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

[Satish Balay]

On Thu, 26 Feb 2015, Satish Balay wrote:

> On Thu, 26 Feb 2015, Nathan Collier wrote:
> 
> > Ok, so I am not sure I got all that, but here goes some more information
> > (attached log).
> > 
> > No packages / master
> > 
> > [nate at titan] time python no_packages.py
> > ...
> > real 14m32.169s
> > user 1m51.527s
> > sys 3m40.734s
> > 
> > [nate at titan] time make all
> > ...
> > real 15m56.004s
> > user 8m8.971s
> > sys 52m42.734s
> 
> BTW: This number for 'make all' looks bizzare [or I don't know how to interpret it].
> 
> - 'sys' time is 52m?
> - 'user' time is 8m - but 'real' is 15m - so its not a parallel build?
>    But configure sets 'MAKE_NP = 24' (with gnumake build).
> 
> Perhaps its all intel compiler license server overhead? :)

Ah - I was misreading. 'sys' is cumilative time across all procs. So
speedup factor with parallel build is something like:

(user+sys)/real - which is 3.8

The 'sys' time - which is 6.5 times more than 'user' time - this is
puzzling. Perhaps this happend if the machine is actively used by many
users?

For Mira the numbers are:

real    1m49.837s
user    26m1.417s
sys     9m11.808s

with (user+sys)/real = 19.2 [with MAKE_NP=32] - and the 'sys' time is
a small fraction of user time.

Satish