GPGPU use cases from the MoBrain community João Rodrigues Postdoctoral Researcher Utrecht University, NL j.rodrigues@uu.nl
MoBrain main activities Task 1: User support and training Task 2: Cryo-EM in the cloud: bringing clouds to the data Task 3: Increasing the throughput efficiency of WeNMR portals via DIRAC4EGI Task 4: Cloud VMs for structural biology Task 5: GPU portals for biomolecular simulations Task 6: Integrating the micro (WeNMR/INSTRUCT) and macroscopic (NeuGRID4you) VRCs
Software our solutions Powerfit Fitting atomic structures in Cryo-EM density maps using a full exhaustive 6D cross-correlation search based on FFT techniques. DisVis Visualization and quantification of accessible interaction space of distance-restrained protein-protein docking based on FFT techniques. GROMACS Versatile package to perform Molecular Dynamics simulations on systems with hundreds to millions of particles. AMBER Package to perform Molecular Dynamics simulations.
Use Case fitting atomic structures in Cryo-EM density maps Powerfit Fitting atomic structures in Cryo-EM density maps using a full exhaustive 6D cross-correlation search based on FFT techniques.
Software powerfit & disvis Core Dependencies Numpy Cython Scipy https://github.com/haddocking/
Software powerfit & disvis Numpy Cython Scipy Accelerated CPU FFTW3 pyFFTW https://github.com/haddocking/
Software powerfit & disvis OpenCL pyOpenCL clFFT Numpy Cython gpyFFT Scipy GPGPU Acceleration FFTW3 pyFFTW https://github.com/haddocking/
Software powerfit & disvis https://github.com/haddocking/
Use Case MD simulation of a large protein system Ferritin is a protein of 450 kDa, consisting of 24 subunits A MD simulation in explicit solvent involves: more than 4000 amino acids more than 36000 water molecules Total atoms: 176000 Test Simulations were run using AMBER 14, with OpenMPI Performance on 2 GPU K20m: 8.66 ns/day
Software GROMACS & AMBER CUDA 4.x MKL CC & CMake FFTW3
Software GROMACS & AMBER CUDA 4.x GBs of data per day per simulation. MKL CC & CMake FFTW3
Queueing & Middleware resources & requirements Example Hardware Cluster based on 3 Worker Nodes: 2x XEON E5-2620 v2 2x K20m 64 Gb RAM Total 36 CPU core and 6 GPU 192 Gb RAM
Queueing & Middleware resources & requirements OpenMPI Middleware Requirements One Job per GPU (AMBER) CPUs must be powerful to match the GPU CPU is still doing some work (e.g. bonded interactions) Discoverable within the e-infrastructure (e.g. jdl requirement) Preferrably containing GPU type (GTX vs K-series, AMD vs NVIDIA) AMD GPUs not supported by MD code (yet) Double-precision only supported by Tesla cards Torque & Maui
Conclusions & Questions OpenCL Scipy pyOpenCL CUDA CC & CMake Numpy clFFT FFTW3 OpenMPI Torque & Maui gpyFFT Cython pyFFTW MKL Thank you for your attention