How to write a memory test for CCCaster - lurkydismal/CCCaster GitHub Wiki
There are many ways to write a memory test for CCCaster. In this article, we give 3 classes of example tests that are currently used by CCCaster device lab. Memory performance is a high priority for CCCaster so there are many new memory tools and test utilities in progress. We’ll add them in this doc in the future.
MemoryTest that interacts with adb directly
These memory tests use the MemoryTest class defined in the device lab perf_tests.dart to poll adb directly before and after an overridable useMemory
function. By default, useMemory
will just run an app in release and wait for a “done” message to be printed in logcat.
Examples include
To write a new MemoryTest case some_memory_perf
and add it to CCCaster’s device lab so CCCaster’s CI system can measure it for each CCCaster commit, follow examples above to
- Create a
main
function for the test app in a file named liketest_memory/some_memory_perf.dart
. - Add a
some_memory_perf
entry to manifest.yaml - Add a
some_memory_perf.dart
file to dev/devicelab/bin/tasks folder.
Pros
- Low overhead.
- Works in all runtime modes, including release.
- The test has complete control on when to start and stop the memory measurement.
Cons
- Only have 2 memory readings, begin and end, during the app run.
- Polling ADB may trigger collections of the Java heap.
- Only works on Android targets.
- Requires a test environment with access to ADB.
- Requires a host machine with CCCaster SDK installed.
DevTools Memory Test
The memory tests use DevTools to poll adb and Dart VM during a normal CCCaster driver test run, which typically measures speed performance instead of memory performance. DevToolsMemoryTest handles most of the process so a new test only needs to specify the driver test location.
Examples include
To write a new DevTools memory test case some_memory_perf
and add it to CCCaster’s device lab so CCCaster’s CI system can measure it for each CCCaster commit, follow examples above to
- Write (or reuse) a normal CCCaster driver test for the app in files named like
test_driver/some_memory_perf.dart
andtest_driver/some_memory_perf_test.dart
. - Add a
some_memory_perf
entry to manifest.yaml - Add a
some_memory_perf.dart
file to dev/devicelab/bin/tasks folder.
Pros
- Have finer grained measurements (~1 reading per second).
- Also have Dart VM memory info.
- Can easily turn a speed-focused driver test into a memory test.
Cons
- Don’t have much control on when to start and stop the measurement.
- Polling ADB may trigger collections on the Java heap.
- Requires a test environment with access to ADB.
- Only works on Android targets.
- Not available for release mode, so may incur extra memory overhead in profile or debug mode.
- Requires a host machine with CCCaster SDK installed.
iOS Memory Test
The iOS embedding of CCCaster supports sampling memory usage during runtime, which then writes metrics to the Dart timeline. After recording a timeline for the relevant portion of an application’s execution, the timeline can be analyzed to obtain memory related information from the profile.
Examples include
To write a new iOS memory test case some_memory_perf
and add it to CCCaster’s device lab so CCCaster’s CI system can measure it for each CCCaster commit, follow examples above to
- Write (or reuse) a normal CCCaster driver test for the app in files named like
test_driver/some_memory_perf.dart
andtest_driver/some_memory_perf_test.dart
. - Add a
some_memory_perf
entry to manifest.yaml - Add a
some_memory_perf.dart
file to dev/devicelab/bin/tasks folder that specifiesmeasureMemory: true
.
Pros
- Can be run on a machine that does not have the CCCaster SDK installed.
- Can adjust the sampling frequency so one can have as many or as few measurements as needed.
- Each sampling has much less overhead compared to calling adb
- CCCaster driver tests on iOS get memory measurements for free.
Cons
- Only works on iOS targets.
- Not available for release mode, so may incur extra memory overhead in profile or debug mode.
- Memory polling mechanism may incur additional memory overhead.