part_conc_spt_abc

Nanoparticle concentration measurements using single particle tracking and Approximate Bayesian Computation.

Instructions: Running the CPU version

The code is developed using Julia 0.6.0. To run the analysis, just run src/cpu/run_part_conc_spt_abc.jl in the following way

julia -p <number_of_cores>

where the parameters (all mandatory) are

• <number_of_cores> : number of CPU cores (threads) to utilize
• <program> : path to run_part_conc_spt_abc.jl
• <input> : input file

The file test/test_part_conc_spt_abc.jl generates a simulated example data set and calls the analysis program.

Instructions: Building & Running the GPU version

Building the GPU version:

First generate the makefiles using premake:

$ external/premake5 gmake

Note that only make-based builds have been tested. The project definition in premake5.lua includes raw nvcc command lines which may or may not need to be adjusted for other build types ("actions") supported by premake.

Next, run make:

$ make -j4

This builds the debug version. An optimized release version can be built via

$ make -j4 config=release_native

This should produce the main binary (accel-<config>.exe) in the bin/ directory.

Running the GPU version:

Run bin/accel-<config>.exe --help for an overview of command line options. The most important options are:

• -i <input> : select input file
• -o <output> : select output file (will be overwritten!)
• -g <gpuspec> : select GPU configuration (see below)

Other command line options can typically be omitted (which will use their default values).

If no output file is specified via -o, the default output file path from the input will be used.

The <gpuspec> is used to select GPUs. It is a comma-separated list (no spaces!) of <GPU number>/<number of queues>. The default is -g 1/30, which uses the first available GPU with 30 asynchronous queues. For a system with two identical GPUs one might use -g 1/30,2/30. Work is distributed evently across the queues, so simple manual "load-balancing" can be performed by lowering the number of queues (e.g., for a system with a GTX1060 + a GTX960, -g 1/30,2/21 turned out to be a good choice).

Available CUDA GPUs can be listed by using the --list-devices command line option.