Repository navigation
[gcstress] HardwareIntrinsics timing out on osx-arm64 and linux-arm #78323
Description
The hw intrinsics tests seem to time out on aforementioned platforms after #74886.
Test run: https://dev.azure.com/dnceng-public/public/_build/results?buildId=81893&view=ms.vss-test-web.build-test-results-tab&runId=1708824&resultId=220189&paneView=debug
Maybe the stripe count needs to be increased @davidwrighton?
Activity
davidwrighton commentedon Nov 28, 2022
@jakobbotsch From looking at the logs, this appears to be caused by some sort of actual deadlock, not a test failure. In particular, the test is making good progress for a short period of time, and then stops making progress. I don't have easy access to the appropriate hardware to easily test this, but I think this needs to be investigated as a product failure, and not just increase the parallelism of the testing.
One thing to be aware of is that in the past, before my change, most hardware intrinsic tests were never run under GCStress. (The tests would mostly be skipped during GCStress execution.)
BruceForstall commentedon Nov 29, 2022
Another set of failures: https://dev.azure.com/dnceng-public/public/_build/results?buildId=94375&view=ms.vss-test-web.build-test-results-tab
Note that these are GCStress=3 failures, so are generally due to VM, not JIT (or timeout due to GCStress=3 being very slow).
Perhaps we should once again disable them under GCStress.
We currently have GCStressIncompatible. Perhaps we should also have GCStressIncompatible_3/GCStressIncompatible_C for more granularity. Actually, this should already be possible when using the xunit attributes. See https://github.com/dotnet/runtime/blob/main/docs/workflow/ci/disabling-tests.md and https://github.com/dotnet/arcade/blob/main/src/Microsoft.DotNet.XUnitExtensions/src/RuntimeTestModes.cs.
jakobbotsch commentedon Jan 4, 2023
I don't really see what David was saying above, from the log indeed good progress is made, but we hit the 4 hour Helix timeout specified here after which the Helix job gets killed. I will disable it under gcstress again.