Skip to content

Latest commit

 

History

History
53 lines (46 loc) · 2.69 KB

thelio-12-v1-variations.md

File metadata and controls

53 lines (46 loc) · 2.69 KB
Raw

Persistent Coarray Buffer for Version 1

NAG comes very close to MPI performance for version 1 of the coarray implementation. Are there any modifications that would improve its performance? Note that each invocation of the gather_aux procedure allocates a local coarray buffer src[*] (and deallocates it on return). The amount of storage associated with the coarray is negligible -- merely a scalar derived type with a single (unassociated) array pointer. However there may be significant overhead associated with allocation of a coarray. So two modifications of version 1 were considered that replaced the local coarray with a persistent coarray that is allocated once and reused by each invocation of gather_aux:

  • Version 1A: The local coarray is replaced by a new allocatable component %src[:] of the index_map type. Source in directory coarray1a.
  • Version 1B: The local coarray is replaced by a module variable. Source in directory coarray1b.
time (ms) MPI v1 v1A v1B
B0-12 0.0075 0.020 0.0050 0.0053
B1-12 0.010 0.022 0.0077 0.0078
B2-12 0.019 0.032 0.016 0.016
B3-12 0.044 0.051 0.035 0.037
B4-12 0.085 0.129 0.120 0.126
  • There is no real difference between 1A and 1B; not surprising.
  • The modified versions are now faster than the original and MPI for all but the largest problem.
  • The B4 problem is very large with a lot of communication. All coarray versions performed about the same. Perhaps some cache effects being seen here that are associated with NAG's shared memory coarray communication implementation.

Design Implications

There are significant design downsides to the 1A and 1B approaches.

  • For 1B it just feels wrong to have multiple instances of index_map sharing the same coarray buffer. Perhaps it is okay, but it may be prone to "threadsafe" questions.
  • For 1B first use can allocate, but who deallocates?
  • For 1A, the passed object for the gather and gather_aux procedures must be changed from intent(in) to intent(inout) because the coarray component is being modified. Not terrible, and is something that is always accommodated when an otherwise intent(in) passed object carries along temporary workspace.
  • The biggest problem with 1A is that the passed object argument to init can no longer be intent(out), which is fundamental to the semantics of an init procedure that is regarded as instantiating an object. This is due to a restriction on dummy arguments that have a coarray component -- apparently the standard writers didn't want allocatable coarray components to be implicitly deallocated on entry.