NUMA balancing

NUMA balancing impact on common benchmarks

NUMA balancing can lead to performance degradation on NUMA-based arm64 systems when tasks migrate,
and their memory accesses now suffer additional latency.

Platform

System	Information
Architecture	aarch64
Processor version	Kunpeng 920-6426
CPUs	128
NUMA nodes	4
Kernel release	5.6.0+
Node name	ARMv2-3

Test results

PerfBenchSchedPipe

perf bench -f simple sched pipe  

Test	Result
numa_balancing-ON	10.012 (usecs/op)
numa_balancing-OFF	10.509 (usecs/op)

PerfBenchSchedMessaging

perf bench -f simple sched messaging -l 10000  

Test	Result
numa_balancing-ON	6.417 (Sec)
numa_balancing-OFF	6.494 (Sec)

PerfBenchMemMemset

perf bench -f simple  mem memset -s 4GB -l 5 -f default  

Test	Result
numa_balancing-ON	17.438783330964565 (GB/sec)
numa_balancing-OFF	17.63163114627642 (GB/sec)

PerfBenchFutexWake

perf bench -f simple futex wake -s -t 1024 -w 1  

Test	Result
numa_balancing-ON	9.2742 (ms)
numa_balancing-OFF	9.2178 (ms)

SysBenchCpu

sysbench cpu --time=10 --threads=64 --cpu-max-prime=10000 run  

Test	Result
numa_balancing-ON	214960.28 (Events/sec)
numa_balancing-OFF	214965.55 (Events/sec)

SysBenchMemory

sysbench memory --memory-access-mode=rnd --threads=64 run  

Test	Result
numa_balancing-ON	1645 (MB/s)
numa_balancing-OFF	1959 (MB/s)

SysBenchThreads

sysbench threads --threads=64 run  

Test	Result
numa_balancing-ON	4604 (Events/sec)
numa_balancing-OFF	5390 (Events/sec)

SysBenchMutex

sysbench mutex --mutex-num=1 --threads=512 run  

Test	Result
numa_balancing-ON	33.2165 (Sec)
numa_balancing-OFF	32.1088 (Sec)