You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Collecting some ideas on additional ways to run our existing tests:
Include code coverage metrics, so we can find the parts of our codebase not being exercised.
Run tests in CPU, OpenMP and GPU mode.
Run at various core counts (e.g. 1, 2, 4 and 8}.
Run with various maximum dimension settings
Multi-rank runs -- Multi-node would be marginally better, but even just multiple ranks on a single node should be sufficient to find most functional bugs. This might require changes to the legate launcher or test driver, as we currently assume mostly rank-per-node. We should also bind each rank to distinct resources (this can be machine-specific, and thus could be handled in quickstart). It is possible we need to enable some option in gasnet to enable communication over shared memory on the same node.
Run with -lg:safe_ctrlrepl 1 on at least 2 ranks, to check for control replication violations. Possibly add some tools that can help pinpoint where the violation comes from.
Run with -lg:partcheck.
Run with some form of memory safety instrumentation, e.g. Legion bounds checks, valgrind, or the LLVM sanitizers.
Run with some instrumentation to detect when we have introduced a reference cycle in the core data structures, like the ones fixed by Object cycle fix #84.
Run some long running tests with checks for resource leaks.
The text was updated successfully, but these errors were encountered:
* Provide requisite workflow permission
- Replace GITHUB_TOKEN with WORKFLOW_TOKEN.
- Trigger script with push also.
(GitHub Apps added to repository can't access the secrets in the repository without adequate permission. The token issued for GitHub Actions doesn't have this permission by default.)
* - Remove the push mechanism. It was only added for testing.
Collecting some ideas on additional ways to run our existing tests:
-lg:safe_ctrlrepl 1
on at least 2 ranks, to check for control replication violations. Possibly add some tools that can help pinpoint where the violation comes from.-lg:partcheck
.The text was updated successfully, but these errors were encountered: