Download presentation
Presentation is loading. Please wait.
Published byVincent Barrett Modified over 9 years ago
1
1 © 2014 The MathWorks, Inc. HPC with MATLAB Making parallel programming simple Jos Martin, Principal Architect, Parallel Computing Tools jos.martin@mathworks.co.uk
2
2 Attributes of MATLAB 1. Mathematically correct 2. Usable 3. Bug-free 4. Fast
3
3 Some Benchmarks HPLFFTEP-Stream Score (Lower is better) 366 Implementation o = A\b;o = fft(v);o = a.*b + c;
4
4 Science (even in HPC) is about the Maths Don’t make it hard to program –Make expressing parallelism easy – parfor, distributed arrays, spmd, gpuArray, batch Make it easy to try out –Local cluster with Parallel Computing Toolbox –Scale out to cluster with no code changes
5
5 “Simple to use” vs. “Lots of control” Level of ControlCPU ParallelGPU Simplebuilt-in to toolboxes gpuArray, associated maths Intermediate parfor, distributed arrays, batch arrayfun(@fun, …) Detailed spmd, jobs and tasks direct integration with CUDA kernels
6
6 Our Users Design and tuning of the control system for the International Linear Collider –Queen Mary College, London Neural Network Design for Matching Heart Transplant Donors with Recipients –Lund University, Sweden Tomographic Reconstruction of Protein Structure –Max Planck Institute of Biochemistry, Germany
7
7 Building and Using Clusters
8
8
9
9 Thank You
Similar presentations
© 2024 SlidePlayer.com Inc.
All rights reserved.