load_balance() fixes for affinity

We noticed in testing that affined workloads do not provide consistent performance under certain circumstances. To isolate the issue, we began testing with a representative CPU workload. In some cases the CPU workload results would vary by a large percentage from run to run. We used JTAG and scheduler trace to try and profile the issue and noticed that at least one core was left idle for the duration of the test in the low score runs.

The specific test methodology used in our investigation involved using an executable that was compiled to fork a specific number of workers. The executable was then affined to a number of cores equal to the number of workers forked, using the userspace taskset utility. The expectation was that all cores in the mask of runnable CPUs would run a process until the work was complete. In practice, for several affinity masks, one to two cores would not receive work until late in the test or not at all.

Our first observation was that it appeared initial cpu assignment of the forked threads appeared suboptimal. We observed that it was highly probable that many or all of the threads would be initially assigned to the same CPU. Our limited investigation into this indicated that the forking of the threads occurs fast enough that load estimates for CPUs have not adjusted to the work they have been assigned. The function select_task_rq_fair() does not take number of running process into consideration and thus can easily assign several tasks spawned close together to the same CPU within the allowed CPUs for a workload. While this seems suboptimal, the load_balance() algorithm run on other idle cores in the affinity mask should resolve imbalances that result from it.

As our test case was leaving cores idle for many seconds at a time, we dug into load_balance() code and specifically the group_imbalance path. We added ftrace to the path in question and noticed that two branches in load_balance() conflict in their assessments of the ability of the algorithm to migrate load. The group_imbalance path correctly sets the flag to indicate the group can not be properly balanced due to affinity, but the redo condition right after this branch incorrectly assumes that there may be other cores with work to be pulled by considering cores outside of the scheduling domain in question.

Consider the following circumstance where (T) represents tasks and (*) represents the allowed CPUs:

A B C
{ 0 1 } { 2 3 } { 4 5 }
T T T T
T
* * * * *

In this case, core 0 in scheduling group 'A' should run load balance on its local scheduling domain and try to pull work from core 1. When it fails to move any work due to affinity, it will mark scheduling group 'A' as "group_imbalance". However, right after doing so, core 0 load_balance() evaluates the redo path with core 1 removed from consideration. This should leave no other cores to consider and result in a jump to "out_all_pinned", leaving the "group_imbalance" flag to be seen by future runs of the algorithm on core 3. As the code stands currently, core 0 evaluates all cores instead of those in the scheduling domain under consideration. It incorrectly sees other cores in the system and attempts a redo. With core 1 removed from the mask, the redo identifies no load imbalance and results in a jump to "out_balanced", clearing the "group_imbalance" flag.

We propose correcting this logic with patch 1 of this series.

Making this correction revealed an issue in the calculate_imbalance() path. Patch 2 removes a branch that does not make sense with the current load_balance() algorithm because it has no scenario where it benifits the "group_imbalance" case in calculate_imbalance() and which causes problems in systems with affined workloads and many idle or lightly loaded CPUs.

Jeffrey Hugo (2):
sched/fair: Fix load_balance() affinity redo path sched/fair: Remove group imbalance from calculate_imbalance()

 kernel/sched/fair.c | 22 ++++++++++++----------
 1 file changed, 12 insertions(+), 10 deletions(-)

Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

If load_balance() fails to migrate any tasks because all tasks were affined, load_balance() removes the source cpu from consideration and attempts to redo and balance among the new subset of cpus.

There is a bug in this code path where the algorithm considers all active cpus in the system (minus the source that was just masked out). This is not valid for two reasons: some active cpus may not be in the current scheduling domain and one of the active cpus is dst_cpu. These cpus should not be considered, as we cannot pull load from them.

Instead of failing out of load_balance(), we may end up redoing the search with no valid cpus and incorrectly concluding the domain is balanced. Additionally, if the group_imbalance flag was just set, it may also be incorrectly unset, thus the flag will not be seen by other cpus in future load_balance() runs as that algorithm intends.

Fix the check by removing cpus not in the current domain and the dst_cpu from considertation, thus limiting the evaluation to valid remaining cpus from which load might be migrated.

---
 kernel/sched/fair.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

index d711093..8f783ba 100644
@@ -8219,8 +8219,19 @@ static int load_balance(int this_cpu, struct rq *this_rq,
 
 		/* All tasks on this runqueue were pinned by CPU affinity */
 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
+			struct cpumask tmp;
+
+			/* Cpumask of all initially possible busiest cpus. */
+			cpumask_copy(&tmp, sched_domain_span(env.sd));
+			cpumask_clear_cpu(env.dst_cpu, &tmp);
+
 			cpumask_clear_cpu(cpu_of(busiest), cpus);
-			if (!cpumask_empty(cpus)) {
+			/*
+			 * Go back to "redo" iff the load-balance cpumask
+			 * contains other potential busiest cpus for the
+			 * current sched domain.
+			 */
+			if (cpumask_intersects(cpus, &tmp)) {
 				env.loop = 0;
 				env.loop_break = sched_nr_migrate_break;
 				goto redo;

Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

So per that Chain Austin wrote the patch, who handed it to Dietmar, who handed it to you. Except I don't see a From: Austin on.

What gives?

Austin and I did the investigations and wrote the initial version. We discussed it with Dietmar, who suggested some significant rewrites which we felt added to the readability of the code. The current version posted on the list you've seen was basically written by all three of us, so I listed the authors in alphabetical order to properly give credit to all involved.

Is there a better way to handle patches which have authorship from multiple people?

Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

Well, Signed-off-by is only a chain of custody thing. It says who handled the patches and that they have the right to publish and that sorts of thing. We have a document describing this.

It does _NOT_ however imply any kind of authorship what so ever. Of course, the author must be the first in the custody chain, how else could the patch 'escape'.

Authorship comes from the Author: header, and there's only 1 of those.

Just mention the people by name in the Changelog or something.

I'm not entirely sure I agree with that assessment, but I'll discuss it with the folks offline and see what we want to roll into the next patch version after seeing what the other comments are.

Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

Well, we need a temp variable to store the intermediate values since the cpumask_* operations are somewhat limited, and require a "storage" parameter.

Do you have any suggestions to meet all of these requirements?

cpus == cpu_active_mask, which we compare against below in just a few lines with the cpumask_intersects check. So, no, I don't think we did forget to mask with cpu_active_mask.

Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

What about we use env.dst_grpmask and check if cpus is an improper subset of env.dst_grpmask? In this case we have to get rid of setting env.dst_grpmask = NULL in case of CPU_NEWLY_IDLE which is IMHO not an issue since it's idle is passed via env into can_migrate_task().
And cpus has to be and'ed with sched_domain_span(env.sd).

I'm not sure if this will work with 'not fully connected NUMA' (SD_OVERLAP) though ...

index a903276fcb62..2ede4c1c9db8 100644
@@ -6737,10 +6737,10 @@ int can_migrate_task(struct task_struct *p, struct lb_env *env)
                 * our sched_group. We may want to revisit it if we couldn't
                 * meet load balance goals by pulling other tasks on src_cpu.
                 *
-                * Also avoid computing new_dst_cpu if we have already computed
-                * one in current iteration.
+                * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
+                * already computed one in current iteration.
                 */
-               if (!env->dst_grpmask || (env->flags & LBF_DST_PINNED))
+               if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
                        return 0;
 
                /* Prevent to re-select dst_cpu via env's cpus */

@@ -8091,14 +8091,7 @@ static int load_balance(int this_cpu, struct rq *this_rq,
                .tasks          = LIST_HEAD_INIT(env.tasks),
        };
 
-       /*
-        * For NEWLY_IDLE load_balancing, we don't need to consider
-        * other cpus in our group
-        */
-       if (idle == CPU_NEWLY_IDLE)
-               env.dst_grpmask = NULL;
-
-       cpumask_copy(cpus, cpu_active_mask);
+       cpumask_and(cpus, cpu_active_mask, sched_domain_span(env.sd));
 
        schedstat_inc(sd->lb_count[idle]);
 

@@ -8220,7 +8213,7 @@ static int load_balance(int this_cpu, struct rq *this_rq,
                /* All tasks on this runqueue were pinned by CPU affinity */
                if (unlikely(env.flags & LBF_ALL_PINNED)) {
                        cpumask_clear_cpu(cpu_of(busiest), cpus);
-                       if (!cpumask_empty(cpus)) {
+                       if (!cpumask_subset(cpus, env.dst_grpmask)) {
                                env.loop = 0;
                                env.loop_break = sched_nr_migrate_break;
                                goto redo;

Hmm. I follow the idea, but I'm not too confident in the SD_OVERLAP case, and looking at your proposed code, it seems invasive to me - changes are needed in what would otherwise be unrelated sections of code. I'd prefer not to go in that direction. Also, it appears that the dst_cpu is still considered as a source for load.

We've got a different idea to address the stack issue, and still keep the change "contained", which I'll roll into a V2 today or tomorrow.

Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.

The group_imbalance path in calculate_imbalance() made sense when it was added back in 2007 with commit 908a7c1b9b80 ("sched: fix improper load balance across sched domain") because busiest->load_per_task factored into the amount of imbalance that was calculated. That is not the case today.

The group_imbalance path can only affect the outcome of calculate_imbalance() when the average load of the domain is less than the original busiest->load_per_task. In this case, busiest->load_per_task is overwritten with the scheduling domain load average. Thus busiest->load_per_task no longer represents actual load that can be moved.

At the final comparison between env->imbalance and busiest->load_per_task, imbalance may be larger than the new busiest->load_per_task causing the check to fail under the assumption that there is a task that could be migrated to satisfy the imbalance. However env->imbalance may still be smaller than the original busiest->load_per_task, thus it is unlikely that there is a task that can be migrated to satisfy the imbalance. Calculate_imbalance() would not choose to run fix_small_imbalance() when we expect it should. In the worst case, this can result in idle cpus.

Since the group imbalance path in calculate_imbalance() is at best a NOP but otherwise harmful, remove it.

---
 kernel/sched/fair.c | 9 ---------
 1 file changed, 9 deletions(-)

index 8f783ba..3283561 100644
@@ -7760,15 +7760,6 @@ static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *s
 	local = &sds->local_stat;
 	busiest = &sds->busiest_stat;
 
-	if (busiest->group_type == group_imbalanced) {
-		/*
-		 * In the group_imb case we cannot rely on group-wide averages
-		 * to ensure cpu-load equilibrium, look at wider averages. XXX
-		 */
-		busiest->load_per_task =
-			min(busiest->load_per_task, sds->avg_load);
-	}
-
 	/*
 	 * Avg load of busiest sg can be less and avg load of local sg can
 	 * be greater than avg load across all sgs of sd because avg load

Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.