Test virtual interface [NOT FOR REVIEW]

Test virtual interface [NOT FOR REVIEW]

BUG=685733

[kraynov, 2017-01-27 14:47:22]

Description was changed from

==========
Test virtual interface

BUG=685733
==========

to

==========
Test virtual interface [NOT FOR REVIEW]

BUG=685733
==========

[kraynov, 2017-01-27 14:50:08]

kraynov@chromium.org changed reviewers:
+ primiano@chromium.org

[kraynov, 2017-01-27 14:50:09]

Seems to have comparable performance, but usually virtual interface version is
slower.

*RESULT TraceEventPerfTest.Submit_10000_TRACE_EVENT_TEST_via_virtual_calls:
events= 22 ms
*RESULT TraceEventPerfTest.Submit_10000_TRACE_EVENT_TEST_via_direct_calls:
events= 16 ms

*RESULT TraceEventPerfTest.Submit_10000_TRACE_EVENT_TEST_via_virtual_calls:
events= 17 ms
*RESULT TraceEventPerfTest.Submit_10000_TRACE_EVENT_TEST_via_direct_calls:
events= 16 ms

*RESULT TraceEventPerfTest.Submit_10000_TRACE_EVENT_TEST_via_virtual_calls:
events= 21 ms
*RESULT TraceEventPerfTest.Submit_10000_TRACE_EVENT_TEST_via_direct_calls:
events= 16 ms

*RESULT TraceEventPerfTest.Submit_10000_TRACE_EVENT_TEST_via_virtual_calls:
events= 16 ms
*RESULT TraceEventPerfTest.Submit_10000_TRACE_EVENT_TEST_via_direct_calls:
events= 21 ms

*RESULT TraceEventPerfTest.Submit_10000_TRACE_EVENT_TEST_via_virtual_calls:
events= 16 ms
*RESULT TraceEventPerfTest.Submit_10000_TRACE_EVENT_TEST_via_direct_calls:
events= 16 ms

[Primiano Tucci (use gerrit), 2017-01-27 21:26:10]

The example is a bit fake, see the comment in the code. can we make it more
realistic?

Also can you:
1) make sure you build with is_debug=false is_component_build=false
dcheck_always_on=false is_official_build=true

2) run this both on arm and x86_64

3) Get more than 5 samples. Is quite hard to do any science with 5 samples

4) Can you counter-check that actually adding some silly busy loop (or just a
asm volatile("mfence") on x86) makes a difference  in the benchmark?
I would like to have confidence that the benchmark is able to actually spot
regressions :)

https://codereview.chromium.org/2656303003/diff/1/components/tracing/test/tra...
File components/tracing/test/tracing_interface_simulator.cc (right):

https://codereview.chromium.org/2656303003/diff/1/components/tracing/test/tra...
components/tracing/test/tracing_interface_simulator.cc:24:
TRACE_EVENT0(category, name);
this is over-penalizing this case. if you use a trace_event0 you will get
another category check and another jump here.

https://codereview.chromium.org/2656303003/diff/1/components/tracing/test/tra...
components/tracing/test/tracing_interface_simulator.cc:28: if
(::strcmp(category, "test_category") == 0)
base::trace_event::TraceLog::GetCategoryGroupEnabled

https://codereview.chromium.org/2656303003/diff/1/components/tracing/test/tra...
File components/tracing/test/tracing_interface_simulator.h (right):

https://codereview.chromium.org/2656303003/diff/1/components/tracing/test/tra...
components/tracing/test/tracing_interface_simulator.h:29: static const bool*
enabled_ptr = nullptr; \
see comments below

https://codereview.chromium.org/2656303003/diff/1/components/tracing/test/tra...
components/tracing/test/tracing_interface_simulator.h:40: static const bool*
enabled_ptr = nullptr; \
see INTERNAL_TRACE_EVENT_GET_CATEGORY_INFO here. it uses the right set of atomic
ops

https://codereview.chromium.org/2656303003/diff/1/components/tracing/test/tra...
components/tracing/test/tracing_interface_simulator.h:41: if (!enabled_ptr) \
ditto. see INTERNAL_TRACE_EVENT_CATEGORY_GROUP_ENABLED, it uses atomic ops

[kraynov, 2017-02-01 18:49:18]

Based on 20 runs.

Desktop:
  Median virtual (events): 83 ms
  Median direct (events): 80 ms
  Tradeoff: 3.75%

Android (Nexus 5):
  Median virtual (events): 258 ms
  Median direct (events): 251 ms
  Tradeoff: 2.79%

So in this test setting we have rougly 3% tradeoff incurred by using virtual
dispatch.

[Primiano Tucci (use gerrit), 2017-02-01 19:17:18]

On 2017/02/01 18:49:18, kraynov wrote:
> Based on 20 runs.
> 
> Desktop:
>   Median virtual (events): 83 ms
>   Median direct (events): 80 ms
>   Tradeoff: 3.75%
> 
> Android (Nexus 5):
>   Median virtual (events): 258 ms
>   Median direct (events): 251 ms
>   Tradeoff: 2.79%
> 
> So in this test setting we have rougly 3% tradeoff incurred by using virtual
> dispatch.

Are these values for how many runs? 50k?
Does this mean that the cost today of a TRACE_EVENT0 is
Android 258/50k = 5.1 us and with virtual dispatch would become 5.02 us ?
Linux = today: 1.60 us, virtual: 1.66 us

that would be a great news

(I'm asking because desktop numbers seem aligned with what I measured in
https://codereview.chromium.org/1947373002/, arm seems 1/4 of those numbers. but
it might be just due to the device. I think I used a n4 there)

[kraynov, 2017-02-01 19:21:58]

Yes, for 50K

Values per TRACE_EVENT0

Desktop:
- today: 1.60 us
- virtual: 1.66 us

Nexus 5:
- today: 50.2 us
- virtual: 51.6 us