[Pw_forum] Help with Building and Installing Quantum Espresso

[Filippo Spiga]

Dear Jean,

please refer to http://qe-forge.org/gf/project/q-e-gpu/
in order to configure and install QE-GPU on your machine.  The procedure is easy and lot of details are reported in the file GPU/READEME.GPU


On Dec 29, 2012, at 11:48 AM, Jean Martel <john.martel9 at gmail.com> wrote:
> I need some help with building and installing Quantum Espresso on my desktop running under Ubuntu 12.04. I am interested in running GPU accelerated TD-DFT calculations to compute excited-state properties, with occasional B3LYP geometry optimizations.

TDDFT is not entirely supported. Only part of it can exploit GPU acceleration thanks to some libraries and shared modules.


> One of the main issues I am having is the with the two different QE-GPU versions, there are version 5.01 and 5.02. However 5.02 contains much fewer and totally different files, and I am not sure which one to use.

I do not thing you are really interested to understand too much details about the implementation. If you have run calculations using QE 5.0.1 and you need to restart/resume from that version then pick up the QE-GPU-5.0.1. If you have to run new calculations use the QE-GPU-5.0.2. Latest version always means better code, less bug, sometimes better/improved performance, less "problems".


> Any guidance or pointers on how to build and install QE  on my system for the types of calculations i am looking for would be great. 
> 
> My System Specifications:
> Ubuntu12.04
> gcc 4.6.3 with gfortran
> Intel MKL version 11 update 1
> Cuda 5.0
> Hexacore i7-3930K
> Dual Gtx 690
> 64 GB RAM


What it is not clear is if you have a single GTX690 card (that has 2 GPU die on the same board) or two different cards. If I understand correctly, the GTX 690 comes with 2 integrate GPU in the same board and 4 GByte of RAM. QE-GPU will detect two distinct GPUs and we will manage them safely. Since your system has so much RAM I do suggest to run QE-GPU in MPI+OpenMP mode in order to use both GPUs together as better as possible. So

$ cd GPU
$ ./configure --enable-parallel --enable-openmp --enable-cuda --with-gpu-arch=30 --with-cuda-dir=<where your CUDA is installed> --enable-magma --enable-phigemm
$ cd ..
$ make -f Makefile.gpu all-gpu

Then you should run using 2 MPI and 3 OpenMP per thread (or 6 OpenMP if hyper-threading is enabled). Please install Open MPI on your system (latest version is fine). I have no idea what kind of binding settings suggest because I have not found any detail about the NUMA topology of the processor. Please read the RREADME.GPU file.

The only TDDFT executable will be turbo_lanczos-gpu.x. I never tested it so.. good luck. Actually I will be happy to receive a test case and some feedback about it.

Few words about some performance expectation: the GTX 690 uses a GK104 GPU chip. This chip has been designed for gaming mainly. So its double precision peak is quite poor compared to another GPU generation chip (GK1110). So please do not expect speedups like 2 or 3 times. GPU will help to have a less time-to-solution but not a big improvement in performance. Moreover the GPU implementation targets compute intensive calculation. If you have small systems composed by less than ~50 atoms then there is not guarantee about any performance benefit.

Happy New Year and Happy Computing

(PLEASE subscribe the QE-GPU mailing list, I will not reply to any further email on this one)

Cheers,
Filippo

--
Mr. Filippo SPIGA, M.Sc.
http://filippospiga.me ~ skype: filippo.spiga

«Nobody will drive us out of Cantor's paradise.» ~ David Hilbert

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20121230/9dde737c/attachment.html>