Aug042016 - openpmix/openpmix GitHub Wiki

Aug 4, 2016

Ralph, Dave, Keita

• Issue #119 - support for “put” updates.

  • We discussed this at some length. The primary concern was over just how much of a use-case we really have for this feature, especially recalling that the objective of PMIx is to eliminate the exchange of such information. Rather than make signficant changes to the code base, we propose instead that we add an info key PMIX_REFRESH_CACHE that can be passed as a directive to the PMIx_Get call to indicate that the local cache is to be refreshed.

  • Our rationale is that the application must have some reason to believe that "put" information may have been updated. This is a rare use-case, if it even exists, and therefore can be treated as an exception.

• Issue #120 - user object for notifications

  • Dave will review

• Continue discussion of RFC #0009 - modification of PMIx_Get behavior

  • Primary concern was over how to handle node-specific information. Our current storage is based on rank, not nodeid, and so is the PMIx_Get function. We discussed ways of indicating that the PMIx_Get call is requesting node-level data, but it rapidly became confusing and clunky.

  • We propose to do the following:

    • Use the PMIx_Query function to retrieve node-level information, thus avoiding further overload of PMIx_Get
    • Modify PMIx_Query to use a new pmix_query_t structure that contains:
      • char **keys - identify information being requested
      • pmix_info_t qualifiers[] - array of qualifiers to direct the search
      • size_t nqual - number of qualifiers in array
    • Return function for PMIx_Query to remain the same, with each pmix_info_t in the array to correspond to a pmix_query_t. Multiple values returned for a single pmix_query_t will be provided in the pmix_data_array_t field in the pmix_info_t's value
    • Create a new hash table for node-level data

• Generalized Data Store

  • Keita asked a few clarifying questions
  • Ralph agreed to provide an example to help illustrate how the function pointer method would work