Please Note: This version of the documentation is outdated. We recommend reading the guide instead. It contains current information on new features.
General information
Motivation
OpenMPI, despite being relatively old, still remains the most dominant programming model used in high-performance computing (HPC). As of today Fiji supports the GPU parallelization through CLIJ and allows executing automated HPC workflows by means of Automated_workflow_for_parallel_Multiview_Reconstruction. However, there are still no genuine internally parallel plugins developed specifically for deployment on large scale parallel machines like HPC clusters or supercomputers. This is a preliminary OpenMPI framework for Fiji and model parallel implementations of the most common image processing operations included in ImageJ.
Extensions Description
This presents a set of Ops adapted to herein presented OpenMPI wrapper, jointly comprising a solution designed to be seamlessly used in any existing ImageJ2 code. Moreover, framework automatically initializes and disposes connections to the OpenMPI environment, making available OpenMPI implementations for the most frequent operations covered by Ops. The package contains example implementations for fundamental image processing functions as well as basic math, logical and statistics operations.
Installation
TBD
How to use
In order to demonstrate convenience of this approach, we prepared an example of parallelized convolution. We used Python for readers’ convenience, however plugins can be invoked from any other scripting language supported by Fiji.
kernel = ops.run("create.img", [3, 3], FloatType())
for p in kernel:
p.set(1.0/kernel.size())
def fn():
output = ops.create().img(input)
ops.filter().convolve(output, Views.extendMirrorSingle(input), kernel)
return output
input = scifio.datasetIO().open(input_path)
output = Measure.benchmark(fn, rounds)
scifio.datasetIO().save(output, output_path)
The Listing shows the OpenMPI implementation with initialization of a 3×3 blur kernel (lines 1-3), definition of a processing function with convolution (it reuses already existing implementation of a convolution filter in ImageJ2) on lines 5-8 and finally its execution in the benchmark mode to measure execution times (the variable rounds defines the number of repetitions).
Noticeably, using the OpenMPI extension does not add any extra layer as long as data are available in a form of IterableInterval.
Debugging
Debugging printouts on the console shows the node allocation and dataset division. The following listing shows the dataset divided to 6 nodes.
Fri Jun 19 08:36:49 2020[1,2]<stdout>:Chunk{data=test.tif, offset=13334, len=6667, chunks=1}
Fri Jun 19 08:36:49 2020[1,0]<stdout>:Chunk{data=test.tif, offset=0, len=6667, chunks=1}
Fri Jun 19 08:36:49 2020[1,5]<stdout>:Chunk{data=test.tif, offset=33335, len=6665, chunks=1}
Fri Jun 19 08:36:49 2020[1,1]<stdout>:Chunk{data=test.tif, offset=6667, len=6667, chunks=1}
Fri Jun 19 08:36:49 2020[1,4]<stdout>:Chunk{data=test.tif, offset=26668, len=6667, chunks=1}
Fri Jun 19 08:36:49 2020[1,3]<stdout>:Chunk{data=test.tif, offset=20001, len=6667, chunks=1}
