summaryrefslogtreecommitdiff
path: root/kernel/sched/fair.c
diff options
context:
space:
mode:
authorMel Gorman <mgorman@techsingularity.net>2020-01-14 10:13:20 +0000
committerIngo Molnar <mingo@kernel.org>2020-01-28 21:36:55 +0100
commitb396f52326de20ec974471b7b19168867b365cbf (patch)
tree6548b26df07836163f9c7f376e29070878e45ff1 /kernel/sched/fair.c
parentebc0f83c78a2d26384401ecf2d2fa48063c0ee27 (diff)
sched/fair: Allow a small load imbalance between low utilisation SD_NUMA domains
The CPU load balancer balances between different domains to spread load and strives to have equal balance everywhere. Communicating tasks can migrate so they are topologically close to each other but these decisions are independent. On a lightly loaded NUMA machine, two communicating tasks pulled together at wakeup time can be pushed apart by the load balancer. In isolation, the load balancer decision is fine but it ignores the tasks data locality and the wakeup/LB paths continually conflict. NUMA balancing is also a factor but it also simply conflicts with the load balancer. This patch allows a fixed degree of imbalance of two tasks to exist between NUMA domains regardless of utilisation levels. In many cases, this prevents communicating tasks being pulled apart. It was evaluated whether the imbalance should be scaled to the domain size. However, no additional benefit was measured across a range of workloads and machines and scaling adds the risk that lower domains have to be rebalanced. While this could change again in the future, such a change should specify the use case and benefit. The most obvious impact is on netperf TCP_STREAM -- two simple communicating tasks with some softirq offload depending on the transmission rate. 2-socket Haswell machine 48 core, HT enabled netperf-tcp -- mmtests config config-network-netperf-unbound baseline lbnuma-v3 Hmean 64 568.73 ( 0.00%) 577.56 * 1.55%* Hmean 128 1089.98 ( 0.00%) 1128.06 * 3.49%* Hmean 256 2061.72 ( 0.00%) 2104.39 * 2.07%* Hmean 1024 7254.27 ( 0.00%) 7557.52 * 4.18%* Hmean 2048 11729.20 ( 0.00%) 13350.67 * 13.82%* Hmean 3312 15309.08 ( 0.00%) 18058.95 * 17.96%* Hmean 4096 17338.75 ( 0.00%) 20483.66 * 18.14%* Hmean 8192 25047.12 ( 0.00%) 27806.84 * 11.02%* Hmean 16384 27359.55 ( 0.00%) 33071.88 * 20.88%* Stddev 64 2.16 ( 0.00%) 2.02 ( 6.53%) Stddev 128 2.31 ( 0.00%) 2.19 ( 5.05%) Stddev 256 11.88 ( 0.00%) 3.22 ( 72.88%) Stddev 1024 23.68 ( 0.00%) 7.24 ( 69.43%) Stddev 2048 79.46 ( 0.00%) 71.49 ( 10.03%) Stddev 3312 26.71 ( 0.00%) 57.80 (-116.41%) Stddev 4096 185.57 ( 0.00%) 96.15 ( 48.19%) Stddev 8192 245.80 ( 0.00%) 100.73 ( 59.02%) Stddev 16384 207.31 ( 0.00%) 141.65 ( 31.67%) In this case, there was a sizable improvement to performance and a general reduction in variance. However, this is not univeral. For most machines, the impact was roughly a 3% performance gain. Ops NUMA base-page range updates 19796.00 292.00 Ops NUMA PTE updates 19796.00 292.00 Ops NUMA PMD updates 0.00 0.00 Ops NUMA hint faults 16113.00 143.00 Ops NUMA hint local faults % 8407.00 142.00 Ops NUMA hint local percent 52.18 99.30 Ops NUMA pages migrated 4244.00 1.00 Without the patch, only 52.18% of sampled accesses are local. In an earlier changelog, 100% of sampled accesses are local and indeed on most machines, this was still the case. In this specific case, the local sampled rates was 99.3% but note the "base-page range updates" and "PTE updates". The activity with the patch is negligible as were the number of faults. The small number of pages migrated were related to shared libraries. A 2-socket Broadwell showed better results on average but are not presented for brevity as the performance was similar except it showed 100% of the sampled NUMA hints were local. The patch holds up for a 4-socket Haswell, an AMD EPYC and AMD Epyc 2 machine. For dbench, the impact depends on the filesystem used and the number of clients. On XFS, there is little difference as the clients typically communicate with workqueues which have a separate class of scheduler problem at the moment. For ext4, performance is generally better, particularly for small numbers of clients as NUMA balancing activity is negligible with the patch applied. A more interesting example is the Facebook schbench which uses a number of messaging threads to communicate with worker threads. In this configuration, one messaging thread is used per NUMA node and the number of worker threads is varied. The 50, 75, 90, 95, 99, 99.5 and 99.9 percentiles for response latency is then reported. Lat 50.00th-qrtle-1 44.00 ( 0.00%) 37.00 ( 15.91%) Lat 75.00th-qrtle-1 53.00 ( 0.00%) 41.00 ( 22.64%) Lat 90.00th-qrtle-1 57.00 ( 0.00%) 42.00 ( 26.32%) Lat 95.00th-qrtle-1 63.00 ( 0.00%) 43.00 ( 31.75%) Lat 99.00th-qrtle-1 76.00 ( 0.00%) 51.00 ( 32.89%) Lat 99.50th-qrtle-1 89.00 ( 0.00%) 52.00 ( 41.57%) Lat 99.90th-qrtle-1 98.00 ( 0.00%) 55.00 ( 43.88%) Lat 50.00th-qrtle-2 42.00 ( 0.00%) 42.00 ( 0.00%) Lat 75.00th-qrtle-2 48.00 ( 0.00%) 47.00 ( 2.08%) Lat 90.00th-qrtle-2 53.00 ( 0.00%) 52.00 ( 1.89%) Lat 95.00th-qrtle-2 55.00 ( 0.00%) 53.00 ( 3.64%) Lat 99.00th-qrtle-2 62.00 ( 0.00%) 60.00 ( 3.23%) Lat 99.50th-qrtle-2 63.00 ( 0.00%) 63.00 ( 0.00%) Lat 99.90th-qrtle-2 68.00 ( 0.00%) 66.00 ( 2.94% For higher worker threads, the differences become negligible but it's interesting to note the difference in wakeup latency at low utilisation and mpstat confirms that activity was almost all on one node until the number of worker threads increase. Hackbench generally showed neutral results across a range of machines. This is different to earlier versions of the patch which allowed imbalances for higher degrees of utilisation. perf bench pipe showed negligible differences in overall performance as the differences are very close to the noise. An earlier prototype of the patch showed major regressions for NAS C-class when running with only half of the available CPUs -- 20-30% performance hits were measured at the time. With this version of the patch, the impact is negligible with small gains/losses within the noise measured. This is because the number of threads far exceeds the small imbalance the aptch cares about. Similarly, there were report of regressions for the autonuma benchmark against earlier versions but again, normal load balancing now applies for that workload. In general, the patch simply seeks to avoid unnecessary cross-node migrations in the basic case where imbalances are very small. For low utilisation communicating workloads, this patch generally behaves better with less NUMA balancing activity. For high utilisation, there is no change in behaviour. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Valentin Schneider <valentin.schneider@arm.com> Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Acked-by: Phil Auld <pauld@redhat.com> Tested-by: Phil Auld <pauld@redhat.com> Link: https://lkml.kernel.org/r/20200114101319.GO3466@techsingularity.net
Diffstat (limited to 'kernel/sched/fair.c')
-rw-r--r--kernel/sched/fair.c41
1 files changed, 29 insertions, 12 deletions
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index fe4e0d775375..25dffc03f0f6 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -8658,10 +8658,6 @@ static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *s
/*
* Try to use spare capacity of local group without overloading it or
* emptying busiest.
- * XXX Spreading tasks across NUMA nodes is not always the best policy
- * and special care should be taken for SD_NUMA domain level before
- * spreading the tasks. For now, load_balance() fully relies on
- * NUMA_BALANCING and fbq_classify_group/rq to override the decision.
*/
if (local->group_type == group_has_spare) {
if (busiest->group_type > group_fully_busy) {
@@ -8701,16 +8697,37 @@ static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *s
env->migration_type = migrate_task;
lsub_positive(&nr_diff, local->sum_nr_running);
env->imbalance = nr_diff >> 1;
- return;
- }
+ } else {
- /*
- * If there is no overload, we just want to even the number of
- * idle cpus.
- */
- env->migration_type = migrate_task;
- env->imbalance = max_t(long, 0, (local->idle_cpus -
+ /*
+ * If there is no overload, we just want to even the number of
+ * idle cpus.
+ */
+ env->migration_type = migrate_task;
+ env->imbalance = max_t(long, 0, (local->idle_cpus -
busiest->idle_cpus) >> 1);
+ }
+
+ /* Consider allowing a small imbalance between NUMA groups */
+ if (env->sd->flags & SD_NUMA) {
+ unsigned int imbalance_min;
+
+ /*
+ * Compute an allowed imbalance based on a simple
+ * pair of communicating tasks that should remain
+ * local and ignore them.
+ *
+ * NOTE: Generally this would have been based on
+ * the domain size and this was evaluated. However,
+ * the benefit is similar across a range of workloads
+ * and machines but scaling by the domain size adds
+ * the risk that lower domains have to be rebalanced.
+ */
+ imbalance_min = 2;
+ if (busiest->sum_nr_running <= imbalance_min)
+ env->imbalance = 0;
+ }
+
return;
}