Modify the behavior of PMIx_Get
This RFC simplifies and extends the processing of PMIx_Get requests:
reserving the PMIx namespace key and enforcing tighter rules on the rank value used to store/retrieve values
When requesting info about a process in another job, ensuring that a copy of the job-level info for that job is included in the reply so that subsequent requests can be locally satisfied
Copyright (c) 2016 Intel, Inc. All rights reserved.
This document is subject to all provisions relating to code contributions to the PMIx community as defined in the community’s LICENSE file. Code Components extracted from this document must include the License text as described in that file.
Initially, PMIx was fairly loose regarding the rank provided when executing the PMIx_Get function. The library would check for the requested key in two locations:
the job-level data provided at time of PMIx_Init. This data could be marked with either PMIX_RANK_WILDCARD, or with a specific rank value. Thus, the check was made against both the wildcard and the specified rank (if other than wildcard).
the data “published” by the application itself via the PMIx_Put function.
Thus, a request for a data item resulted in up to three separate hash table lookups, thereby impacting scalable launch times. This RFC places restrictions on the rank and key values that allow the “get” operation to predictably complete with only a single lookup.
The restrictions enacted by this RFC are:
keys beginning with “pmix” are solely controlled by the PMIx community – i.e., the “pmix” namespace is reserved. All PMIx namespace keys will be stored in the job-level data according to the following rule:
data pertaining to process-level (e.g. PMIX_LOCAL_RANK) information must be marked with the corresponding rank.
data that differs by process (e.g., PMIX_APP_NUM, PMIX_LOCAL_SIZE) due to the location of the process or its membership within an application must be marked with the corresponding rank.
data pertaining to job-level information (e.g., PMIX_UNIV_SIZE) will be marked with the PMIX_RANK_WILDCARD rank.
data published by an application via the PMIx_Put function must have a key that lies outside the reserved namespace – i.e., the key cannot begin with “pmix”
In addition to enforcing these restrictions, this RFC extends the PMIx_Get behavior to ensure that a request for information from a process in another namespace will return the job-level info for that namespace in addition to any info published by that process via PMIx_Put. This allows any subsequent request for job-level info to be locally satisfied without an additional communication.
The PMIx library implementation is covered in the Simplify the PMIx_Get processing pull request. The prototype has been tested against Open MPI and is currently committed into the master repo of that project.
Ralph H. Castain