[lttng-dev] [PATCH 14/16] urcu-qsbr: batch concurrent synchronize_rcu()
Mathieu Desnoyers
mathieu.desnoyers at efficios.com
Tue Nov 20 14:40:27 EST 2012
Here are benchmarks on batching of synchronize_rcu(), and it leads to
very interesting scalability improvement and speedups, e.g., on a
24-core AMD, with a write-heavy scenario (4 readers threads, 20 updater
threads, each updater using synchronize_rcu()):
* Serialized grace periods :
./test_urcu_qsbr 4 20 20
SUMMARY ./test_urcu_qsbr testdur 20 nr_readers 4
rdur 0 wdur 0 nr_writers 20 wdelay 0
nr_reads 20251412728 nr_writes 1826331 nr_ops 20253239059
* Batched grace periods :
./test_urcu_qsbr 4 20 20
SUMMARY ./test_urcu_qsbr testdur 20 nr_readers 4
rdur 0 wdur 0 nr_writers 20 wdelay 0
nr_reads 15141994746 nr_writes 9382515 nr_ops 15151377261
For a 9382515/1826331 = 5.13 speedup for 20 updaters.
Of course, we can see that readers have slowed down, probably due to
increased update traffic, given there is no change to the read-side code
whatsoever.
Now let's see the penality of managing the stack for single-updater.
With 4 readers, single updater:
* Serialized grace periods :
./test_urcu_qsbr 4 1 20
SUMMARY ./test_urcu_qsbr testdur 20 nr_readers 4
rdur 0 wdur 0 nr_writers 1 wdelay 0
nr_reads 19240784755 nr_writes 2130839 nr_ops 19242915594
* Batched grace periods :
./test_urcu_qsbr 4 1 20
SUMMARY ./test_urcu_qsbr testdur 20 nr_readers 4
rdur 0 wdur 0 nr_writers 1 wdelay 0
nr_reads 19160162768 nr_writes 2253068 nr_ops 1916241583
2253068 vs 2137036 -> a couple of runs show that this difference lost in
the noise for single updater.
CC: Paul E. McKenney <paulmck at linux.vnet.ibm.com>
CC: Lai Jiangshan <laijs at cn.fujitsu.com>
CC: Alan Stern <stern at rowland.harvard.edu>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers at efficios.com>
---
urcu-qsbr.c | 151 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 151 insertions(+)
diff --git a/urcu-qsbr.c b/urcu-qsbr.c
index 5b341b5..7f747ed 100644
--- a/urcu-qsbr.c
+++ b/urcu-qsbr.c
@@ -36,6 +36,7 @@
#include <poll.h>
#include "urcu/wfcqueue.h"
+#include "urcu/wfstack.h"
#include "urcu/map/urcu-qsbr.h"
#define BUILD_QSBR_LIB
#include "urcu/static/urcu-qsbr.h"
@@ -78,6 +79,35 @@ DEFINE_URCU_TLS(unsigned int, rcu_rand_yield);
static CDS_LIST_HEAD(registry);
+/*
+ * Number of busy-loop attempts before waiting on futex for grace period
+ * batching.
+ */
+#define RCU_AWAKE_ATTEMPTS 1000
+
+enum adapt_wakeup_state {
+ /* AWAKE_WAITING is compared directly (futex compares it). */
+ AWAKE_WAITING = 0,
+ /* non-zero are used as masks. */
+ AWAKE_WAKEUP = (1 << 0),
+ AWAKE_AWAKENED = (1 << 1),
+ AWAKE_TEARDOWN = (1 << 2),
+};
+
+struct gp_waiters_thread {
+ struct cds_wfs_node node;
+ int32_t wait_futex;
+};
+
+/*
+ * Stack keeping threads awaiting to wait for a grace period. Contains
+ * struct gp_waiters_thread objects.
+ */
+static struct cds_wfs_stack gp_waiters = {
+ .head = CDS_WFS_END,
+ .lock = PTHREAD_MUTEX_INITIALIZER,
+};
+
static void mutex_lock(pthread_mutex_t *mutex)
{
int ret;
@@ -116,6 +146,58 @@ static void wait_gp(void)
NULL, NULL, 0);
}
+/*
+ * Note: urcu_adaptative_wake_up needs "value" to stay allocated
+ * throughout its execution. In this scheme, the waiter owns the futex
+ * memory, and we only allow it to free this memory when it receives the
+ * AWAKE_TEARDOWN flag.
+ */
+static void urcu_adaptative_wake_up(int32_t *value)
+{
+ cmm_smp_mb();
+ assert(uatomic_read(value) == AWAKE_WAITING);
+ uatomic_set(value, AWAKE_WAKEUP);
+ if (!(uatomic_read(value) & AWAKE_AWAKENED))
+ futex_noasync(value, FUTEX_WAKE, 1, NULL, NULL, 0);
+ /* Allow teardown of "value" memory. */
+ uatomic_or(value, AWAKE_TEARDOWN);
+}
+
+/*
+ * Caller must initialize "value" to AWAKE_WAITING before passing its
+ * memory to waker thread.
+ */
+static void urcu_adaptative_busy_wait(int32_t *value)
+{
+ unsigned int i;
+
+ /* Load and test condition before read futex */
+ cmm_smp_rmb();
+ for (i = 0; i < RCU_AWAKE_ATTEMPTS; i++) {
+ if (uatomic_read(value) != AWAKE_WAITING)
+ goto skip_futex_wait;
+ caa_cpu_relax();
+ }
+ futex_noasync(value, FUTEX_WAIT, AWAKE_WAITING, NULL, NULL, 0);
+skip_futex_wait:
+
+ /* Tell waker thread than we are awakened. */
+ uatomic_or(value, AWAKE_AWAKENED);
+
+ /*
+ * Wait until waker thread lets us know it's ok to tear down
+ * memory allocated for value.
+ */
+ for (i = 0; i < RCU_AWAKE_ATTEMPTS; i++) {
+ if (uatomic_read(value) & AWAKE_TEARDOWN)
+ break;
+ caa_cpu_relax();
+ }
+ while (!(uatomic_read(value) & AWAKE_TEARDOWN))
+ poll(NULL, 0, 10);
+ assert(uatomic_read(value) & AWAKE_TEARDOWN);
+}
+
static void wait_for_readers(struct cds_list_head *input_readers,
struct cds_list_head *cur_snap_readers,
struct cds_list_head *qsreaders)
@@ -198,6 +280,9 @@ void synchronize_rcu(void)
CDS_LIST_HEAD(cur_snap_readers);
CDS_LIST_HEAD(qsreaders);
unsigned long was_online;
+ struct gp_waiters_thread gp_waiters_thread;
+ struct cds_wfs_head *gp_waiters_head;
+ struct cds_wfs_node *waiters_iter, *waiters_iter_n;
was_online = URCU_TLS(rcu_reader).ctr;
@@ -214,8 +299,26 @@ void synchronize_rcu(void)
else
cmm_smp_mb();
+ /*
+ * Add ourself to gp_waiters stack of threads awaiting to wait
+ * for a grace period. Proceed to perform the grace period only
+ * if we are the first thread added into the stack.
+ */
+ cds_wfs_node_init(&gp_waiters_thread.node);
+ gp_waiters_thread.wait_futex = AWAKE_WAITING;
+ if (cds_wfs_push(&gp_waiters, &gp_waiters_node) != 0) {
+ /* Not first in stack: will be awakened by another thread. */
+ urcu_adaptative_busy_wait(&gp_waiters_thread.wait_futex);
+ goto gp_end;
+ }
+
mutex_lock(&rcu_gp_lock);
+ /*
+ * Pop all waiters into our local stack head.
+ */
+ gp_waiters_head = __cds_wfs_pop_all(&gp_waiters);
+
if (cds_list_empty(®istry))
goto out;
@@ -272,6 +375,19 @@ void synchronize_rcu(void)
out:
mutex_unlock(&rcu_gp_lock);
+ /* Wake all waiters in our stack head, excluding ourself. */
+ cds_wfs_for_each_blocking_safe(gp_waiters_head, waiters_iter,
+ waiters_iter_n) {
+ struct gp_waiters_thread *wt;
+
+ wt = caa_container_of(waiters_iter,
+ struct gp_waiters_thread, node);
+ if (wt == &gp_waiters_thread)
+ continue;
+ urcu_adaptative_wake_up(&wt->wait_futex);
+ }
+
+gp_end:
/*
* Finish waiting for reader threads before letting the old ptr being
* freed.
@@ -286,6 +402,9 @@ void synchronize_rcu(void)
{
CDS_LIST_HEAD(qsreaders);
unsigned long was_online;
+ struct gp_waiters_thread gp_waiters_thread;
+ struct cds_wfs_head *gp_waiters_head;
+ struct cds_wfs_node *waiters_iter, *waiters_iter_n;
was_online = URCU_TLS(rcu_reader).ctr;
@@ -299,7 +418,26 @@ void synchronize_rcu(void)
else
cmm_smp_mb();
+ /*
+ * Add ourself to gp_waiters stack of threads awaiting to wait
+ * for a grace period. Proceed to perform the grace period only
+ * if we are the first thread added into the stack.
+ */
+ cds_wfs_node_init(&gp_waiters_thread.node);
+ gp_waiters_thread.wait_futex = AWAKE_WAITING;
+ if (cds_wfs_push(&gp_waiters, &gp_waiters_thread.node) != 0) {
+ /* Not first in stack: will be awakened by another thread. */
+ urcu_adaptative_busy_wait(&gp_waiters_thread.wait_futex);
+ goto gp_end;
+ }
+
mutex_lock(&rcu_gp_lock);
+
+ /*
+ * Pop all waiters into our local stack head.
+ */
+ gp_waiters_head = __cds_wfs_pop_all(&gp_waiters);
+
if (cds_list_empty(®istry))
goto out;
@@ -334,6 +472,19 @@ void synchronize_rcu(void)
out:
mutex_unlock(&rcu_gp_lock);
+ /* Wake all waiters in our stack head, excluding ourself. */
+ cds_wfs_for_each_blocking_safe(gp_waiters_head, waiters_iter,
+ waiters_iter_n) {
+ struct gp_waiters_thread *wt;
+
+ wt = caa_container_of(waiters_iter,
+ struct gp_waiters_thread, node);
+ if (wt == &gp_waiters_thread)
+ continue;
+ urcu_adaptative_wake_up(&wt->wait_futex);
+ }
+
+gp_end:
if (was_online)
rcu_thread_online();
else
--
1.7.10.4
More information about the lttng-dev
mailing list