[petsc-dev] Seeking OLCF users complaining about poor build times

[Barry Smith]

> On Feb 26, 2015, at 6:04 PM, Jed Brown <jed at jedbrown.org> wrote:
> 
> Barry Smith <bsmith at mcs.anl.gov> writes:
> 
>>  Ok, based on what we have so far we need to completely rethink our process, our previous test results are pretty much useless Kill any tests you have outstanding at the moment.
>> 
>>  1) Forget about maint branch!  
>>  2) Forget about --download-superlu_dist since it does not have a parallel make
>> 
>>   So we want runs on master with --download-hypre --download-metis --download-parmetis --with-debugging=0 and good optimization flags and proper TMPDIRs, if appropriate, on all systems.
> 
> Why download these external packages?

  My thinking was I want to know the time to build a complete USABLE PETSc install (which generally includes a few external packages). From the point of view of a user this is all they care about; they could give a hoot WHY the total time is slow.

  Note that hypre and the metis libraries do have parallel builds

>  That part should be parallel (if
> their build systems handle it; it's a problem we know how to solve in
> any case), so it's measuring more-or-less the same thing as PETSc's
> build.  I'd rather see stripped-down configure because that is the part
> that is characteristically different from PETSc build, and also likely
> to show the greatest discrepancy.

  You are right that the stripped down configure gives you two distinct phases of computation to help understand which is slow. 

  Note that Satish's stripped down numbers for Mira already demonstrate that Mira is a fine system for configuring and compilering. Less than 7 minutes for configure and less than 2 for compiler. From my perspective we no longer have a right to complain about Mira. (Yes it would be nice if the 7 was smaller but the 7 is completely acceptable).

  Barry