[PATCH v4 09/10] sched/fair: disable stealing if too many NUMA nodes

From: Steve Sistare
Date: Thu Dec 06 2018 - 16:39:12 EST


The STEAL feature causes regressions on hackbench on larger NUMA systems,
so disable it on systems with more than sched_steal_node_limit nodes
(default 2). Note that the feature remains enabled as seen in features.h
and /sys/kernel/debug/sched_features, but stealing is only performed if
nodes <= sched_steal_node_limit. This arrangement allows users to activate
stealing on reboot by setting the kernel parameter sched_steal_node_limit
on kernels built without CONFIG_SCHED_DEBUG. The parameter is temporary
and will be deleted when the regression is fixed.

Details of the regression follow. With the STEAL feature set, hackbench
is slower on many-node systems:

X5-8: 8 sockets * 18 cores * 2 hyperthreads = 288 CPUs
Intel(R) Xeon(R) CPU E7-8895 v3 @ 2.60GHz
Average of 10 runs of: hackbench <groups> processes 50000

--- base -- --- new ---
groups time %stdev time %stdev %speedup
1 3.627 15.8 3.876 7.3 -6.5
2 4.545 24.7 5.583 16.7 -18.6
3 5.716 25.0 7.367 14.2 -22.5
4 6.901 32.9 7.718 14.5 -10.6
8 8.604 38.5 9.111 16.0 -5.6
16 7.734 6.8 11.007 8.2 -29.8

Total CPU time increases. Profiling shows that CPU time increases
uniformly across all functions, suggesting a systemic increase in cache
or memory latency. This may be due to NUMA migrations, as they cause
loss of LLC cache footprint and remote memory latencies.

The domains for this system and their flags are:

domain0 (SMT) : 1 core
SD_LOAD_BALANCE SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK
SD_SHARE_PKG_RESOURCES SD_PREFER_SIBLING SD_SHARE_CPUCAPACITY
SD_WAKE_AFFINE

domain1 (MC) : 1 socket
SD_LOAD_BALANCE SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK
SD_SHARE_PKG_RESOURCES SD_PREFER_SIBLING
SD_WAKE_AFFINE

domain2 (NUMA) : 4 sockets
SD_LOAD_BALANCE SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK
SD_SERIALIZE SD_OVERLAP SD_NUMA
SD_WAKE_AFFINE

domain3 (NUMA) : 8 sockets
SD_LOAD_BALANCE SD_BALANCE_NEWIDLE
SD_SERIALIZE SD_OVERLAP SD_NUMA

Schedstats point to the root cause of the regression. hackbench is run
10 times per group and the average schedstat accumulation per-run and
per-cpu is shown below. Note that domain3 moves are zero because
SD_WAKE_AFFINE is not set there.

NO_STEAL
--- domain2 --- --- domain3 ---
grp time %busy sched idle wake steal remote move pull remote move pull
1 20.3 10.3 28710 14346 14366 0 490 3378 0 4039 0 0
2 26.4 18.8 56721 28258 28469 0 792 7026 12 9229 0 7
3 29.9 28.3 90191 44933 45272 0 5380 7204 19 16481 0 3
4 30.2 35.8 121324 60409 60933 0 7012 9372 27 21438 0 5
8 27.7 64.2 229174 111917 117272 0 11991 1837 168 44006 0 32
16 32.6 74.0 334615 146784 188043 0 3404 1468 49 61405 0 8

STEAL
--- domain2 --- --- domain3 ---
grp time %busy sched idle wake steal remote move pull remote move pull
1 20.6 10.2 28490 14232 14261 18 3 3525 0 4254 0 0
2 27.9 18.8 56757 28203 28562 303 1675 7839 5 9690 0 2
3 35.3 27.7 87337 43274 44085 698 741 12785 14 15689 0 3
4 36.8 36.0 118630 58437 60216 1579 2973 14101 28 18732 0 7
8 48.1 73.8 289374 133681 155600 18646 35340 10179 171 65889 0 34
16 41.4 82.5 268925 91908 177172 47498 17206 6940 176 71776 0 20

Cross-numa-node migrations are caused by load balancing pulls and
wake_affine moves. Pulls are small and similar for no_steal and steal.
However, moves are significantly higher for steal, and rows above with the
highest moves have the worst regressions for time; see for example grp=8.

Moves increase for steal due to the following logic in wake_affine_idle()
for synchronous wakeup:

if (sync && cpu_rq(this_cpu)->nr_running == 1)
return this_cpu; // move the task

The steal feature does a better job of smoothing the load between idle
and busy CPUs, so nr_running is 1 more often, and moves are performed
more often. For hackbench, cross-node affine moves early in the run are
good because they colocate wakers and wakees from the same group on the
same node, but continued moves later in the run are bad, because the wakee
is moved away from peers on its previous node. Note that even no_steal
is far from optimal; binding an instance of "hackbench 2" to each of the
8 NUMA nodes runs much faster than running "hackbench 16" with no binding.

Clearing SD_WAKE_AFFINE for domain2 eliminates the affine cross-node
migrations and eliminates the difference between no_steal and steal
performance. However, overall performance is lower than WA_IDLE because
some migrations are helpful as explained above.

I have tried many heuristics in a attempt to optimize the number of
cross-node moves in all conditions, with limited success. The fundamental
problem is that the scheduler does not track which groups of tasks talk to
each other. Parts of several groups become entrenched on the same node,
filling it to capacity, leaving no room for either group to pull its peers
over, and there is neither data nor mechanism for the scheduler to evict
one group to make room for the other.

For now, disable STEAL on such systems until we can do better, or it is
shown that hackbench is atypical and most workloads benefit from stealing.

Signed-off-by: Steve Sistare <steven.sistare@xxxxxxxxxx>
---
kernel/sched/fair.c | 16 +++++++++++++---
kernel/sched/sched.h | 2 +-
kernel/sched/topology.c | 25 +++++++++++++++++++++++++
3 files changed, 39 insertions(+), 4 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 1476ae8..1efd9c4 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -3735,11 +3735,21 @@ static inline void rq_idle_stamp_clear(struct rq *rq)
rq->idle_stamp = 0;
}

+static inline bool steal_enabled(void)
+{
+#ifdef CONFIG_NUMA
+ bool allow = static_branch_likely(&sched_steal_allow);
+#else
+ bool allow = true;
+#endif
+ return sched_feat(STEAL) && allow;
+}
+
static void overload_clear(struct rq *rq)
{
struct sparsemask *overload_cpus;

- if (!sched_feat(STEAL))
+ if (!steal_enabled())
return;

rcu_read_lock();
@@ -3753,7 +3763,7 @@ static void overload_set(struct rq *rq)
{
struct sparsemask *overload_cpus;

- if (!sched_feat(STEAL))
+ if (!steal_enabled())
return;

rcu_read_lock();
@@ -9902,7 +9912,7 @@ static int try_steal(struct rq *dst_rq, struct rq_flags *dst_rf)
int stolen = 0;
struct sparsemask *overload_cpus;

- if (!sched_feat(STEAL))
+ if (!steal_enabled())
return 0;

if (!cpu_active(dst_cpu))
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index eacf5db..2a28340 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -936,7 +936,6 @@ static inline int cpu_of(struct rq *rq)
#endif
}

-
#ifdef CONFIG_SCHED_SMT

extern struct static_key_false sched_smt_present;
@@ -1185,6 +1184,7 @@ enum numa_topology_type {
#endif

#ifdef CONFIG_NUMA
+extern struct static_key_true sched_steal_allow;
extern void sched_init_numa(void);
extern void sched_domains_numa_masks_set(unsigned int cpu);
extern void sched_domains_numa_masks_clear(unsigned int cpu);
diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
index 89a78ce..259d659 100644
--- a/kernel/sched/topology.c
+++ b/kernel/sched/topology.c
@@ -1344,6 +1344,30 @@ static void init_numa_topology_type(void)
}
}

+DEFINE_STATIC_KEY_TRUE(sched_steal_allow);
+static int sched_steal_node_limit;
+#define SCHED_STEAL_NODE_LIMIT_DEFAULT 2
+
+static int __init steal_node_limit_setup(char *buf)
+{
+ get_option(&buf, &sched_steal_node_limit);
+ return 0;
+}
+
+early_param("sched_steal_node_limit", steal_node_limit_setup);
+
+static void check_node_limit(void)
+{
+ int n = num_possible_nodes();
+
+ if (sched_steal_node_limit == 0)
+ sched_steal_node_limit = SCHED_STEAL_NODE_LIMIT_DEFAULT;
+ if (n > sched_steal_node_limit) {
+ static_branch_disable(&sched_steal_allow);
+ pr_debug("Suppressing sched STEAL. To enable, reboot with sched_steal_node_limit=%d", n);
+ }
+}
+
void sched_init_numa(void)
{
int next_distance, curr_distance = node_distance(0, 0);
@@ -1492,6 +1516,7 @@ void sched_init_numa(void)
sched_max_numa_distance = sched_domains_numa_distance[level - 1];

init_numa_topology_type();
+ check_node_limit();
}

void sched_domains_numa_masks_set(unsigned int cpu)
--
1.8.3.1