We encountered an issue related to NR_CPUMASK_BITS on ARM64. It turned out CPUMASK_OFFSTACK was not enabled and NR_CPUS is set to 4096 in our aarch64 config, so each cpumask operation is unnecessarily expensive esp. on small VMs. There are decent number of cpumask operations in the scheduler. Many of them are on the hot paths. The cumulative effect is quite siginificant. With CONFIG_CPUMASK_OFFSTACKT set, we saw ~10% gain on ApacheBench and a few percentage on redis in a 32-core ARM64 VM with 5.16 kernel. One may argue why not just set NR_CPUS to smaller values, the thing is, for example, we support VMs with flexible shapes. The customer can specify the number of CPUs ranging from 1 to 160. You cannot just prepare 160 kernel images for each release with one for each shape. And kernel with NR_CPUS=160 doesn't perform well on a 1-CPU VM. It's important that we can set CPUMASK_OFFSTACK=y so that the kernel can dynamically allocate cpumasks based on the actual number of CPUs in the system. The problem I am trying to address in this patch is currently CPUMASK_OFFSTACK depends on DEBUG_PER_CPU_MAPS except for x86 and it doesn't even show up in kconfig menu as well as kconfig without DEBUG_PER_CPU_MAPS=y. We should remove such outdated, unnecessary dependency. Of course, I am open to other ideas. And people who are more familar with this matter can shed a light on why this dependency has been kept for so long. My goal here is to determine if DEBUG_PER_CPU_MAPS is absoultely necessary for CPUMASK_OFFSTACK. Libo Chen (1): lib/Kconfig: remove DEBUG_PER_CPU_MAPS dependency for CPUMASK_OFFSTACK lib/Kconfig | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) -- 2.27.0
