keichi / mpedm Goto Github PK

A scalable parallel implementation of Empirical Dynamic Modeling

License: MIT License

C++ 80.15% CMake 5.62% C 0.28% Python 12.96% Dockerfile 0.99%

parallel-computing high-performance-computing empirical-dynamic-modeling gpu-computing

mpedm's Introduction

mpEDM

This project has been superseded by kEDM and will no longer be maintained. kEDM has generally higher performance than mpEDM, and provides Python bindings for ease of use.

GPU-accelerated implementation of Empirical Dynamic Modeling

Requirements

CMake 3.10+
C++ compiler that supports C++11 (Intel C++ Compiler is strongly recommended for building the CPU backend)
HDF5 (parallel build if building the multi-node benchmarks)
(optional) MPI if building the multi-node benchmarks
(optional) ArrayFire 3.6.2+ if building the GPU backend

Installation

Install dependencies:
- Install HDF5.
- Install ArrayFire if building the GPU backend.

Clone the source code:

$ git clone --recursive https://github.com/keichi/mpEDM.git

Run cmake:
```
$ cd mpEDM
$ mkdir build
$ cd build
$ cmake -DCMAKE_BUILD_TYPE=Release ..
```
- -DCMAKE_BUILD_TYPE can be either Release, RelWithDebInfo, Debug or MinSizeRel.
- Add -DCMAKE_CXX_COMPILER=/path/to/c++ to select the C++ compiler to use.
- Add -DCMAKE_CXX_FLAGS="..." to customize the C++ compiler flags.
- Add -DHDF5_DIR=/path/to/hdf5 if HDF5 is not installed in a standard path.
- Add -DArrayFire_DIR=/path/to/arrayfire if ArrayFire is not installed in a standard path.
Run make:
```
$ make
```

Literature

For a detailed description of the algorithm and performance measurements using up to 2,044 GPUs, please refer to the following paper:

Wassapon Watanakeesuntorn, Keichi Takahashi, Kohei Ichikawa, Joseph Park, George Sugihara, Ryousei Takano, Jason Haga, Gerald M. Pao, "Massively Parallel Causal Inference of Whole Brain Dynamics at Single Neuron Resolution", 26th International Conference on Parallel and Distributed Systems (ICPADS 2020), Dec. 2020. 10.1109/ICPADS51040.2020.00035

mpedm's People

Contributors

Stargazers

Watchers

Forkers

lgtm-migrator

mpedm's Issues

Packaging for production run on ABCI

Package cuEDM as a docker or singularity container for easier deployment on ABCI.

Set thread affinity

Bind OpenMP threads to cores and see if it improves performance: https://www.openmp.org/spec-html/5.0/openmpse52.html

Implement benchmark for cross mapping on CPU

Implement GPU-based cross mapping

Add tests for Simplex projection

Investigate mysterious SGE error on ITC cluster

The jobs @boom10899 is running on ITC occasionally hang with the following mysterious error:

execution daemon on host XXX didn't accept task

Is this a problem with our application or a problem in SGE?
https://stackoverflow.com/questions/14775451/submitting-open-mpi-jobs-to-sge

Example test

Hi @keichi, awesome implementation! Could you please provide some example on how to run the tool? For example a Jupyter notebook tutorial or something along that line? Thank you.

Prediction accuracy degrades if library size and prediction size are different

Test GPU kernels in simplex_test and cross_mapping_test

Add tests for k-NN calculation

Change binary name

cross_mapping_bench and cross_mapping_mpi_bench are apps rather than just benchmarks. We should change the binary name to something like edm-cm, edm-xmap, edm-crossmap, etc.

Add option to transpose the input dataset

Add CLI options to select the number of CPUs and GPUs to use

Add option to select the dataset in an HDF5 file

Write out cross mapping results from cross_mapping_mpi_bench

Use parallel HDF5 to write from the workers concurrently.

Investigate MPI scalability problem

Fix yet another scalability bug Boom has found on ABCI. Initial profiling results indicate that MPI-IO routine called from the workers is the bottleneck.