Re: [gmx-users] GPU/CPU perfomance loss

[Mark Abraham]

Hi,

And it's very CPU bound because the processor is fairly antiquated and the
GPU fairly recent. Current versions of GROMACS can't work very well in that
situation.

Mark

On Tue, 10 May 2016 20:18 Szilárd Páll  wrote:

> Hi Justin,
>
> Your run is simply *very* CPU-bound. Your GPU is so much faster than the
> CPU that even after increasing the cut-off to 2 nm to shift PME load away
> from the CPU, there is still a lot of GPU idling.
>
> For details, please have a look at on of our recent papers, e.g.
>
> http://arxiv.org/abs/1506.00716
>
> http://arxiv.org/abs/1507.00898
>
> Cheers,
>
> --
> Szilárd
>
> On Mon, May 9, 2016 at 7:00 PM, Van Zyl, J, Mnr  <
> justi...@sun.ac.za> wrote:
>
> > Good day,
> >
> >
> > I am running simulations from a workstation and keep getting a gromacs
> > note at the end of the run saying "The GPU has >25% less load than the
> CPU.
> > This imbalance causes performance loss."
> >
> >
> > I have read a previous post pertaining to this problem but the only thing
> > that seems to increase performance is by changing nstlist from 150 to 20
> > however the gromacs note still shows. I will post the log file of that
> run
> > soon however below is the log file of the first run:
> >
> > http://pastebin.com/3CUDh960
> >
> >
> > Is there anything I can do to increase perfomance or is this simply as
> > good as it will get for my system (22467 atoms)?
> >
> >
> > Thanks for the help,
> >
> > Regards,
> >
> > Justin van Zyl
> >
> > [http://pastebin.com/i/facebook.png]
> >
> > md_1.log - Pastebin.com
> > pastebin.com
> >
> >
> >
> >
> > Justin van Zyl
> > MSc. Candidate
> > Department of Chemistry and Polymer Science,
> > Stellenbosch University, Private Bag X1 Matieland,
> > 7602, South Africa
> > Cell: +27 72 9372 755
> > justi...@sun.ac.za
> > http://www.klumperman-group.net/
> >
> > --
> > 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-requ...@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-requ...@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-requ...@gromacs.org.

Re: [gmx-users] GPU/CPU perfomance loss

[Szilárd Páll]

Hi Justin,

Your run is simply *very* CPU-bound. Your GPU is so much faster than the
CPU that even after increasing the cut-off to 2 nm to shift PME load away
from the CPU, there is still a lot of GPU idling.

For details, please have a look at on of our recent papers, e.g.

http://arxiv.org/abs/1506.00716

http://arxiv.org/abs/1507.00898

Cheers,

--
Szilárd

On Mon, May 9, 2016 at 7:00 PM, Van Zyl, J, Mnr  <
justi...@sun.ac.za> wrote:

> Good day,
>
>
> I am running simulations from a workstation and keep getting a gromacs
> note at the end of the run saying "The GPU has >25% less load than the CPU.
> This imbalance causes performance loss."
>
>
> I have read a previous post pertaining to this problem but the only thing
> that seems to increase performance is by changing nstlist from 150 to 20
> however the gromacs note still shows. I will post the log file of that run
> soon however below is the log file of the first run:
>
> http://pastebin.com/3CUDh960
>
>
> Is there anything I can do to increase perfomance or is this simply as
> good as it will get for my system (22467 atoms)?
>
>
> Thanks for the help,
>
> Regards,
>
> Justin van Zyl
>
> [http://pastebin.com/i/facebook.png]
>
> md_1.log - Pastebin.com
> pastebin.com
>
>
>
>
> Justin van Zyl
> MSc. Candidate
> Department of Chemistry and Polymer Science,
> Stellenbosch University, Private Bag X1 Matieland,
> 7602, South Africa
> Cell: +27 72 9372 755
> justi...@sun.ac.za
> http://www.klumperman-group.net/
>
> --
> 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-requ...@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-requ...@gromacs.org.

Re: [gmx-users] GPU+CPU

[Szilárd Páll]

On Sun, Sep 20, 2015 at 7:11 PM, Parker de Waal 
wrote:

> Hello Everyone,
>
> I recently started to explore GROMACS (switching over from AMBER) and need
> some help understanding how to launch GPU+CPU simulations.
>
> GROMACS 5.0.6 was compiled with the following cmake arguments:
> cmake .. -DGMX_BUILD_OWN_FFTW=ON
> -DCUDA_TOOLKIT_ROOT_DIR=/cm/shared/apps/cuda70/toolkit/7.0.28 -DGMX_MPI=ON
> -DCMAKE_INSTALL_PREFIX=/home/parker.dewaal/applications/gromacs-5.0.6
> -DGMX_GPU=ON
>
> Now I would like to run a simulation on a box with 28 CPU cores and 2
> titan K80s (4 GPU threads), what would be the difference between running
> the following:
>

I assume you mean Tesla K80s.


>
> Mdrun
>
> Mdrun –ntmpi 4 –ntomp 7
>
> Mpirun –np 4 mdrun_mpi
>

The former uses multi-threading based built-in MPI, the latter an external
MPI library that implements ranks as processes. This leads to technical
differences between the two runs and the general outcome is not trivial to
predict. Typically the thread-MPI based run will be slightly more
efficient, but with modern MPI implementations the performance should be
close.

Cheers,
--
Szilárd


>
> Cheers,
> Parker
>
>
> --
> 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-requ...@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-requ...@gromacs.org.

Re: [gmx-users] GPU/CPU load imbalance

[Pappu Kumar]

The GPU has much less load than CPU. Could you please give me some idea how to 
port code to GPU? Thank you. 


 On Tuesday, 7 July 2015 2:26 AM, Szilárd Páll pall.szil...@gmail.com 
wrote:
   

 What kind of load imbalance? I assume you are referring to CPU-GPU imbalance 
as your runs are CPU-bound. In that case the only thing that helps is either 
porting code to CUDA or a faster CPU.
Reaction field electrostatics will help too in getting better performance in 
CPU-bound runs, but beware that you'll have to justify the choice of deviating 
form the most established way of treating electrostatics!
--
Szilárd
On Mon, Jul 6, 2015 at 1:32 PM, Pappu Kumar papu...@yahoo.com wrote:

I have 3x GTX 970 for 1x Intel i7-5820K with X99-E WS motherboard to remove 
PCIe limitations. So I have 4 threads/gpu and I face load imbalances.

I am wondering if it will be a good idea to use reaction-field instead of PME 
in order to reduce the processor load (It would be less accurate). Let me know 
if you have any suggestions for proper load balancing. Thank you.



  
-- 
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-requ...@gromacs.org.

Re: [gmx-users] GPU/CPU load imbalance

[Szilárd Páll]

What kind of load imbalance? I assume you are referring to CPU-GPU
imbalance as your runs are CPU-bound. In that case the only thing that
helps is either porting code to CUDA or a faster CPU.

Reaction field electrostatics will help too in getting better performance
in CPU-bound runs, but beware that you'll have to justify the choice of
deviating form the most established way of treating electrostatics!

--
Szilárd

On Mon, Jul 6, 2015 at 1:32 PM, Pappu Kumar papu...@yahoo.com wrote:

 I have 3x GTX 970 for 1x Intel i7-5820K with X99-E WS motherboard to
 remove PCIe limitations. So I have 4 threads/gpu and I face load imbalances.

 I am wondering if it will be a good idea to use reaction-field instead of
 PME in order to reduce the processor load (It would be less accurate). Let
 me know if you have any suggestions for proper load balancing. Thank you.

 --
 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-requ...@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-requ...@gromacs.org.