This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
[PATCH v3 3/3] Mutex: Replace trylock by read only while spinning
- From: Kemi Wang <kemi dot wang at intel dot com>
- To: Adhemerval Zanella <adhemerval dot zanella at linaro dot org>, Florian Weimer <fweimer at redhat dot com>, Rical Jason <rj at 2c3t dot io>, Carlos Donell <carlos at redhat dot com>, Glibc alpha <libc-alpha at sourceware dot org>
- Cc: Dave Hansen <dave dot hansen at linux dot intel dot com>, Tim Chen <tim dot c dot chen at intel dot com>, Andi Kleen <andi dot kleen at intel dot com>, Ying Huang <ying dot huang at intel dot com>, Aaron Lu <aaron dot lu at intel dot com>, Lu Aubrey <aubrey dot li at intel dot com>, Kemi Wang <kemi dot wang at intel dot com>
- Date: Wed, 23 May 2018 17:22:34 +0800
- Subject: [PATCH v3 3/3] Mutex: Replace trylock by read only while spinning
- References: <1527067354-13333-1-git-send-email-kemi.wang@intel.com>
The pthread adaptive spin mutex spins on the lock for a while before
calling into the kernel to block. But, in the current implementation of
spinning, the spinners go straight back to LLL_MUTEX_TRYLOCK(cmpxchg) when
the lock is contended, it is not a good idea on many targets as that will
force expensive memory synchronization among processors and penalize other
running threads. For example, it constantly floods the system with "read
for ownership" requests, which are much more expensive to process than a
single read. Thus, we only use MO read until we observe the lock to not be
acquired anymore, as suggeusted by Andi Kleen.
Performance impact:
Significant mutex performance improvement is not expected for this patch,
though, it probably bring some benefit for the scenarios with severe lock
contention on many architectures, the whole system performance can benefit
from this modification because a number of unnecessary "read for ownership"
requests which stress the cache system via read and invalidate broadcast
are eliminated during spinning.
Meanwhile, it may have some tiny performance regression on the lock holder
transformation for the case of lock acquisition via spinning gets, because
the lock state is checked before acquiring the lock via trylock. In the
worst case, the extra latency of read and pause is added except that of
trylock when lock is available.
Similar mechanism has been implemented for pthread spin lock.
Test machine:
2-sockets Skylake platform, 112 cores with 62G RAM
Test case: mutex-adaptive-thread (Contended pthread adaptive spin mutex
with global update)
Usage: make bench BENCHSET=mutex-adaptive-thread
Test result:
+----------------+-----------------+-----------------+------------+
| Configuration | Base | Head | % Change |
| | Total iteration | Total iteration | base->head |
+----------------+-----------------+-----------------+------------+
| | Critical section size: 1x |
+----------------+------------------------------------------------+
|1 thread | 2.76681e+07 | 2.7965e+07 | +1.1% |
|2 threads | 3.29905e+07 | 3.55279e+07 | +7.7% |
|3 threads | 4.38102e+07 | 3.98567e+07 | -9.0% |
|4 threads | 1.72172e+07 | 2.09498e+07 | +21.7% |
|28 threads | 1.03732e+07 | 1.05133e+07 | +1.4% |
|56 threads | 1.06308e+07 | 5.06522e+07 | +14.6% |
|112 threads | 8.55177e+06 | 1.02954e+07 | +20.4% |
+----------------+------------------------------------------------+
| | Critical section size: 10x |
+----------------+------------------------------------------------+
|1 thread | 1.57006e+07 | 1.54727e+07 | -1.5% |
|2 threads | 1.8044e+07 | 1.75601e+07 | -2.7% |
|3 threads | 1.35634e+07 | 1.46384e+07 | +7.9% |
|4 threads | 1.21257e+07 | 1.32046e+07 | +8.9% |
|28 threads | 8.09593e+06 | 1.02713e+07 | +26.9% |
|56 threads | 9.09907e+06 | 4.16203e+07 | +16.4% |
|112 threads | 7.09731e+06 | 8.62406e+06 | +21.5% |
+----------------+------------------------------------------------+
| | Critical section size: 100x |
+----------------+------------------------------------------------+
|1 thread | 2.87116e+06 | 2.89188e+06 | +0.7% |
|2 threads | 2.23409e+06 | 2.24216e+06 | +0.4% |
|3 threads | 2.29888e+06 | 2.29964e+06 | +0.0% |
|4 threads | 2.26898e+06 | 2.21394e+06 | -2.4% |
|28 threads | 1.03228e+06 | 1.0051e+06 | -2.6% |
|56 threads | 1.02953 +06 | 1.6344e+07 | -2.3% |
|112 threads | 1.01615e+06 | 1.00134e+06 | -1.5% |
+----------------+------------------------------------------------+
| | Critical section size: 1000x |
+----------------+------------------------------------------------+
|1 thread | 316392 | 315635 | -0.2% |
|2 threads | 302806 | 303469 | +0.2% |
|3 threads | 298506 | 294281 | -1.4% |
|4 threads | 292037 | 289945 | -0.7% |
|28 threads | 155188 | 155250 | +0.0% |
|56 threads | 190657 | 183106 | -4.0% |
|112 threads | 210818 | 220342 | +4.5% |
+----------------+-----------------+-----------------+------------+
* nptl/pthread_mutex_lock.c: Optimize adaptive spin mutex
ChangLog:
V2->V3:
a) Drop the idea of blocking spinners if fail to acquire a lock, since
this idea would not be an universal winner. E.g. several threads
contend for a lock which protects a small critical section, thus,
probably any thread can acquire the lock via spinning.
b) Fix the format issue AFAIC
V1->V2: fix format issue
Suggested-by: Andi Kleen <andi.kleen@intel.com>
Signed-off-by: Kemi Wang <kemi.wang@intel.com>
---
ChangeLog | 5 +++++
nptl/pthread_mutex_lock.c | 41 +++++++++++++++++++++++++----------------
2 files changed, 30 insertions(+), 16 deletions(-)
diff --git a/ChangeLog b/ChangeLog
index e2991e9..3bafb0e 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,10 @@
2018-05-23 Kemi Wang <kemi.wang@intel.com>
+ * nptl/pthread_mutex_lock.c: Replace trylock by read only while
+ spinning.
+
+2018-05-23 Kemi Wang <kemi.wang@intel.com>
+
* benchtests/bench-mutex-adaptive-thread.c: Microbenchmark for adaptive
spin mutex.
* benchmark/Makefile: Add adaptive spin mutex benchmark.
diff --git a/nptl/pthread_mutex_lock.c b/nptl/pthread_mutex_lock.c
index 1519c14..7ce50f6 100644
--- a/nptl/pthread_mutex_lock.c
+++ b/nptl/pthread_mutex_lock.c
@@ -26,6 +26,7 @@
#include <atomic.h>
#include <lowlevellock.h>
#include <stap-probe.h>
+#include <pthread_mutex_conf.h>
#ifndef lll_lock_elision
#define lll_lock_elision(lock, try_lock, private) ({ \
@@ -123,22 +124,30 @@ __pthread_mutex_lock (pthread_mutex_t *mutex)
if (LLL_MUTEX_TRYLOCK (mutex) != 0)
{
- int cnt = 0;
- int max_cnt = MIN (MAX_ADAPTIVE_COUNT,
- mutex->__data.__spins * 2 + 10);
- do
- {
- if (cnt++ >= max_cnt)
- {
- LLL_MUTEX_LOCK (mutex);
- break;
- }
- atomic_spin_nop ();
- }
- while (LLL_MUTEX_TRYLOCK (mutex) != 0);
-
- mutex->__data.__spins += (cnt - mutex->__data.__spins) / 8;
- }
+ int val = 0;
+ int cnt = 0;
+ int max_cnt = MIN (__mutex_aconf.spin_count,
+ mutex->__data.__spins * 2 + 10);
+
+ do
+ {
+ if (cnt >= max_cnt)
+ {
+ LLL_MUTEX_LOCK (mutex);
+ break;
+ }
+ /* Read only while spinning unless lock is available. */
+ do
+ {
+ atomic_spin_nop ();
+ val = atomic_load_relaxed (&mutex->__data.__lock);
+ }
+ while (val != 0 && ++cnt < max_cnt);
+ }
+ while (LLL_MUTEX_TRYLOCK (mutex) != 0);
+
+ mutex->__data.__spins += (cnt - mutex->__data.__spins) / 8;
+ }
assert (mutex->__data.__owner == 0);
}
else
--
2.7.4