> structure to cumulatively calculate amount of time a drm_gpu_scheduler
> spend on serving a job.
>
> Using least used drm scheduler to choose a run queue
> improves drm_sched_entity_get_free_sched()'s job distribution
>
> Below are test results after running amdgpu_test from mesa drm
>
> Before this patch:
>
> sched_name     num of many times it got scheduled
> =========      ==================================
> sdma0	       314
> sdma1          32
> comp_1.0.0     56
>
> After this patch:
>
> sched_name     num of many times it got scheduled
> =========      ==================================
> sdma0	       113
> sdma1          383
> comp_1.0.0     9
> comp_1.0.1     9
> comp_1.1.0     8
> comp_1.1.1     8
> comp_1.2.0    12
> comp_1.2.1    13
> comp_1.3.0    16
> comp_1.3.1    9
>
> Signed-off-by: Nirmoy Das <nirmoy.das at amd.com>

Well that is a nice start, but a couple of comments below.

> ---
>   drivers/gpu/drm/scheduler/sched_entity.c | 9 +++++----
>   drivers/gpu/drm/scheduler/sched_main.c   | 2 ++
>   include/drm/gpu_scheduler.h              | 2 ++
>   3 files changed, 9 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/scheduler/sched_entity.c b/drivers/gpu/drm/scheduler/sched_entity.c
> index 2e3a058fc239..b5555af787d0 100644
> --- a/drivers/gpu/drm/scheduler/sched_entity.c
> +++ b/drivers/gpu/drm/scheduler/sched_entity.c
> @@ -130,7 +130,7 @@ static struct drm_sched_rq *
>   drm_sched_entity_get_free_sched(struct drm_sched_entity *entity)
>   {
>   	struct drm_sched_rq *rq = NULL;
> -	unsigned int min_jobs = UINT_MAX, num_jobs;
> +	uint64_t min_time_consumed = -1, total_consumed_time;
>   	int i;
>   
>   	for (i = 0; i < entity->num_sched_list; ++i) {
> @@ -141,9 +141,9 @@ drm_sched_entity_get_free_sched(struct drm_sched_entity *entity)
>   			continue;
>   		}
>   
> -		num_jobs = atomic_read(&sched->num_jobs);
> -		if (num_jobs < min_jobs) {
> -			min_jobs = num_jobs;
> +		total_consumed_time = sched->total_consumed_time;
> +		if (total_consumed_time < min_time_consumed) {
> +			min_time_consumed = total_consumed_time;
>   			rq = &entity->sched_list[i]->sched_rq[entity->priority];

You might want to remove num_jobs now completely since it unused, but 
consider the whole algorithm first. See below for that.

>   		}
>   	}
> @@ -499,6 +499,7 @@ void drm_sched_entity_push_job(struct drm_sched_job *sched_job,
>   
>   	trace_drm_sched_job(sched_job, entity);
>   	atomic_inc(&entity->rq->sched->num_jobs);
> +	sched_job->start_time = ktime_get_ns();
>   	WRITE_ONCE(entity->last_user, current->group_leader);
>   	first = spsc_queue_push(&entity->job_queue, &sched_job->queue_node);
>   
> diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
> index 3fad5876a13f..67fdf4f248d4 100644
> --- a/drivers/gpu/drm/scheduler/sched_main.c
> +++ b/drivers/gpu/drm/scheduler/sched_main.c
> @@ -653,6 +653,7 @@ static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
>   	struct drm_sched_job *s_job = container_of(cb, struct drm_sched_job, cb);
>   	struct drm_sched_fence *s_fence = s_job->s_fence;
>   	struct drm_gpu_scheduler *sched = s_fence->sched;
> +	uint64_t end = ktime_get_ns();
>   
>   	atomic_dec(&sched->hw_rq_count);
>   	atomic_dec(&sched->num_jobs);
> @@ -660,6 +661,7 @@ static void drm_sched_process_job(struct dma_fence *f, struct dma_fence_cb *cb)
>   	trace_drm_sched_process_job(s_fence);
>   
>   	drm_sched_fence_finished(s_fence);
> +	s_job->sched->total_consumed_time += end - s_job->start_time;

Don't use s_job->sched here, use the local variable sched (s_job might 
already be destroyed).

The next problem is that you somehow need to protect total_consumed_time 
from concurrent updates, a 64bit atomic should probably do it.

Another problem from the algorithm point of view is that the 
total_time_consumed of each scheduler doesn't necessary say anything 
about a good scheduling decision for the future.

For example there could have been a lot of work from a single queue to 
SDMA0 in the past, but now both SDMA0 and SDMA1 are idle.

Scheduling everything to SDMA1 until we catch up with SDMA0 is certainly 
not a good idea. Since this counters the whole idea of load balancing.

Regards,
Christian.

>   	wake_up_interruptible(&sched->wake_up_worker);
>   }
>   
> diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
> index 96a1a1b7526e..496d9b209d12 100644
> --- a/include/drm/gpu_scheduler.h
> +++ b/include/drm/gpu_scheduler.h
> @@ -191,6 +191,7 @@ struct drm_sched_job {
>   	struct dma_fence_cb		finish_cb;
>   	struct list_head		node;
>   	uint64_t			id;
> +	uint64_t			start_time;
>   	atomic_t			karma;
>   	enum drm_sched_priority		s_priority;
>   	struct drm_sched_entity  *entity;
> @@ -285,6 +286,7 @@ struct drm_gpu_scheduler {
>   	atomic_t                        num_jobs;
>   	bool			ready;
>   	bool				free_guilty;
> +	uint64_t			total_consumed_time;
>   };
>   
>   int drm_sched_init(struct drm_gpu_scheduler *sched,