[v2,0/5] Add a dirty logging performance test

Message

Ben Gardon Nov. 3, 2020, 11:49 p.m. UTC

Currently KVM lacks a simple, userspace agnostic, performance benchmark for
dirty logging. Such a benchmark will be beneficial for ensuring that dirty
logging performance does not regress, and to give a common baseline for
validating performance improvements. The dirty log perf test introduced in
this series builds on aspects of the existing demand paging perf test and
provides time-based performance metrics for enabling and disabling dirty
logging, getting the dirty log, and dirtying memory.

While the test currently only has a build target for x86, I expect it will
work on, or be easily modified to support other architectures.

This series was tested by running the following invocations on an Intel
Skylake machine after apply all commits in the series:
dirty_log_perf_test -b 20m -i 100 -v 64
dirty_log_perf_test -b 20g -i 5 -v 4
dirty_log_perf_test -b 4g -i 5 -v 32
demand_paging_test -b 20m -v 64
demand_paging_test -b 20g -v 4
demand_paging_test -b 4g -v 32
All behaved as expected.

v1 -> v2 changes:
(in response to comments from Peter Xu)
- Removed pr_debugs from main test thread while waiting on vCPUs to reduce
  log spam
- Fixed a bug in iteration counting that caused the population stage to be
  counted as part of the first dirty logging pass
- Fixed a bug in which the test failed to wait for the population stage for all
  but the first vCPU.
- Refactored the common code in perf_test_util.c/h
- Moved testing description to cover letter
- Renamed timespec_diff_now to timespec_elapsed

Ben Gardon (5):
  KVM: selftests: Remove address rounding in guest code
  KVM: selftests: Factor code out of demand_paging_test
  KVM: selftests: Simplify demand_paging_test with timespec_diff_now
  KVM: selftests: Add wrfract to common guest code
  KVM: selftests: Introduce the dirty log perf test

 tools/testing/selftests/kvm/.gitignore        |   1 +
 tools/testing/selftests/kvm/Makefile          |   7 +-
 .../selftests/kvm/demand_paging_test.c        | 231 ++---------
 .../selftests/kvm/dirty_log_perf_test.c       | 381 ++++++++++++++++++
 .../selftests/kvm/include/perf_test_util.h    |  51 +++
 .../testing/selftests/kvm/include/test_util.h |   2 +
 .../selftests/kvm/lib/perf_test_util.c        | 166 ++++++++
 tools/testing/selftests/kvm/lib/test_util.c   |  22 +-
 8 files changed, 661 insertions(+), 200 deletions(-)
 create mode 100644 tools/testing/selftests/kvm/dirty_log_perf_test.c
 create mode 100644 tools/testing/selftests/kvm/include/perf_test_util.h
 create mode 100644 tools/testing/selftests/kvm/lib/perf_test_util.c

Comments

Andrew Jones Nov. 11, 2020, 12:34 p.m. UTC | #1

On Tue, Nov 03, 2020 at 03:49:47PM -0800, Ben Gardon wrote:
> Currently KVM lacks a simple, userspace agnostic, performance benchmark for
> dirty logging. Such a benchmark will be beneficial for ensuring that dirty
> logging performance does not regress, and to give a common baseline for
> validating performance improvements. The dirty log perf test introduced in
> this series builds on aspects of the existing demand paging perf test and
> provides time-based performance metrics for enabling and disabling dirty
> logging, getting the dirty log, and dirtying memory.
> 
> While the test currently only has a build target for x86, I expect it will
> work on, or be easily modified to support other architectures.
> 
> This series was tested by running the following invocations on an Intel
> Skylake machine after apply all commits in the series:
> dirty_log_perf_test -b 20m -i 100 -v 64
> dirty_log_perf_test -b 20g -i 5 -v 4
> dirty_log_perf_test -b 4g -i 5 -v 32
> demand_paging_test -b 20m -v 64
> demand_paging_test -b 20g -v 4
> demand_paging_test -b 4g -v 32
> All behaved as expected.
> 
> v1 -> v2 changes:

Considering v1 got applied,

> (in response to comments from Peter Xu)
> - Removed pr_debugs from main test thread while waiting on vCPUs to reduce
>   log spam

I didn't look at this. Maybe you and Peter can decide if a pr_debug
cleanup patch needs to be sent.

> - Fixed a bug in iteration counting that caused the population stage to be
>   counted as part of the first dirty logging pass
> - Fixed a bug in which the test failed to wait for the population stage for all
>   but the first vCPU.

I didn't try to confirm that these fixes were in. I do see that Paolo
added 6d6a18fdde8b ("KVM: selftests: allow two iterations of
dirty_log_perf_test"), which sounds like it might be fixing the same thing
as your first "Fixed" bullet above. Anyway, you may want to check the 
current code to see if any additional fixes should be sent.

> - Refactored the common code in perf_test_util.c/h

I did this part in my "Cleanups, take 2" series

> - Moved testing description to cover letter
> - Renamed timespec_diff_now to timespec_elapsed

These two would have been nice, but oh well...

Thanks,
drew