On 7/23/2021 2:11 PM, Wanpeng Li wrote:
On Fri, 23 Jul 2021 at 13:41, Zeng Guang <guang.zeng@xxxxxxxxx> wrote: -------------------------------------- IPI microbenchmark: (https://lore.kernel.org/kvm/20171219085010.4081-1-ynorov@xxxxxxxxxxxxxxxxxx) 2 vCPUs, 1:1 pin vCPU to pCPU, guest VM runs with idle=poll, x2APIC mode Improve the performance for unicast ipi is as expected, however, I wonder whether the broadcast performance is worse than PV IPIs/Thomas's IPI shorthands(IPI shorthands are supported by upstream linux apic/x2apic driver). The hardware acceleration is not always outstanding on AMD(https://lore.kernel.org/kvm/CANRm+Cx597FNRUCyVz1D=B6Vs2GX3Sw57X7Muk+yMpi_hb+v1w@xxxxxxxxxxxxxx/), how about your Intel guys? Please try a big VM at least 96 vCPUs as below or more bigger.
Intel IPIv target to accelerate unicast ipi process, not benefit to broadcast performance.
As to IPI benchmark, it's not big different to test with large or small scale of vCPUs. In essential, Normal IPI test try to send ipi to any other online CPU in sequence. The cost on IPI process itself should be similar.
Result with IPIv enabled: Dry-run: 0, 272798 ns Self-IPI: 5094123, 11114037 ns Normal IPI: 131697087, 173321200 ns Broadcast IPI: 0, 155649075 ns Broadcast lock: 0, 161518031 ns Result with IPIv disabled: Dry-run: 0, 272766 ns Self-IPI: 5091788, 11123699 ns Normal IPI: 145215772, 174558920 ns Broadcast IPI: 0, 175785384 ns Broadcast lock: 0, 149076195 ns As IPIv can benefit unicast IPI to other CPU, Noraml IPI test case gain about 9.73% time saving on average out of 15 test runs when IPIv is enabled. w/o IPIv w/ IPIv Normal IPI: 145944306.6 ns 131742993.1 ns %Reduction -9.73% -------------------------------------- hackbench: 8 vCPUs, guest VM free run, x2APIC mode ./hackbench -p -l 100000 w/o IPIv w/ IPIv Time: 91.887 74.605 %Reduction: -18.808% 96 vCPUs, guest VM free run, x2APIC mode ./hackbench -p -l 1000000 w/o IPIv w/ IPIv Time: 287.504 235.185 %Reduction: -18.198%Good to know this. Wanpeng
