[Nek5000-users] survey of apps
nek5000-users at lists.mcs.anl.gov
nek5000-users at lists.mcs.anl.gov
Thu May 2 11:13:05 CDT 2013
Dear Neks,
As part of a strategic planning exercise, I'd like to
get a survey of performance profiles over a broad range
of applications. I'm particularly interested in production
runs involving lots of compute cycles, since these would
be the most important for future optimization (i.e., cases
where you are running 100s of simulations of more or less
the same type). I'm wonder if I could ask each of you
to send one or at most two logfiles directly to me -
fischer at mcs.anl.gov
with a very brief description giving a ballpark estimate
of utilization, i.e., something like:
We have run 100 of these cases, typically 6 hours each, or
whatever suits. I don't really need to know anything about
the app.
The one thing I do need, however, is that your logfile contain
the statistics that are printed at the end when the job terminates
with nek's standard termination criterion (i.e., it does not stop
because your run time expired). When you
grep time logfile
you should see something like:
163.0805685520172 7.9530425400815632E-002
eslv time 121259 153.8100430965424 7.5009415695555526E-002
pres time 121259 433.6815834045410 0.2114959564030525
crsl time 122729 0.7530927658081055 3.6726501852911642E-004
hmhz time 363777 469.4056932926178 0.2289177355990086
spro time 121260 14.56137013435364 7.1012259246232062E-003
usbc time 606294 54.14203882217407 2.6403755014022586E-002
axhm time 2070607 497.7768969535828 0.2427536813301520
advc time 363777 252.1085748672485 0.1229472179574250
vdss time 486498 7.609275102615356 3.7108583277339125E-003
dsum time 2438769 18.98749208450317 9.2597379085877490E-003
dadd time 0 1.966142654418945 9.5883993472067994E-004
ddsl time 122729 49.81083679199219 2.4291533165547920E-002
solv time 122729 22.48464226722717 1.0965212964212575E-002
prep time 121259 10.51901149749756 5.1298659712803868E-003
total elapsed time : 2.05059E+03 sec
total solver time incl. I/O : 2.05054E+03 sec
time/timestep : 1.69104E-02 sec
CPU seconds/timestep/gridpt : 4.23184E-06 sec
Preferably, the case in question should be 3D and have run
for at least 100 timesteps (the more the better), so that
the statistics have more significance.
My plan is to build a histogram with pts/core on the x-axis, time
spent where on the y axis, etc. --- something like this... By
gathering this on a variety of platforms we'll be able to make more
informed decisions.
Thank you all!
Paul
More information about the Nek5000-users
mailing list