The garbage collector is an integral part of the .NET runtime. Changes to the .NET runtime can have significant effects on program correctness and performance. As such, GC has strong implementation and testing requirements.
We strongly recommend that if you want to make a change to GC, that you share your proposal prior to implementation. This will ensure that if there are any questions or concerns, that they are addressed before significant work is done.
The native implementation of the GC - including the GCHeap and gc_heap classes.
Required Testing: (Instructions for running tests are below)
- gc_heap class: Changes to gc_heap requires a 48 hour stress run on a debug build. This helps to ensure correctness of the change given that the GC runs in many different configurations and interacts with many other sub-systems within the runtime. For some simple changes a stress run may not be required. If you believe your change does not manipulate the heap itself, mention this in your pull request and we can tell you if stress is required or not.
- GCHeap class: GCHeap changes generally do not require stress runs, but do require functional testing of the affected code.
- Performance changes: Performance specific changes require a performance test run.
Managed APIs in System.GC and System.Runtime.GCSettings. These APIs allow an application developer to view and modify GC options.
Required Testing: Validation of the behavior of the affected APIs.
Stress testing must run for at least 48 hours against a debug build.
Stress testing for checked and release builds can be run locally. Please following the instructions described in gc-stress-run-readme.md. You can also request it on pull requests with The .NET CI infrastructure with the trigger phrase:
@dotnet_bot test <platform> <flavor> gc_reliability_framework
This will run the stress framework for the default amount of time (15 hours) on the given platform and build flavor.
A functional test run executes the same code as a stress run, but only runs for 30 minutes.
The "Long GC" tests are a series of GC tests whose running time is too long or memory usage is too high to run with the rest of the Priority 0 unit tests. The "Standalone GC" build mode builds and runs the GC in a semi-standalone manner (see https://github.com/dotnet/coreclr/projects/3).
You may also wish to run the GC Simulator tests. They may take up to 24 hours to complete and are known to sometimes fail on Ubuntu due to poor interactions with the Linux OOM killer. However, they have proven to be quite useful in finding bugs in the past:
@dotnet_bot test windows Release gcsimulator
@dotnet_bot test Ubuntu Release gcsimulator
@dotnet_bot test OSX10.12 Release gcsimulator
Performance tests can be run using the GC Benchmarking Infrastructure hosted within the benchmarks section of the performance repo. This tool allows you to run the tests capturing traces through PerfView, and later analyze such traces for detailed results on GC's, heaps, and related behaviors. These tests also allow you to directly compare behaviors and metrics between different .NET builds and machines.
The README.md located at the link mentioned above contains detailed instructions on how to setup and use the tool. Additionally, there is a docs section with all the information to perform further analysis on the traces.