DragonFlyBSD To See Better Low-Level Lock Performance When Heavily Contested
Software running on DragonFlyBSD and making use of pthreads is set to see better performance around low-level locks when heavily contested.
This commit has the details on the change by DragonFlyBSD founder Matthew Dillon. But long story short pthreads-using software should benefit from this low-level lock performance improvement.
As his test case, Dillon found that when running QEMU with the NVMM hypervisor that this change dropped the DragonFlyBSD build-all time from 9:10 to 8:20. This is just one example and an important one given the widespread use of QEMU. QEMU with NVMM turns out to have significant contention issues with high CPU core counts while software not seeing heavy lock contention is unlikely to see any real difference from these optimizations.
This commit has the details on the change by DragonFlyBSD founder Matthew Dillon. But long story short pthreads-using software should benefit from this low-level lock performance improvement.
As his test case, Dillon found that when running QEMU with the NVMM hypervisor that this change dropped the DragonFlyBSD build-all time from 9:10 to 8:20. This is just one example and an important one given the widespread use of QEMU. QEMU with NVMM turns out to have significant contention issues with high CPU core counts while software not seeing heavy lock contention is unlikely to see any real difference from these optimizations.
3 Comments