How to create a Raspberry Pi cluster using MPI, NFS, and Python
install from here
Navigate to the download and uncompress it
tar -xzf <mpich-install-folder>
Enter mpi folder
cd <mpich-folder>
Install. Disabling fortran shortens install time, but it still takes a while!
./configure --disable-fortran
Build and install mpi (This takes a looong time)
make; sudo make install
For info on your finished install:
mpiexec --version
More info from mpi tutorials
MPI is orginally meant to be used with C++, but I wanted to use python. This is a handy library for that!
sudo pip install mpi4py
A hello world to test mpi4py
mpiexec -n 5 python -m mpi4py.bench helloworld
More mpi4py information on the docs
Aliases will allow you to refer to each node without using its IP address.
sudo nano /etc/hosts
Add <IP-address> <alias> to the file for each node
Alias all nodes to numbered work nodes. One node should also be aliased to master.
New users with the same name on all nodes makes setup easier
sudo adduser mpiuser
Give mpiuser sudo access
sudo usermod -aG sudo mpiuser
Login to mpiuser
su - mpiuser
Enable SSH using sudo raspi-config
Generate keys on the master node
ssh-keygen
Send the key to each node
ssh-copy-id <node>
NFS, or Network File System, allows the work nodes to access the file system of the master node as their own, to ease the exchange of data.
sudo apt-get install nfs-kernel-server
Make the directory to share
mkdir cluster
sudo nano /etc/exports
Add the line /home/mpiuser/cluster *(rw,sync,no_root_squash,no_subtree_check)
To share the folder every time there is a change:
sudo exportfs -a
sudo apt-get install nfs-common
Create the folder on the work node
mkdir cluster
Mount the shared folder
sudo mount -t nfs master:/home/mpiuser/cluster ~/cluster
Use df -h to check the shared directory
To mount access when booting
sudo nano /etc/fstab
Add the line master:/home/mpiuser/cluster /home/mpiuser/cluster nfs x-systemd.automount
To run your programs, use
mpiexec -n 10 -machinefile nodes python -m mpi4py script.py
where 10 is the amount of repitiations, nodes is a file with a list of the work nodes, and script.py is the job.
For comparison,
mpiexec -n 10 python -m mpi4py script.py
Runs the program only on the master node.
In shell, time <command> can be used to run a command and return the time taken to complete the job.
To power off, you can give each user sudo priveleges and shutdown each easily over ssh.