[gmx-users] time accounting in log file with GPU
Sikandar Mashayak
symashayak at gmail.com
Fri Jul 25 19:31:28 CEST 2014
Thanks Szilárd.
I am bit confused about the -resethway or -resetstep options. Do they
exclude the time spent on initialization and load-balancing from the total
time reported in the log file, i.e., the time reported is the total time
spent only in the loop/iterations over time-steps?
Thanks,
Sikandar
On Thu, Jul 24, 2014 at 4:30 PM, Szilárd Páll <pall.szilard at gmail.com>
wrote:
> On Fri, Jul 25, 2014 at 12:48 AM, Sikandar Mashayak
> <symashayak at gmail.com> wrote:
> > Thanks Mark. -noconfout option helps.
>
> For benchmarking purposes, additionally to -noconfout I suggest also using:
> * -resethway or -resetstep: to exclude initialization and
> load-balancing at the beginning of the run to get a more realistic
> performance measurement from a short run
> * -nsteps N or -maxh: the former is useful if you want to directly
> compare (e.g. two-sided diff) the timings from the end of the log
> between multiple runs
>
> Cheers,
> --
> Szilárd
>
> >
> > --
> > Sikandar
> >
> >
> > On Thu, Jul 24, 2014 at 3:25 PM, Mark Abraham <mark.j.abraham at gmail.com>
> > wrote:
> >
> >> On Fri, Jul 25, 2014 at 12:12 AM, Sikandar Mashayak <
> symashayak at gmail.com>
> >> wrote:
> >>
> >> > Hi
> >> >
> >> > I am running a benchmark test with the GPU. The system consists of
> simple
> >> > LJ atoms.
> >> > And I am running only very basic simulation with NVE ensemble and not
> >> > writing any
> >> > trajectories or energy values. My grompp.mdp file is attached below.
> >> >
> >> > However, in the time accounting table in the md.log, I observe that
> write
> >> > traj. and comm energies
> >> > operations take 40% of time each. So, my question is that even if I
> have
> >> > specified not to write
> >> > trajectories and energies, why is 80% of time being spent on those
> >> > operations?
> >> >
> >>
> >> Because you're writing a checkpoint file (hint, use mdrun -noconfout),
> and
> >> that load is imbalanced so the other cores wait for it in the global
> >> communication stage in Comm. energies (fairly clear, since they have the
> >> same "Wall time"). Hint - make benchmarks run for about a minute, so you
> >> are not dominated by setup and load-balancing time. Your compute time
> was
> >> about 1/20 of a second...
> >>
> >> Mark
> >>
> >>
> >> > Thanks,
> >> > Sikandar
> >> >
> >> > R E A L C Y C L E A N D T I M E A C C O U N T I N G
> >> >
> >> > On 2 MPI ranks
> >> >
> >> > Computing: Num Num Call Wall time
> Giga-Cycles
> >> > Ranks Threads Count (s) total sum
> %
> >> >
> >> >
> >>
> -----------------------------------------------------------------------------
> >> > Domain decomp. 2 1 11 0.006 0.030
> >> 2.1
> >> > DD comm. load 2 1 2 0.000 0.000
> >> 0.0
> >> > Neighbor search 2 1 11 0.007 0.039
> >> 2.7
> >> > Launch GPU ops. 2 1 202 0.007 0.036
> >> 2.5
> >> > Comm. coord. 2 1 90 0.002 0.013
> >> 0.9
> >> > Force 2 1 101 0.001 0.003
> >> 0.2
> >> > Wait + Comm. F 2 1 101 0.004 0.020
> >> 1.4
> >> > Wait GPU nonlocal 2 1 101 0.004 0.020
> >> 1.4
> >> > Wait GPU local 2 1 101 0.000 0.002
> >> 0.2
> >> > NB X/F buffer ops. 2 1 382 0.001 0.008
> >> 0.6
> >> > Write traj. 2 1 1 0.108 0.586
> >> 40.2
> >> > Update 2 1 101 0.005 0.025
> >> 1.7
> >> > Comm. energies 2 1 22 0.108 0.588
> >> 40.3
> >> > Rest 0.016 0.087
> >> 5.9
> >> >
> >> >
> >>
> -----------------------------------------------------------------------------
> >> > Total 0.269 1.459
> >> 100.0
> >> >
> >> >
> >>
> -----------------------------------------------------------------------------
> >> >
> >> >
> >> > grompp.mdp file:
> >> >
> >> > integrator = md-vv
> >> > dt = 0.001
> >> > nsteps = 100
> >> > nstlog = 0
> >> > nstcalcenergy = 0
> >> > cutoff-scheme = verlet
> >> > ns_type = grid
> >> > nstlist = 10
> >> > pbc = xyz
> >> > rlist = 0.7925
> >> > vdwtype = Cut-off
> >> > rvdw = 0.7925
> >> > rcoulomb = 0.7925
> >> > gen_vel = yes
> >> > gen_temp = 296.0
> >> > --
> >> > Gromacs Users mailing list
> >> >
> >> > * Please search the archive at
> >> > http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> >> > posting!
> >> >
> >> > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> >> >
> >> > * For (un)subscribe requests visit
> >> > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> >> > send a mail to gmx-users-request at gromacs.org.
> >> >
> >> --
> >> Gromacs Users mailing list
> >>
> >> * Please search the archive at
> >> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> >> posting!
> >>
> >> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> >>
> >> * For (un)subscribe requests visit
> >> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> >> send a mail to gmx-users-request at gromacs.org.
> >>
> > --
> > Gromacs Users mailing list
> >
> > * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> posting!
> >
> > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> >
> > * For (un)subscribe requests visit
> > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> send a mail to gmx-users-request at gromacs.org.
> --
> Gromacs Users mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> send a mail to gmx-users-request at gromacs.org.
>
More information about the gromacs.org_gmx-users
mailing list