https://github.com/mre/cudampi

Large hybrid CPU/GPU sorting network using CUDA and MPI
https://github.com/mre/cudampi

algorithms bucket bucketsort cuda filesystem gpu hybrid-cpu mpi parallel sorting-network

Last synced: 26 days ago
JSON representation

Large hybrid CPU/GPU sorting network using CUDA and MPI

• Host: GitHub
• URL: https://github.com/mre/cudampi
• Owner: mre
• License: lgpl-3.0
• Created: 2013-10-10T17:20:44.000Z (about 11 years ago)
• Default Branch: master
• Last Pushed: 2013-10-10T20:43:41.000Z (about 11 years ago)
• Last Synced: 2024-10-11T09:48:28.698Z (about 1 month ago)
• Topics: algorithms, bucket, bucketsort, cuda, filesystem, gpu, hybrid-cpu, mpi, parallel, sorting-network
• Language: C++
• Size: 1.28 MB
• Stars: 4
• Watchers: 3
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
CUDAMPI

=======

A large hybrid CPU/GPU sorting network using CUDA and MPI.

The sorting network uses a standard Quicksort for CPUs and a custom

Bitonic Sort for GPUs. These two algorithms were the fastest

in a number of prior benchmarks.

We execute the first step of a bucketsort algorithm to presort the data.

A bucket only contains numbers in a given range.

We put each number into its corresponding bucket. This can be done in parallel.

Now each bucket can be sorted on either a CPU or a GPU.

The sorting network uses the filesystem as a process management solution.

Therefore no explicit locks are required.

MPI/IO is used to write to the filesystem in parallel.

The result is a number of sorted files.