StanfordLegion/legion

deferred_allocs consistently fails with the UCX backend when separate channels are used for small vs. large messages

Opened this issue · 1 comments

https://gitlab.com/StanfordLegion/legion/-/merge_requests/950
https://gitlab.com/StanfordLegion/legion/-/jobs/5240234828

The deferred_allocs test incorrectly depends on the order of active messages sent/received by the network module. When we use 2 separate workers (channels/end points) to separate small messages form the large ones, the test consistently fails with `

test/realm/deferred_allocs.cc:178: void directed_test_memory(const TestConfig&, Realm::Memory, Realm::Processor, const char*, const char*): Assertion `ii.alloc_result' failed.

Will disable the test for now as per suggestion from @streichler

This should probably be added to #1305