[PATCH v10 00/25] Speculative page faults
Punit Agrawal
punit.agrawal at arm.com
Thu May 3 00:17:19 AEST 2018
Hi Laurent,
One query below -
Laurent Dufour <ldufour at linux.vnet.ibm.com> writes:
[...]
>
> Ebizzy:
> -------
> The test is counting the number of records per second it can manage, the
> higher is the best. I run it like this 'ebizzy -mTRp'. To get consistent
> result I repeated the test 100 times and measure the average result. The
> number is the record processes per second, the higher is the best.
>
> BASE SPF delta
> 16 CPUs x86 VM 12405.52 91104.52 634.39%
> 80 CPUs P8 node 37880.01 76201.05 101.16%
How do you measure the number of records processed? Is there a specific
version of ebizzy that reports this? I couldn't find a way to get this
information with the ebizzy that's included in ltp.
>
> Here are the performance counter read during a run on a 16 CPUs x86 VM:
> Performance counter stats for './ebizzy -mRTp':
> 860074 faults
> 856866 spf
> 285 pagefault:spf_pte_lock
> 1506 pagefault:spf_vma_changed
> 0 pagefault:spf_vma_noanon
> 73 pagefault:spf_vma_notsup
> 0 pagefault:spf_vma_access
> 0 pagefault:spf_pmd_changed
>
> And the ones captured during a run on a 80 CPUs Power node:
> Performance counter stats for './ebizzy -mRTp':
> 722695 faults
> 699402 spf
> 16048 pagefault:spf_pte_lock
> 6838 pagefault:spf_vma_changed
> 0 pagefault:spf_vma_noanon
> 277 pagefault:spf_vma_notsup
> 0 pagefault:spf_vma_access
> 0 pagefault:spf_pmd_changed
>
> In ebizzy's case most of the page fault were handled in a speculative way,
> leading the ebizzy performance boost.
A trial run showed increased fault handling when SPF is enabled on an
8-core ARM64 system running 4.17-rc3. I am using a port of your x86
patch to enable spf on arm64.
SPF
---
Performance counter stats for './ebizzy -vvvmTRp':
1,322,736 faults
1,299,241 software/config=11/
10.005348034 seconds time elapsed
No SPF
-----
Performance counter stats for './ebizzy -vvvmTRp':
708,916 faults
0 software/config=11/
10.005807432 seconds time elapsed
Thanks,
Punit
[...]
More information about the Linuxppc-dev
mailing list