Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

From: George Bosilca (bosilca_at_[hidden])
Date: 2007-06-21 15:14:30


Well, I might be the only one shocked by this but I'll let you know
my feelings anyway.

Having to modify the base mca in order to add one global variable
that is never used for anything useful except by the orted, doesn't
look like the best coding approach to me. Note that orterun use this
variable too, but for a different purpose: it set it to true [which
is anyway the default value] just to make sure that the default value
is correctly understood by that little nasty processor ...

So, I propose to update the documentation for the
mca_base_param_recache_files function in the header file to something
closer to the actual behavior of the function:

This function mostly reload/recache the MCA parameters file(s).
Except, if the obscure and undocumented
opal_mca_base_param_use_amca_sets global is set to false in which
case it will only recache half of the files you expect, in a way that
not even the original developer might have imagined.

Now, if I look at most of the tools in the orte/tools directory they
apparently take special care of how they initialize the opal layer.
They all call mca_base_param_init before calling the opal_init. Is
there a real reason for that ? I miss to see it, as the execution is
equivalent with the case where we call directly the opal_init ...

I try to figure out if there is a consistent way to deal with the
amca stuff, and as far as I see there is not. Is there a reason to
have the orted behave differently than the others tools ? Btw, the
call to mca_base_param_init in orted is completely useless, and
wrongly documented. In fact the call will not do anything because the
opal_init_util was called few lines prior and it already called the
mca_base_param_init function (which has set the initialized status to
true).

   george.

On Jun 20, 2007, at 7:00 AM, jjhursey_at_[hidden] wrote:

> Author: jjhursey
> Date: 2007-06-20 10:00:40 EDT (Wed, 20 Jun 2007)
> New Revision: 15144
> URL: https://svn.open-mpi.org/trac/ompi/changeset/15144
>
> Log:
> Fix an AMCA parameter regression introduced (as a side effect of)
> in r14449
> (and, due to lack of in code documentation, in r14661).
>
> The {{{opal_mca_base_param_use_amca_sets}}} flag tells the orted
> that it should
> not look at the parameter files just yet since it may have an AMCA
> parameter
> file to look at first. So we need to set this to {{{false}}} before
> initializing
> the MCA paras, then quickly turn around and re-init them when we
> have the full
> information.
>
> This commit fixes #1058
>
>
>
> Text files modified:
> trunk/orte/tools/orted/orted.c | 10 ++++++----
> 1 files changed, 6 insertions(+), 4 deletions(-)
>
> Modified: trunk/orte/tools/orted/orted.c
> ======================================================================
> ========
> --- trunk/orte/tools/orted/orted.c (original)
> +++ trunk/orte/tools/orted/orted.c 2007-06-20 10:00:40 EDT (Wed, 20
> Jun 2007)
> @@ -83,8 +83,6 @@
> /*
> * Globals
> */
> -
> -extern bool opal_mca_base_param_use_amca_sets;
> orted_globals_t orted_globals;
>
> static struct opal_event term_handler;
> @@ -196,6 +194,12 @@
> /* initialize the globals */
> memset(&orted_globals, 0, sizeof(orted_globals_t));
>
> + /* Need to set this so that the orted does not throw a warning
> message
> + * about missing AMCA param files that are located in the
> relative or
> + * absolute paths (e.g., not in the package directory).
> + */
> + opal_mca_base_param_use_amca_sets = false;
> +
> /* Ensure that enough of OPAL is setup for us to be able to
> run */
> if (OPAL_SUCCESS != opal_init_util()) {
> fprintf(stderr, "OPAL failed to initialize -- orted
> aborting\n");
> @@ -209,8 +213,6 @@
> * Do not parse the Aggregate Parameter Sets in this pass.
> * we will get to them in a moment
> */
> - /* GMS: what does this do and why is it gone now? */
> - /* opal_mca_base_param_use_mca_sets = false; */
> mca_base_param_init();
>
> /* setup to check common command line options that just report
> and die */
> _______________________________________________
> svn mailing list
> svn_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/svn



  • application/pkcs7-signature attachment: smime.p7s