To: J3                                                     J3/19-244
From: Malcolm Cohen
Subject: Collective subroutines and STAT=
Date: 2019-October-15
Reference: 19-181

1. Introduction

Paper 19-181 contains five purported interpretation requests glommed into
a single paper. This paper discusses the fifth of them.


2. The alleged problem

From 19-181:
    QUESTION 5:
    A collective subroutine is defined (3.143.2) as an
        "intrinsic subroutine that performs a calculation on a team of
         images without requiring synchronization",
    and collective subroutines are invoked
        "on all active images of its current team in segments that are not
         ordered with respect to each other"
    (16.6 par. 1). Is the guarantee provided by 16.6 par. 7:
        "If the STAT argument is present and ... the current team contained
         failed images, an error condition occurs and the value
         STAT_FAILED_IMAGE from the intrinsic module ISO_FORTRAN_ENV is
         assigned to the STAT argument"
    too strong? It seems to imply that all images in the team will have
    knowledge of a failed image if there is one.

    ANSWER 5:
    Yes. It was not intended to require a synchronization of all the images
    involved. Rather than
        "if the current team contained failed images",
    the standard should say
        "if the current team is detected to contain failed images".
    This is particularly important for CO_BROADCAST, because the failure of
    one image does not prevent others receiving the correct result, and
    each should be allowed to continue execution as soon as it has the
    result.

    EDIT 5:
    In 16.6 Collective subroutines, para 7, sentence 2, change
        "if the current team contained failed images"
    to
        "if the current team is detected to contain failed images".
    so that the sentence becomes
        "Otherwise, if the current team is detected to contain failed
         images, an error condition occurs and the value STAT_FAILED_IMAGE
         from the intrinsic module ISO_FORTRAN_ENV is assigned to the STAT
         argument."


3. Response from /INTERP

The question is without merit, and the answer does not make sense.

(a) According to 16.6, the value assigned to each STAT argument is
    consistent. It would be a disaster if some part of the program were to
    be told that the CO_WHATEVER was successfully executed but another
    part was told that it was not.

(b) If there are any failed images in the team, CO_WHATEVER is going to
    find them as it involves communication with each image in the team.
    Thus the edit suggested is unlikely to make any difference to the
    synchronisation argument.

(c) The assertion that failed image detection necessarily requires full
    synchronisation is false. Collectives do necessarily require some
    internal synchronisation in any case.

(d) Minor nitpickery, but the standard says "contains". Not past tense.

(e) In the case of CO_BROADCAST, as well as all the other collective
    subroutines, the variable does not receive "the correct result" in the
    case of an error condition occurring, but becomes undefined (16.6p6).

(f) My view is that "Unreliable collective subroutines" would be a new
    feature, different from the one we currently have, and should not be
    added via the interpretation process.

Therefore /INTERP politely declines to accept this interpretation request
in its current form at this time.

Instead, /HPC is requested to examine these issues and determine whether
there any issue here that needs addressing in the current standard, or
whether there are any issues here that should be addressed in the next
revision.

===END===