[Trilinos-Users] Teuchos::TimeMonitor::summarize problem with local communicators

[Mark Hoemmen]

On Dec 8, 2011, at 12:00 PM, <trilinos-users-request at software.sandia.gov> wrote:
> Message: 4
> Date: Thu, 08 Dec 2011 16:22:07 +0100
> From: "Georg Hammerl" <hammerl at lnm.mw.tum.de>
> Subject: [Trilinos-Users] Teuchos::TimeMonitor::summarize problem with
> 	local communicators
> To: trilinos-users at software.sandia.gov
> Message-ID: <4EE0D61F.9000804 at lnm.mw.tum.de>
> Content-Type: text/plain; charset=iso-8859-1; format=flowed
> 
> Hi Teuchos,
> 
> I have a question regarding the Teuchos::TimeMonitor.
> 
> Is there a special reason why the summarize method takes more or less 
> automatically the MPI_COMM_WORLD communicator.

1. Because it was a legacy interface which I got stuck supporting ;-)

2. So that summarize() works in packages that have no communicator concept (e.g., Anasazi and Belos).

> I'm working with a split communicator and not all of my processes 
> call the summarize. Hence, some of my processes wait in front of several 
> "barrier (*pComm)" calls within the summarize but the remaining 
> processes will never reach these barriers and my whole simulation stops.
> 
> An alternative would be to add a further argument to the summarize 
> method that includes a potential local communicator. The default can of 
> course be the MPI_COMM_WORLD.

This is not very hard for us to do.  We may have to give the new method a different name or otherwise change the interface so as to retain backwards compatibility.

I'm the person who would likely have to implement this.  You are welcome to communicate with me directly regarding the urgency of this request, so that I can prioritize tasks.

Best,
mfh