search

ccsb-scripps / AutoDock-GPU

AutoDock for GPUs and other accelerators
https://ccsb.scripps.edu/autodock
GNU General Public License v2.0
392 stars 108 forks source link

MPI parallelization across GPUs of a multi-GPU in AutoDock-GPU #182

Open vlkale opened 2 years ago

vlkale commented 2 years ago

Description of Problem

AutoDock-GPU is a good use case for OpenCL (or OpenMP) parallelization across multiple GPUs (or xPUs) on a node of a supercomputer. We want to know how well an MPI parallelization would work across multiple xPUs, but AFAIK, MPI is not used in AutoDock-GPU.

Solution we Would Like

Is there is an easy way to replace the OpenMP parallelization across GPUs with an MPI parallelization across GPUs (using OpenMP offload or CUDA for parallelization within each GPU)? This would allow AutoDock-GPU to (a) have an alternative to OpenCL/OpenMP parallelization across GPUs and (b) make a AutoDock-GPU a prime use case that could potentially show the benefits of an MPI parallelization over an OpenMP parallelization across GPUs of a node.

Alternatives Considered

One could consider the recent NCCL instead of MPI for parallelization across xPUs. However, NCCL is not as widely supported on supercomputers and small clusters.

carlosfamilia commented 2 years ago

I have found the following paper:

https://pubmed.ncbi.nlm.nih.gov/21527034/

It seems that an MPI version of autodock already exists, though the download page is currently offline:

https://autodock.scripps.edu/downloads/multilevel-parallel-autodock4.2

vlkale commented 2 years ago

Hi Carlos,

Thanks for this.

We are specifically looking for an AutoDock-GPU with MPI, e.g., a version that has both MPI and CUDA, and more specifically for what we are interested in with OpenMP offload, an MPI+OpenMP offload version of AutoDock-GPU.

Having looked through my notes and talked with Mathialakan Thavappiragasam, what you show is an MPI+OpenMP version running on clusters of multi-core processors (CPUs) only, and not clusters of nodes with multi-core processors (CPUs) and GPUs.

Vivek

On Mar 1, 2022, at 17:18, Carlos Familia @.***> wrote:

 I have found the following paper:

https://pubmed.ncbi.nlm.nih.gov/21527034/

It seems that an MPI version of autodock already exists, though the download page is currently offline:

https://autodock.scripps.edu/downloads/multilevel-parallel-autodock4.2

— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.

carlosfamilia commented 2 years ago

Sorry, didn't pay attention to the GPU part of the ticket.