From ongoing tests of this patch, it still certainly shows to address
most of the Linux 5.11 performance regression previously encountered
when using Schedutil. Additionally, for a number of workloads where not
seeing a regression from 5.10 to 5.11 Git is still showing even better
performance with this patch. The power monitoring on the AMD EPYC server
is showing higher power spikes but the average power consumption rate is
roughly comparable to that of Linux 5.11 Git, which is higher than 5.10
by just about 3%.
So this patch still seems to be working out well and indeed taking care
of some wide losses seen otherwise on Linux 5.11 when using Schedutil on
AMD Zen2/Zen3. Still have some other tests running but so far no
unexpected results.
Michael
AMD EPYC 7F72 2P
On an EPYC 7F72 2P server[1] across 147 test cases I am finding the
patched Linux 5.11 kernel to be just over 1% faster than 5.10 stable
compared to the unpatched 5.11 Git being just behind 5.10. For the
workloads on that server where Linux 5.11 is slower with Schedutil, the
patch indeed is largely addressing that regression and also providing
other improvements.
During that testing, the amd_energy interface was monitored. Linux 5.11
with Schedutil AMD freq invariance did show on average 10 Watts (~3.7%)
higher power consumption on average than Linux 5.10 with Schedutil. But
with this patch, that average is still basically the same. The peak
power consumption during any of the tests was higher at 530~549 Watts
compared to 501 Watts with Linux 5.10. Overall the performance is
looking good but given amd_energy still not working for consumer models,
I don't have much power data to share at the moment.
Ryzen 9 5950X
Expanding on the prior testing with the 5950X, I ran some follow-up
tests[2]. Of 221 test cases there, the current Linux 5.11 Git
performance came around 2% slower on a geo mean basis than Linux 5.10
while the patched performance pulls it to ~2.5% faster than 5.10. There
still are some cases where Linux 5.10 is faster, but overall at least
the patched Linux 5.11 performance doesn't show nearly as many
regressions. In a number of test cases, the Linux 5.11 patched
performance is outright better than Linux 5.10 even where 5.11
(un-patched) hadn't regressed or by that much.
Ryzen 5 4500U
For something at the lower end of the spectrum I also ran a number of
tests on a Ryzen 5 4500U notebook[3]. Linux 5.11 (unpatched) doesn't see
as many regressions as on the larger systems but still the patched
performance did help in a number of tests, particularly around
graphics/gaming. In the heavier multi-core core tests are still a number
of cases where Linux 5.10 is faster than patched/unpatched 5.11. The
patched kernel in those 90 tests came out to being about 4% faster than
5.10.
(Result highlights below, results with little change set to hidden by
default.)
[1]
https://openbenchmarking.org/result/2101248-HA-AMDEPYC7F52&grs&hlc=1&hnr=1&hlc=1
[2] https://openbenchmarking.org/result/2101242-HA-RYZEN959530&grs&hlc=1
[3]
https://openbenchmarking.org/result/2101232-PTS-RENOIRLI89&grs&hnr=1&hlc=1
On 1/22/21 2:40 PM, Giovanni Gherdovich wrote:
Michael Larabel from Phoronix.com graciously tested v1, see results at:
AMD EPYC 7702 -
https://openbenchmarking.org/result/2101210-PTS-LINUX51178
AMD Ryzen 9 5950X -
https://openbenchmarking.org/result/2101212-HA-RYZEN959566
The reported regression is recovered, and some workloads even report an
improvement over the v5.10 results.
Thanks Michael for the feedback!
v1 at https://lore.kernel.org/lkml/20210121003223.20257-1-ggherdovich@xxxxxxx/
Changes wrt v1:
- move code around so that it builds for non-x86 architectures too
Giovanni Gherdovich (1):
x86,sched: On AMD EPYC set freq_max = max_boost in schedutil invariant
formula
drivers/cpufreq/acpi-cpufreq.c | 64 +++++++++++++++++++++++++++++++-
drivers/cpufreq/cpufreq.c | 3 ++
include/linux/cpufreq.h | 5 +++
kernel/sched/cpufreq_schedutil.c | 8 +++-
4 files changed, 76 insertions(+), 4 deletions(-)
