How to Install and Run

Installation

In order to provide Fiji an OpenMPI support we need to add scijava-parallel-mpi-1.0-SNAPSHOT.jar to its classpath.

Build from Source Codes

To build the extension from the source codes, download the code from Github https://github.com/fiji-hpc/scijava-parallel-mpi and run build.sh script which will take care of compilation and installation of the final jar file into the default Fiji plugin directory.

git clone https://github.com/fiji-hpc/scijava-parallel-mpi

cd scijava-parallel-mpi

./build.sh

After the script is complete, you should be able to find the package installed in Fiji.app/jars/scijava-parallel-mpi-1.0-SNAPSHOT.jar

Install from Update Sites

TBD

Running

This section illustrates the parallel execution on IT4Innovations clusters. Nevertheless, OpenMPI environment is so widespread environment that the following commands will work (with some parameter modifications) on almost any cluster or supercomputer.

1. Start a new MPI interactive job on Salomon with 2 nodes and 1 hour walltime qsub -q qexp -l select=2:ncpus=24:mpiprocs=1:ompthreads=24,walltime=01:00:00 -I -X

2. Set up an environment with Java and OpenMPI libraries module load Java/1.8.0_144 module load OpenMPI/4.0.0-GCC-6.3.0-2.27

In order to run the script without OpenMPI you would normally use, e.g., ~/Fiji.app/ImageJ-linux64 --ij2 --headless --console --run scijava-parallel-mpi/scripts/canny.py 'input_path="/home/user/FluorescentCells.tif",output_path="output_canny.tif"'

3. To execute the script parallel you need to run it in OpenMPI environment using mpirun and specify the number of instances -np 2 ( as we have in the previous command allocated 2 nodes with qsub call). mpirun -np 2 --bind-to none ~/Fiji.app/ImageJ-linux64 --ij2 --headless --console --run scijava-parallel-mpi/scripts/canny.py 'input_path="/home/user/FluorescentCells.tif",output_path="output_canny.tif"'

Because Fiji internally uses threads, it is recommended to add --bind-to none option.

Debugging OpenMPI Run

For the debugging purposes we can add additional options, where --merge-stderr-to-stdout merges stderr to stdout for each process, --timestamp-output puts timestamp to each line of output to stdout, stderr, and stddiag, --tag-output tags each line with the process jobid and MPI_COMM_WORLD rank of the process that generated the output, and the channel which generated it. Final command with debug outputs running on 2 nodes using OpenMPI can look like

mpirun -np 2 --bind-to none --tag-output --timestamp-output --merge-stderr-to-stdout ~/Fiji.app/ImageJ-linux64 --ij2 --headless --console --run scijava-parallel-mpi/scripts/canny.py 'input_path="/home/user/FluorescentCells.tif",output_path="output_canny.tif"'