[gmx-users] gromacs 3.3.3 vs 4.0.3 performance

[David van der Spoel]

Dimitris Dellis wrote:
> 
> Justin A. Lemkul wrote:
>>
>>
>> Dimitris Dellis wrote:
>>> Hi.
>>> I run the same  (exactly) simulations with v3.3.3 and v4.0.3, on the 
>>> same 64bit Q6600/DDR2-1066 machine, gcc-4.3.2 ,fftw-3.2.
>>> I found that the performance of 4.0.3 is roughly 30% lower than 3.3.3 
>>> (30% higher hours/ns), for few systems (512 molecules of 5-15 sites, 
>>> nstlist=10) I tried.
>>> This happens with single precision serial and parallel, np=2,4 
>>> (openmpi  1.3) versions and only when electrostatics (PME) are present.
>>> With Simple LJ potentials the performance is exactly the same.
>>> Is there any speed comparison 3.3.3 vs 4.0.3 available ?
>>> D.D.
>>>
>>
>> Can you show us your .mdp file?  What did grompp report about the 
>> relative PME load?  These topics have been discussed a few times; 
>> you'll find lots of pointers on optimizing performance in the list 
>> archive.

try turning off optimize_fft

>>
>> -Justin
>>
> Hi Justin,
> These are from the small system, no I/O only 1k steps.
> grompp.mdp
> ===========
> integrator               = md
> dt                       = 0.0010
> nsteps                   = 1000
> nstxout                  = 0
> nstvout                  = 0
> nstlog                   = 1000
> nstcomm                  = 10
> nstenergy                = 0
> nstxtcout                = 0
> nstlist                  = 10
> ns_type                  = grid
> dispcorr                 = AllEnerPres
> tcoupl                   = berendsen
> tc-grps                  = System
> ref_t                    = 293.15
> gen_temp                 = 293.15
> tau_t                    = 0.2
> gen_vel                  = no
> gen_seed                 = 123456
> constraints              = none
> constraint_algorithm     = shake
> dispcorr                 = AllEnerPres
> energygrps               = System 
> rlist                    = 1.6
> vdw-type                 = Cut-off
> rvdw                     = 1.6
> coulombtype              = PME
> fourierspacing           = 0.12
> pme_order                = 4
> ewald_rtol               = 1.0e-5
> optimize_fft             = yes
> rcoulomb                 = 1.6
> 
> related 4.0.3 grompp output
> Estimate for the relative computational load of the PME mesh part: 0.19
> 
> 4.0.3 mdrun serial timings (near zero omitted)
> 
>  Coul(T) + LJ                     576.513824       31708.260    71.5
>  Outer nonbonded loop               8.489390          84.894     0.2
>  Calc Weights                       6.006000         216.216     0.5
>  Spread Q Bspline                 128.128000         256.256     0.6
>  Gather F Bspline                 128.128000        1537.536     3.5
>  3D-FFT                          1088.769682        8710.157    19.6
>  Solve PME                         18.531513        1186.017     2.7
> 
> parallel 4.0.3 np=4
>  Average load imbalance: 5.2 %
>  Part of the total run time spent waiting due to load imbalance: 2.3 %
>  Performance:     96.086      7.380     14.414      1.665
> 
> 3.3.3 mdrun serial timings
>  Coul(T) + LJ                    576.529632    31709.129760    72.0
>  Outer nonbonded loop              8.487860       84.878600     0.2
>  Spread Q Bspline                128.128000      256.256000     0.6
>  Gather F Bspline                128.128000     1537.536000     3.5
>  3D-FFT                         1088.769682     8710.157456    19.8
>  Solve PME                        17.986469     1151.133984     2.6
> 
> parallel 3.3.3 np=4
> Performance:    144.132     12.556     21.600      1.111
> 
> D.D.
> 
>>>
>>> ------------------------------------------------------------------------
>>>
>>> _______________________________________________
>>> gmx-users mailing list    gmx-users at gromacs.org
>>> http://www.gromacs.org/mailman/listinfo/gmx-users
>>> Please search the archive at http://www.gromacs.org/search before 
>>> posting!
>>> Please don't post (un)subscribe requests to the list. Use the www 
>>> interface or send it to gmx-users-request at gromacs.org.
>>> Can't post? Read http://www.gromacs.org/mailing_lists/users.php
>>
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> gmx-users mailing list    gmx-users at gromacs.org
> http://www.gromacs.org/mailman/listinfo/gmx-users
> Please search the archive at http://www.gromacs.org/search before posting!
> Please don't post (un)subscribe requests to the list. Use the 
> www interface or send it to gmx-users-request at gromacs.org.
> Can't post? Read http://www.gromacs.org/mailing_lists/users.php


-- 
David van der Spoel, Ph.D., Professor of Biology
Molec. Biophys. group, Dept. of Cell & Molec. Biol., Uppsala University.
Box 596, 75124 Uppsala, Sweden. Phone:	+46184714205. Fax: +4618511755.
spoel at xray.bmc.uu.se	spoel at gromacs.org   http://folding.bmc.uu.se