[Mpi-comments] Tags for Nonblocking Collective Operations

[Qingyu Meng]

Draft URL: http://meetings.mpi-forum.org/mpi3.0_draft_2.pdf
Page 196 5.12 Nonblocking Collective Operations

The proposed MPI-3 non-blocking collective operations are designed for
better overlapping computation and communication. The matching of
these operations is based on the order of the calls instead of using
tags. This limits the possibility to use dynamic code/task execution
when the order of calls to collective operations should be decided at
run-time. When running with multiple threads, our computational
framework, Uintah (www.uintah.utah.edu) has to duplicate MPI
communicators to ensure collective messages get to the right buffer.

The current draft version of the MPI-3 standard states, “The use of
tags for collective operations can prevent certain hardware
optimizations.” (Page. 108, Line. 14). However, the standard does
allow the use of communicators to match collective
operations. Does this usage also prevent certain hardware
optimizations? Duplicating communicators for message matching is
inefficient, as communicators must be reserved before use and also
cause additional overhead.

We would like to request further consideration in supporting the use
of tags in nonblocking collective operations, to avoid the
inefficiencies related to communicator duplication. This would not
only benefit our computational framework, but other task-based,
parallel frameworks that rely on collective operations as we move
toward exascale.


Qingyu Meng
Scientific Computing and Imaging Institute, University of Utah