Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] MALLOC_MMAP_MAX (and MALLOC_MMAP_THRESHOLD)
From: Eugene Loh (Eugene.Loh_at_[hidden])
Date: 2010-01-25 14:28:14


Jeff Squyres wrote:
On Jan 17, 2010, at 12:13 PM, Ashley Pittman wrote:
  
It is something bear in mind when analysing application performance and diagnosing memory problems with applications, a FAQ entry should be considered the minimum IMHO.
    
What, specifically, do you think we should put in the FAQ?
(it's easy to add stuff to the FAQ; it's more difficult to decide exactly what should be there)
  
One possibility is to have a section that discusses how different MPI implementations may differ in performance in "non-MPI" kinds of ways.  E.g.,

*) Some MPI implementations add optimization flags to their mpicc/mpif90 wrappers by default.  So, default compilation with one MPI gives you optimization, while another does not.

*) Different MPI implementations have different default process binding behaviors.

*) This whole MALLOC issue.

Arguably, these issues aren't inherent to MPI.  They are general HPC performance tuning considerations.  It's a slippery slope to get into these issues (our users care about performance and these things are important to performance, therefore we should tell our users about these things, but where do we stop?).  On the other hand, given that these are knobs that various MPI implementations touch (and touch in different ways), they may make sense for us to discuss.

Looking into some such FAQ additions is on one of my "to do" lists, but I'm not territorial and can't promise if/when I would get to any of this.