| Message ID | 20221024212634.27230-1-luben.tuikov@amd.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v1] drm/scheduler: Set the FIFO scheduling policy as the default | expand |
Pushed to drm-misc-next. Christian. Am 24.10.22 um 23:26 schrieb Luben Tuikov: > The currently default Round-Robin GPU scheduling can result in starvation > of entities which have a large number of jobs, over entities which have > a very small number of jobs (single digit). > > This can be illustrated in the following diagram, where jobs are > alphabetized to show their chronological order of arrival, where job A is > the oldest, B is the second oldest, and so on, to J, the most recent job to > arrive. > > ---> entities > j | H-F-----A--E--I-- > o | --G-----B-----J-- > b | --------C-------- > s\/ --------D-------- > > WLOG, assuming all jobs are "ready", then a R-R scheduling will execute them > in the following order (a slice off of the top of the entities' list), > > H, F, A, E, I, G, B, J, C, D. > > However, to mitigate job starvation, we'd rather execute C and D before E, > and so on, given, of course, that they're all ready to be executed. > > So, if all jobs are ready at this instant, the order of execution for this > and the next 9 instances of picking the next job to execute, should really > be, > > A, B, C, D, E, F, G, H, I, J, > > which is their chronological order. The only reason for this order to be > broken, is if an older job is not yet ready, but a younger job is ready, at > an instant of picking a new job to execute. For instance if job C wasn't > ready at time 2, but job D was ready, then we'd pick job D, like this: > > 0 +1 +2 ... > A, B, D, ... > > And from then on, C would be preferred before all other jobs, if it is ready > at the time when a new job for execution is picked. So, if C became ready > two steps later, the execution order would look like this: > > ......0 +1 +2 ... > A, B, D, E, C, F, G, H, I, J > > This is what the FIFO GPU scheduling algorithm achieves. It uses a > Red-Black tree to keep jobs sorted in chronological order, where picking > the oldest job is O(1) (we use the "cached" structure), and balancing the > tree is O(log n). IOW, it picks the *oldest ready* job to execute now. > > The implementation is already in the kernel, and this commit only changes > the default GPU scheduling algorithm to use. > > This was tested and achieves about 1% faster performance over the Round > Robin algorithm. > > Cc: Christian König <christian.koenig@amd.com> > Cc: Alex Deucher <Alexander.Deucher@amd.com> > Cc: Direct Rendering Infrastructure - Development <dri-devel@lists.freedesktop.org> > Signed-off-by: Luben Tuikov <luben.tuikov@amd.com> > Reviewed-by: Christian König <christian.koenig@amd.com> > --- > drivers/gpu/drm/scheduler/sched_main.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c > index 2fab218d708279..d0ff9e11cb69fa 100644 > --- a/drivers/gpu/drm/scheduler/sched_main.c > +++ b/drivers/gpu/drm/scheduler/sched_main.c > @@ -62,13 +62,13 @@ > #define to_drm_sched_job(sched_job) \ > container_of((sched_job), struct drm_sched_job, queue_node) > > -int drm_sched_policy = DRM_SCHED_POLICY_RR; > +int drm_sched_policy = DRM_SCHED_POLICY_FIFO; > > /** > * DOC: sched_policy (int) > * Used to override default entities scheduling policy in a run queue. > */ > -MODULE_PARM_DESC(sched_policy, "Specify schedule policy for entities on a runqueue, " __stringify(DRM_SCHED_POLICY_RR) " = Round Robin (default), " __stringify(DRM_SCHED_POLICY_FIFO) " = use FIFO."); > +MODULE_PARM_DESC(sched_policy, "Specify the scheduling policy for entities on a run-queue, " __stringify(DRM_SCHED_POLICY_RR) " = Round Robin, " __stringify(DRM_SCHED_POLICY_FIFO) " = FIFO (default)."); > module_param_named(sched_policy, drm_sched_policy, int, 0444); > > static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a, > > base-commit: 27b2ae654370e1a8e446b0e48c4e406abed12ca1
diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c index 2fab218d708279..d0ff9e11cb69fa 100644 --- a/drivers/gpu/drm/scheduler/sched_main.c +++ b/drivers/gpu/drm/scheduler/sched_main.c @@ -62,13 +62,13 @@ #define to_drm_sched_job(sched_job) \ container_of((sched_job), struct drm_sched_job, queue_node) -int drm_sched_policy = DRM_SCHED_POLICY_RR; +int drm_sched_policy = DRM_SCHED_POLICY_FIFO; /** * DOC: sched_policy (int) * Used to override default entities scheduling policy in a run queue. */ -MODULE_PARM_DESC(sched_policy, "Specify schedule policy for entities on a runqueue, " __stringify(DRM_SCHED_POLICY_RR) " = Round Robin (default), " __stringify(DRM_SCHED_POLICY_FIFO) " = use FIFO."); +MODULE_PARM_DESC(sched_policy, "Specify the scheduling policy for entities on a run-queue, " __stringify(DRM_SCHED_POLICY_RR) " = Round Robin, " __stringify(DRM_SCHED_POLICY_FIFO) " = FIFO (default)."); module_param_named(sched_policy, drm_sched_policy, int, 0444); static __always_inline bool drm_sched_entity_compare_before(struct rb_node *a,
The currently default Round-Robin GPU scheduling can result in starvation of entities which have a large number of jobs, over entities which have a very small number of jobs (single digit). This can be illustrated in the following diagram, where jobs are alphabetized to show their chronological order of arrival, where job A is the oldest, B is the second oldest, and so on, to J, the most recent job to arrive. ---> entities j | H-F-----A--E--I-- o | --G-----B-----J-- b | --------C-------- s\/ --------D-------- WLOG, assuming all jobs are "ready", then a R-R scheduling will execute them in the following order (a slice off of the top of the entities' list), H, F, A, E, I, G, B, J, C, D. However, to mitigate job starvation, we'd rather execute C and D before E, and so on, given, of course, that they're all ready to be executed. So, if all jobs are ready at this instant, the order of execution for this and the next 9 instances of picking the next job to execute, should really be, A, B, C, D, E, F, G, H, I, J, which is their chronological order. The only reason for this order to be broken, is if an older job is not yet ready, but a younger job is ready, at an instant of picking a new job to execute. For instance if job C wasn't ready at time 2, but job D was ready, then we'd pick job D, like this: 0 +1 +2 ... A, B, D, ... And from then on, C would be preferred before all other jobs, if it is ready at the time when a new job for execution is picked. So, if C became ready two steps later, the execution order would look like this: ......0 +1 +2 ... A, B, D, E, C, F, G, H, I, J This is what the FIFO GPU scheduling algorithm achieves. It uses a Red-Black tree to keep jobs sorted in chronological order, where picking the oldest job is O(1) (we use the "cached" structure), and balancing the tree is O(log n). IOW, it picks the *oldest ready* job to execute now. The implementation is already in the kernel, and this commit only changes the default GPU scheduling algorithm to use. This was tested and achieves about 1% faster performance over the Round Robin algorithm. Cc: Christian König <christian.koenig@amd.com> Cc: Alex Deucher <Alexander.Deucher@amd.com> Cc: Direct Rendering Infrastructure - Development <dri-devel@lists.freedesktop.org> Signed-off-by: Luben Tuikov <luben.tuikov@amd.com> Reviewed-by: Christian König <christian.koenig@amd.com> --- drivers/gpu/drm/scheduler/sched_main.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) base-commit: 27b2ae654370e1a8e446b0e48c4e406abed12ca1