Description
Per #9298 and #9299, it was discovered that --cpu-set
does not behave the same way that it did in Open MPI 3.1.x. Instead, it behaves like --cpu-list
.
The 2 linked PRs nominally fix the --cpu-set
behavior to be correct, but were ultimately rejected for two reasons:
- The code is very complex and high risk. It feels like every time we have touched this code in the past, we have broken it.
- We'd be changing behavior in the middle of two release series. This may be small behavior that may not affect many people, but it is a change in behavior from "something that results in a successful code run" to "some other behavior that also results in a successful code run." It feels dangerous to do that.
The compromise that we came up with is to keep the --cpu-set
operational (because disabling it may change user scripts to suddenly stop working) but document that --cpu-set
in 4.x is accidentally doing exactly the same thing as --cpu-list
. There are probably 3 places that need to be updated:
- README.md
- mpirun(1) man page
- The output from
mpirun --help
- ...? (maybe others?)
The fix for restoring --cpu-set
behavior will be included in v5.0.0.
@gpaulsen volunteered to document the issue. Please verify with @awlauria that my description above is correct, and update all documentation in v4.0.x and v4.1.x accordingly.