On 2019/7/3 19:58, Jan Glauber wrote:
> Hi Alex,
> I've tried this series on arm64 (ThunderX2 with up to SMT=4  and 224 CPUs)
> with the borderline testcase of accessing a single file from all
> threads. With that
> testcase the qspinlock slowpath is the top spot in the kernel.
> 
> The results look really promising:
> 
> CPUs    normal    numa-qspinlocks
> ---------------------------------------------
> 56        149.41          73.90
> 224      576.95          290.31
> 
> Also frontend-stalls are reduced to 50% and interconnect traffic is
> greatly reduced.
> Tested-by: Jan Glauber <jglau...@marvell.com>

Tested this patchset on Kunpeng920 ARM64 server (96 cores,
4 NUMA nodes), and with the same test case from Jan, I can
see 150%+ boost! (Need to add a patch below [1].)

For the real workload such as Nginx I can see about 10%
performance improvement as well.

Tested-by: Hanjun Guo <guohan...@huawei.com>

Please cc me for new versions and I'm willing to test it.

Thanks
Hanjun

[1]
diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index 657bbc5..72c1346 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -792,6 +792,20 @@ config NODES_SHIFT
          Specify the maximum number of NUMA Nodes available on the target
          system.  Increases memory reserved to accommodate various tables.

+config NUMA_AWARE_SPINLOCKS
+ bool "Numa-aware spinlocks"
+ depends on NUMA
+ default y
+ help
+   Introduce NUMA (Non Uniform Memory Access) awareness into
+   the slow path of spinlocks.
+
+   The kernel will try to keep the lock on the same node,
+   thus reducing the number of remote cache misses, while
+   trading some of the short term fairness for better performance.
+
+   Say N if you want absolute first come first serve fairness.
+
 config USE_PERCPU_NUMA_NODE_ID
        def_bool y
        depends on NUMA
diff --git a/kernel/locking/qspinlock_cna.h b/kernel/locking/qspinlock_cna.h
index 2994167..be5dd44 100644
--- a/kernel/locking/qspinlock_cna.h
+++ b/kernel/locking/qspinlock_cna.h
@@ -4,7 +4,7 @@
 #endif

 #include <linux/random.h>
-
+#include <linux/topology.h>
 /*
  * Implement a NUMA-aware version of MCS (aka CNA, or compact NUMA-aware lock).
  *
@@ -170,7 +170,7 @@ static __always_inline void cna_init_node(struct 
mcs_spinlock *node, int cpuid,
                                          u32 tail)
 {
        if (decode_numa_node(node->node_and_count) == -1)
-           store_numa_node(node, numa_cpu_node(cpuid));
+         store_numa_node(node, cpu_to_node(cpuid));
        node->encoded_tail = tail;
 }

Reply via email to