Re: [PATCH 00/16] Remove hash page table slot tracking from linux PTE

Paul Mackerras [off-list ref] writes:

On Fri, Oct 27, 2017 at 10:57:13AM +0530, Aneesh Kumar K.V wrote:

quoted

On 10/27/2017 10:04 AM, Paul Mackerras wrote:

quoted

How do we interpret these numbers?  Are they times, or speed?  Is
larger better or worse?

Sorry for not including the details. They are time in seconds. Test case is
a modified mmap_bench included in powerpc/selftest.

quoted

Can you give us the mean and standard deviation for each set of 5
please?

powernv without patch
median= 51.432255
stdev = 0.370835

with patch
median = 50.739922
stdev = 0.06419662

pseries without patch
median = 116.617884
stdev = 3.04531023

with patch no hcall
median = 119.42494
stdev = 0.85874552

with patch and hcall
median = 117.735808
stdev = 2.7624151

So on powernv, the patch set *improves* performance by about 1.3%
(almost 2 standard deviations).  Do we know why that is?

I haven't looked at that closely. I was considering it within runtime
variance (no impact with patch series). I will get perf record collected
and will see if that points to any details.

On pseries, performance is about 2.4% worse without new hcalls, but
that is less than 1 standard deviation.  With new hcalls, performance
is 0.95% worse, only a third of a standard deviation.  I think we need
to do more measurements to try to get a more accurate picture here.

Were the pseries numbers done on KVM or PowerVM?  Could you do a set
of measurements on the other one too please?  (I assume the numbers
with the new hcall were done on KVM, and can't be done on PowerVM.)

The above pseries numbers were collected on KVM.

PowerVM numbers on a different machine:
Without patch
31.194165                           
31.372913                           
31.253494
31.416198
31.199180
MEDIAN = 31.253494
STDEV = 0.1018900

With patch series
31.538281
31.385996
31.492737
31.452514
31.259461
MEDIAN = 31.452514
STDEV  = 0.108511

-aneesh