- changed milestone to 2018.09.30 release
a standalone test for completions
currently the test for rput and rget and vis and atomics are acting as secondary tests for completions.
From the material I have seen, how a completion is handled is somewhat tied up to what the operation is meant to do. but it can be deceiving when, say, vis.cpp test fails, and the error is really over in completions.hpp
This relates to what I have heard referred to as "lower triangular dependence" where fundamental features have tests, then operations that have dependence on more fundamental features get tests that can safely assume the lower API levels are complete.
but can we really test completion in isolation? Perhaps we need a gasnet ack function that just pings a remote rank and then sets the handle to complete, and we can write a test harness for upcxx on the ack.
Comments (3)
-
-
- changed milestone to Deferred indefinitely
This issue was triaged at the 2018-06-13 Pagoda meeting and assigned a new milestone/priority.
It was noted completions do not have a hardened internal API, and its likely to change with optimizations like rput::as_rpc getting special treatment.
-
- changed status to resolved
Recently added tests test/rput-cover.cpp and test/copy-cover.cpp meet the spirit of this proposal for rput and copy, which specify all three forms of events and have the most "interesting" set of underlying implementations for the various completions.
As mentioned in the original request, the implementation of each completion variant is often intimately tied to the communication operation itself, so this is probably the closest we will get to a "completions test".
- Log in to comment
Mass roll-over of unresolved issues to the next milestone.