for sched-devel.git
cover for sched-devel.git
Hi Ingo,
Would you stick these into sched-devel.
The first patch should address the latency isolation issue. While the second rectifies a massive brainfart :-)
PATCH 1/2 sched: fair: virtual deadline scheduling
Change CFS into a virtual deadline scheduler.
By flattening the grouping hierarchy into a single level we now end up with tasks that have varying latency requirements. Tasks from group A might have a larger latency period than those from group B as the period depends on the number of runnable tasks within a group.
The current scheduling criteria does not take that into account - it assumes a single latency period. In order to accommodate these varying latencies in the scheduling decision, move to EDF [*] scheduling. We treat the tasks need + its latency period as the deadline it has to meet. This includes the latency into the scheduling decision.
[*] - EDF is correct up until load 1, after that it is not a closed system so improvement is possible here. It is usable because the system strives to generate the load 1 situation.
--- include/linux/sched.h | 1 + kernel/sched_fair.c | 6 +++++- 2 files changed, 6 insertions(+), 1 deletion(-)
include/linux/sched.h +1 -0
@@ -925,6 +925,7 @@ struct sched_entity {u64 exec_start;u64 sum_exec_runtime;u64 vruntime; +u64 vperiod;u64 prev_sum_exec_runtime;#ifdef CONFIG_SCHEDSTATS
kernel/sched_fair.c +5 -1
@@ -220,9 +220,11 @@ static inline u64 min_vruntime(u64 min_vstatic inline s64 entity_key (struct cfs_rq *cfs_rq,struct sched_entity *se) { -return se->vruntime - cfs_rq->min_vruntime; +return se->vruntime + se->vperiod - cfs_rq->min_vruntime; } +static u64 sched_vslice_add (struct cfs_rq *cfs_rq,struct sched_entity *se); +/* * Enqueue an entity into the rb-tree: */
@@ -240,6 +242,8 @@ static void __enqueue_entity(struct cfs_if (se == cfs_rq->curr)return ; + se->vperiod =sched_vslice_add (cfs_rq, se); + cfs_rq = &rq_of (cfs_rq)->cfs; link = &cfs_rq->tasks_timeline.rb_node;
--
PATCH 2/2 sched: fair: fix calc_delta_asym
In reply to Peter ZijlstraThe goal of calc_delta_asym() is to be asymetrically around NICE_0_LOAD, in that it favours >=0 over <0. The current implementation does not achieve that.
-20 |
|
0 --------+-------
.'19 .'
--- kernel/sched_fair.c | 13 +++++++++---- 1 file changed, 9 insertions(+), 4 deletions(-)
kernel/sched_fair.c +9 -4
@@ -417,12 +417,17 @@ calc_delta_fair(unsigned long delta, strstatic inline unsigned long calc_delta_asym (unsigned long delta,struct sched_entity *se) { -unsigned long fair =calc_delta_fair (delta, se); +if (se->load.weight <NICE_0_LOAD ) { +struct load_weight lw = { + .weight =NICE_0_LOAD , + .inv_weight =1UL << (WMULT_SHIFT -NICE_0_SHIFT ) + }; -if (fair > delta) - fair = delta; + delta =calc_delta_mine (delta,cfs_rq_of (se)->load.weight, &lw); + se =parent_entity (se); + } -return fair; +return calc_delta_fair (delta, se); }/*
--
Quoted 6 lines · Peter Zijlstra
Hi Ingo,
Would you stick these into sched-devel.
The first patch should address the latency isolation issue. While the second rectifies a massive brainfart :-)
hehe :-) applied.
Ingo