Skip to content
Maxim Yurkin edited this page Nov 15, 2020 · 13 revisions

Introduction

If you plan to run ADDA on a cluster, MPI is probably already installed on your system. You should consult someone familiar with the particular MPI package. ADDA usage of MPI is based on the MPI 2.1 standard, and it should work with any implementation that is compliant with this or higher versions of the standard. Version of MPI standard is checked for conformity by ADDA both during compilation and at runtime.

ADDA will run on any hardware compatible with MPI, but, in principle, it may run more efficiently on hardware with shared memory address space, e.g. multi-core PCs. However, such hardware also has its drawbacks, e.g. the performance may be limited by memory access speed. A few tests performed on different computers showed that currently using two cores of a dual-core PC results in computational time from 60% to 75% of that for sequential execution on the same machine. We plan to optimize ADDA specifically for such hardware, using e.g. OpenMP.

If you plan to run a parallel version of ADDA on a single computer using a multi-core processor, you need first to install some implementation of MPI. Installation instruction can be found in the manual of a particular MPI package. In the following we briefly describe several examples for various operating systems. Note that modern MPI implementations, such as MPICH and Open MPI have a lot of advanced features, like combining several single- and multi-core computers in a cluster, which are not discussed here.

Unix

The easiest is to install MPICH or Open MPI (or other implementation) through the package manager - see, e.g., the full list for MPICH.

Alternatively, here is a brief instructions for compiling MPICH yourself. Note, however, that they were tested on old version 2-1.4, but should work similarly for later versions as well. A detailed description for compiling MPICH yourself are given in its Installer's Guide, see especially Section 2.2 "From A Standing Start to Running an MPI Program". Below we summarize the most important points.

  • Download the source code of MPICH and unzip it.
  • If you are installing MPI under a user account, choose installation directory and create it, for instance:
    mkdir $HOME/mpich-install
    
  • Cd into the directory with MPICH source (which we here use also as a build directory) and type
    configure -prefix=$HOME/mpich-install -enable-fast
    
    where -prefix specifies installation directory, and -enable-fast is recommended for production runs (at expense of some error reporting). Other useful configuration options to consider can be found in the Installer's Guide.
  • Type
    make
    make install
    
  • Add corresponding bin directory to the $PATH. For example, if you are using BASH, add
    export PATH=$HOME/mpich-install/bin:$PATH
    
    to your .bash_profile and relogin. Now you should be able to compile MPI-version of ADDA using mpicc (automatically detected by mpi/Makefile) and run it using mpiexec.

macOS

Windows

The following guidelines may be outdated - they are based on MPICH2 v. 1.4, while later versions of MPICH2 (or MPICH) seem to lack native support for Windows. We plan to provide instructions for its predecessor - Microsoft MPI (MS-MPI) soon.

MPICH2 Installer's Guide contains instructions for Windows - Section 9 "Windows Version". However, they seem somewhat incomplete in comparison with the following practical guide.

  • Download the installer of MPICH2 for your combination of Windows and hardware and install it.

  • Add bin subdirectory to Windows environmental variable PATH. This will give you direct access to mpiexec and smpd.

  • On some systems you may need to manually install smpd as a system service (check smpd -status). For that open Windows console (cmd) with administrator's privileges (e.g. right-clicking on its icon and choosing corresponding option). Then run

    smpd -install -phrase behappy
    

    The (pass)phrase should be the same for all computers running in a cluster. Specifying it at this moment will stop mpiexec from asking for it at every run.

  • First time you run mpiexec you may be prompted for Windows account name and password, which will be used to run MPI processes. These data is then stored encrypted in Windows registry. MPICH2 can not use an account with empty password. So if that is the case for your account, you need to either add a password (you may then also turn-on automatic login) or create a new account specially to be used by MPICH2.

  • Moreover, you will be once prompted by the Windows firewall, whether to allow mpiexec and smpd to access network. If you plan to run ADDA only on a single PC, you may block them. Otherwise, you should unblock them. You may also get similar prompts for the parallel program itself, e.g., adda_mpi.

The above actions are necessary to use precompiled adda_mpi, included in Windows packages. If you want to compile MPI version of ADDA yourself, then additionally

  • Since Windows package of MPICH2 does not contain compiler wrapper mpicc, either
Clone this wiki locally