Assigned
Status Update
No update yet.
Comments
cc...@google.com
Note that the equivalent to passing flags = 0x10 to Debug.startMethodTracing for am start / am profile is passing --clock-type wall, which I just checked was added in API 34
# from `am help`:
--clock-type <TYPE>: type can be wall / thread-cpu / dual. Specify
the clock that is used to report the timestamps when profiling
The default value is dual. (use with --start-profiler)
ap...@google.com
Project: platform/frameworks/support
Branch: androidx-main
commit 9619685e2a7c8060de62280ea15ef3f4a22d0a30
Author: Chris Craik <ccraik@google.com>
Date: Fri Mar 15 15:03:31 2024
Enable Debug.startMethodTracing 0x10 flag to improve perf on supported platforms
Bug: 329146942
Test: LazyListScrollingBenchmark#scrollProgramatically // API 24, 34
```
// flags = 0x0
LazyListScrollingBenchmark.scrollProgrammatically_newItemComposed[LazyRow]
timeNs min 898,308.0, median 952,308.0, max 1,108,001.0
profilerTimeNs min 141,052,553.0, median 148,459,631.5, max 241,766,179.0
// flags = 0x10
LazyListScrollingBenchmark.scrollProgrammatically_newItemComposed[LazyColumn]
timeNs min 880,538.0, median 961,347.5, max 1,303,154.0
profilerTimeNs min 46,547,619.0, median 50,705,466.5, max 150,416,324.0
// measured on Mokey API 34, using profiling.perfCompare.enable = true
```
Change-Id: Icc567a6fbe5825a012eb2dad8d774cbcef36ac8c
M benchmark/benchmark-common/src/main/java/androidx/benchmark/Profiler.kt
https://android-review.googlesource.com/3005963
Branch: androidx-main
commit 9619685e2a7c8060de62280ea15ef3f4a22d0a30
Author: Chris Craik <ccraik@google.com>
Date: Fri Mar 15 15:03:31 2024
Enable Debug.startMethodTracing 0x10 flag to improve perf on supported platforms
Bug: 329146942
Test: LazyListScrollingBenchmark#scrollProgramatically // API 24, 34
```
// flags = 0x0
LazyListScrollingBenchmark.scrollProgrammatically_newItemComposed[LazyRow]
timeNs min 898,308.0, median 952,308.0, max 1,108,001.0
profilerTimeNs min 141,052,553.0, median 148,459,631.5, max 241,766,179.0
// flags = 0x10
LazyListScrollingBenchmark.scrollProgrammatically_newItemComposed[LazyColumn]
timeNs min 880,538.0, median 961,347.5, max 1,303,154.0
profilerTimeNs min 46,547,619.0, median 50,705,466.5, max 150,416,324.0
// measured on Mokey API 34, using profiling.perfCompare.enable = true
```
Change-Id: Icc567a6fbe5825a012eb2dad8d774cbcef36ac8c
M benchmark/benchmark-common/src/main/java/androidx/benchmark/Profiler.kt
ap...@google.com
Project: platform/frameworks/support
Branch: androidx-main
commit e8c21d81df2242168e21b1310b62609887e70406
Author: Chris Craik <ccraik@google.com>
Date: Mon Mar 11 09:58:23 2024
Enable measuring the overhead of profilers during the profiling phase
Bug: 329146942
Test: BenchmarkStateConfigTest
Test: LazyListScrollingBenchmark # with argument enabled
Relnote: """
* Added inst argument `androidx.benchmark.profiling.skipWhenDurationRisksAnr`, can be set to false to avoid skipping method traces when expected duration may cause an ANR - strongly recommended to avoid in CI runs.
* Added experimental inst argument `androidx.benchmark.profiling.perfCompare.enable`, can be set to true to run comparison timing between measurement and profiling phases. Useful in evaluating overhead of e.g. method tracing.
"""
Sample output, showing method tracing is roughly 150 times slower than the measurement just before of AOT code:
```
LazyListScrollingBenchmark.scrollProgrammatically_newItemComposed[LazyColumn]
timeNs min 915,461.0, median 1,030,923.0, max 1,775,153.0
profilerTimeNs min 141,567,169.0, median 148,904,938.5, max 241,511,871.0
```
Change-Id: I61fb4703f52e8451bf87abf7d50f6f3b1facdabf
M benchmark/benchmark-common/api/current.txt
M benchmark/benchmark-common/api/restricted_current.txt
M benchmark/benchmark-common/src/androidTest/java/androidx/benchmark/BenchmarkStateConfigTest.kt
M benchmark/benchmark-common/src/main/java/androidx/benchmark/Arguments.kt
M benchmark/benchmark-common/src/main/java/androidx/benchmark/BenchmarkState.kt
M benchmark/benchmark-common/src/main/java/androidx/benchmark/MetricCapture.kt
M benchmark/benchmark-common/src/main/java/androidx/benchmark/MicrobenchmarkPhase.kt
M benchmark/benchmark-common/src/main/java/androidx/benchmark/Profiler.kt
https://android-review.googlesource.com/2998123
Branch: androidx-main
commit e8c21d81df2242168e21b1310b62609887e70406
Author: Chris Craik <ccraik@google.com>
Date: Mon Mar 11 09:58:23 2024
Enable measuring the overhead of profilers during the profiling phase
Bug: 329146942
Test: BenchmarkStateConfigTest
Test: LazyListScrollingBenchmark # with argument enabled
Relnote: """
* Added inst argument `androidx.benchmark.profiling.skipWhenDurationRisksAnr`, can be set to false to avoid skipping method traces when expected duration may cause an ANR - strongly recommended to avoid in CI runs.
* Added experimental inst argument `androidx.benchmark.profiling.perfCompare.enable`, can be set to true to run comparison timing between measurement and profiling phases. Useful in evaluating overhead of e.g. method tracing.
"""
Sample output, showing method tracing is roughly 150 times slower than the measurement just before of AOT code:
```
LazyListScrollingBenchmark.scrollProgrammatically_newItemComposed[LazyColumn]
timeNs min 915,461.0, median 1,030,923.0, max 1,775,153.0
profilerTimeNs min 141,567,169.0, median 148,904,938.5, max 241,511,871.0
```
Change-Id: I61fb4703f52e8451bf87abf7d50f6f3b1facdabf
M benchmark/benchmark-common/api/current.txt
M benchmark/benchmark-common/api/restricted_current.txt
M benchmark/benchmark-common/src/androidTest/java/androidx/benchmark/BenchmarkStateConfigTest.kt
M benchmark/benchmark-common/src/main/java/androidx/benchmark/Arguments.kt
M benchmark/benchmark-common/src/main/java/androidx/benchmark/BenchmarkState.kt
M benchmark/benchmark-common/src/main/java/androidx/benchmark/MetricCapture.kt
M benchmark/benchmark-common/src/main/java/androidx/benchmark/MicrobenchmarkPhase.kt
M benchmark/benchmark-common/src/main/java/androidx/benchmark/Profiler.kt
Description
There are new flags we can pass to improve method tracing performance, such as
0x10for the wall clock, and we should also look into streaming traces, if that improves performance.As part of this, want to enable measuring cost of method tracing easily with microbenchmark and macrobenchmark.