[PATCH v12 0/5] Add NUMA-awareness to qspinlock
From: Alex Kogan
Date: Tue Nov 17 2020 - 18:21:24 EST
Minor change from v11:
----------------
Fix documentation issue, as requested by Randy Dunlap and Longman.
The rest of the series is unchanged.
Summary
-------
Lock throughput can be increased by handing a lock to a waiter on the
same NUMA node as the lock holder, provided care is taken to avoid
starvation of waiters on other NUMA nodes. This patch introduces CNA
(compact NUMA-aware lock) as the slow path for qspinlock. It is
enabled through a configuration option (NUMA_AWARE_SPINLOCKS).
CNA is a NUMA-aware version of the MCS lock. Spinning threads are
organized in two queues, a primary queue for threads running on the same
node as the current lock holder, and a secondary queue for threads
running on other nodes. Threads store the ID of the node on which
they are running in their queue nodes. After acquiring the MCS lock and
before acquiring the spinlock, the MCS lock holder checks whether the next
waiter in the primary queue (if exists) is running on the same NUMA node.
If it is not, that waiter is detached from the main queue and moved into
the tail of the secondary queue. This way, we gradually filter the primary
queue, leaving only waiters running on the same preferred NUMA node. Note
that certain priortized waiters (e.g., in irq and nmi contexts) are
excluded from being moved to the secondary queue. We change the NUMA node
preference after a waiter at the head of the secondary queue spins for a
certain amount of time. We do that by flushing the secondary queue into
the head of the primary queue, effectively changing the preference to the
NUMA node of the waiter at the head of the secondary queue at the time of
the flush.
More details are available at https://arxiv.org/abs/1810.05600.
We have done some performance evaluation with the locktorture module
as well as with several benchmarks from the will-it-scale repo.
The following locktorture results are from an Oracle X5-4 server
(four Intel Xeon E7-8895 v3 @ 2.60GHz sockets with 18 hyperthreaded
cores each). Each number represents an average (over 25 runs) of the
total number of ops (x10^7) reported at the end of each run. The
standard deviation is also reported in (), and in general is about 3%
from the average. The 'stock' kernel is v5.10.0-rc2,
commit 4ef8451b3326, compiled in the default configuration.
'patch-CNA' is the modified kernel with NUMA_AWARE_SPINLOCKS set;
the speedup is calculated dividing 'patch-CNA' by 'stock'.
#thr stock patch-CNA speedup (patch-CNA/stock)
1 2.672 (0.109) 2.679 (0.110) 1.002
2 3.247 (0.161) 3.276 (0.136) 1.009
4 4.376 (0.139) 4.820 (0.181) 1.101
8 5.134 (0.137) 7.125 (0.164) 1.388
16 5.875 (0.113) 8.903 (0.209) 1.515
32 6.298 (0.105) 9.911 (0.254) 1.574
36 6.439 (0.125) 9.972 (0.226) 1.549
72 6.249 (0.109) 10.375 (0.209) 1.660
108 6.082 (0.063) 10.511 (0.190) 1.728
142 5.822 (0.058) 10.448 (0.177) 1.795
The following tables contain throughput results (ops/us) from the same
setup for will-it-scale/open1_threads:
#thr stock patch-CNA speedup (patch-CNA/stock)
1 0.508 (0.002) 0.507 (0.002) 0.999
2 0.759 (0.016) 0.767 (0.019) 1.011
4 1.397 (0.032) 1.411 (0.029) 1.011
8 1.694 (0.079) 1.656 (0.120) 0.978
16 1.867 (0.107) 1.809 (0.121) 0.969
32 1.006 (0.056) 1.752 (0.093) 1.742
36 0.934 (0.099) 1.724 (0.064) 1.846
72 0.804 (0.045) 1.632 (0.073) 2.030
108 0.828 (0.036) 1.690 (0.065) 2.041
142 0.784 (0.035) 1.701 (0.074) 2.168
and will-it-scale/lock2_threads:
#thr stock patch-CNA speedup (patch-CNA/stock)
1 1.590 (0.004) 1.603 (0.005) 1.008
2 2.802 (0.057) 2.802 (0.063) 1.000
4 5.478 (0.144) 5.396 (0.299) 0.985
8 4.166 (0.304) 4.131 (0.402) 0.992
16 4.147 (0.137) 3.983 (0.173) 0.961
32 2.492 (0.067) 3.888 (0.125) 1.560
36 2.471 (0.094) 3.908 (0.112) 1.581
72 1.886 (0.092) 3.926 (0.106) 2.081
108 1.883 (0.101) 3.935 (0.096) 2.089
142 1.801 (0.112) 3.907 (0.111) 2.169
Our evaluation shows that CNA also improves performance of user
applications that have hot pthread mutexes. Those mutexes are
blocking, and waiting threads park and unpark via the futex
mechanism in the kernel. Given that kernel futex chains, which
are hashed by the mutex address, are each protected by a
chain-specific spin lock, the contention on a user-mode mutex
translates into contention on a kernel level spinlock.
Here are the throughput results (ops/us) for the leveldb ‘readrandom’
benchmark:
#thr stock patch-CNA speedup (patch-CNA/stock)
1 0.532 (0.013) 0.533 (0.023) 1.004
2 0.836 (0.050) 0.843 (0.031) 1.009
4 1.039 (0.163) 1.087 (0.151) 1.046
8 1.095 (0.181) 1.178 (0.165) 1.076
16 1.002 (0.144) 1.196 (0.019) 1.194
32 0.726 (0.034) 1.163 (0.026) 1.601
36 0.691 (0.030) 1.163 (0.020) 1.683
72 0.627 (0.014) 1.136 (0.022) 1.812
108 0.613 (0.014) 1.143 (0.023) 1.865
142 0.610 (0.014) 1.120 (0.018) 1.838
Further comments are welcome and appreciated.
Alex Kogan (5):
locking/qspinlock: Rename mcs lock/unlock macros and make them more
generic
locking/qspinlock: Refactor the qspinlock slow path
locking/qspinlock: Introduce CNA into the slow path of qspinlock
locking/qspinlock: Introduce starvation avoidance into CNA
locking/qspinlock: Avoid moving certain threads between waiting queues
in CNA
Documentation/admin-guide/kernel-parameters.txt | 19 ++
arch/arm/include/asm/mcs_spinlock.h | 6 +-
arch/x86/Kconfig | 20 ++
arch/x86/include/asm/qspinlock.h | 4 +
arch/x86/kernel/alternative.c | 4 +
include/asm-generic/mcs_spinlock.h | 4 +-
kernel/locking/mcs_spinlock.h | 20 +-
kernel/locking/qspinlock.c | 82 ++++-
kernel/locking/qspinlock_cna.h | 421 ++++++++++++++++++++++++
kernel/locking/qspinlock_paravirt.h | 2 +-
10 files changed, 559 insertions(+), 23 deletions(-)
create mode 100644 kernel/locking/qspinlock_cna.h
--
2.7.4