libpmi: move some complexity to broker #2172

Problem: our PMI-2 library is incomplete and is an attractive nuisance at this point. It was added to allow OpenMPI jobs to be launched under flux when OpenMPI, through its MCA plugins, determines that it should dlopen slurm's libpmi2.so. OpenMPI could be tricked via LD_LIBRARY_PATH into opening a Flux libpmi2.so that provides the minimal API needed by OpenMPI, and maps it to the simple PMI-1 wire protocol. For more background see flux-framework#746 Since we submitted a flux MCA module that helps OpenMPI choose the right PMI in the flux environment, and shipping this libpmi2.so with flux is misleading as it doesn't implement the full PMI-2 "spec", let's get rid of it. Fixes flux-framework#2082

Problem: function for converting array of integers to comma-separated string exists in both test/pminfo.c and test/clique.c. Refactor to pmi_cliquetostr() in clique.[ch].

Problem: the PMI implementation in libpmi is a source of confusion because it's used by the flux broker to bootstrap in non-Flux environments, AND it is the PMI library provided to users for launching under Flux. We can significantly dumb this thing down if it is assumed to only be used by users running under Flux and then it will be more understandable when looked at only from that perspective. Drop singleton and dlopen (wrap) methods, and in fact drop the method abstraction and just assume the "simple client" is the only method. Update pmi unit tests.

Problem: PMI library doesn't consistently protect against NULL parameters used to return values (segfault). Return PMI_ERR_INVALID_ARG in those cases.

Problem: strtoul() and related functions are used without proper error handling. As recommended in strtoul(3), one must set errno = 0 before the call and check for errno != 0 after the call.

Problem: the code for obtaining the "clique" information from the 'PMI_process_mapping' key hardwires canonical PMI API functions, but a user (internal for now) might want to use the handled simple_client functions instead. Move this code to the simple_client class. Currently this is used behind the canonical clique functions in the Flux libpmi.so, but not internally. In the past it was used to assist with PGM wireum. Update test/pminfo.

Problem: support dlopen()ing libpmi.so and for falling back to singleton operation was removed from the Flux libpmi.so, but the broker needs this capability. Re-add this capability, with significantly reduced complexity given that it need not be a faithful PMI-1 API interface.

Problem: libpmi/test/pminfo.c didn't call PMI_Get_universe_size(). Add a call.

Pull out the server thread to its own reusable module, and change it slightly so that it could support multiple clients.

Problem: the PMI_*() API functions don't have unit tests per se. Add a unit test for these interfaces.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

libpmi: move some complexity to broker #2172

libpmi: move some complexity to broker #2172

Commits on Jun 5, 2019