aluminum

High-performance, GPU-aware communication library

https://github.com/llnl/aluminum

Science Score: 72.0%

This score indicates how likely this project is to be science-related based on various indicators:

  • CITATION.cff file
    Found CITATION.cff file
  • codemeta.json file
    Found codemeta.json file
  • .zenodo.json file
    Found .zenodo.json file
  • DOI references
  • Academic publication links
    Links to: ieee.org
  • Committers with academic emails
    8 of 10 committers (80.0%) from academic institutions
  • Institutional organization owner
    Organization llnl has institutional domain (software.llnl.gov)
  • JOSS paper metadata
  • Scientific vocabulary similarity
    Low similarity (11.7%) to scientific vocabulary

Keywords

cpp cuda gpu hpc mpi
Last synced: 6 months ago · JSON representation ·

Repository

High-performance, GPU-aware communication library

Basic Info
Statistics
  • Stars: 86
  • Watchers: 6
  • Forks: 22
  • Open Issues: 3
  • Releases: 21
Topics
cpp cuda gpu hpc mpi
Created over 7 years ago · Last pushed about 1 year ago
Metadata Files
Readme License Citation

README.md

Al Aluminum

Aluminum is a high-performance communication library for CPUs, GPUs, and other accelerator platforms. It leverages existing libraries, such as MPI, NCCL, and RCCL, plus its own infrastructure, to deliver performance and accelerator-centric communication.

Aluminum is open-source and maintained by the Lawrence Livermore National Laboratory. If you use Aluminum, please cite our paper: @inproceedings{dryden2018aluminum, title={Aluminum: An Asynchronous, {GPU}-Aware Communication Library Optimized for Large-Scale Training of Deep Neural Networks on {HPC} Systems}, author={Dryden, Nikoli and Maruyama, Naoya and Moon, Tim and Benson, Tom and Yoo, Andy and Snir, Marc and Van Essen, Brian}, booktitle={Proceedings of the Workshop on Machine Learning in HPC Environments (MLHPC)}, year={2018} }

Features

  • Support for blocking and non-blocking collective and point-to-point operations
  • Accelerator-centric communication
  • Supported communication backends:
    • MPI: Uses the Message Passing Interface and supports any hardware your underlying MPI library supports.
    • NCCL: Uses either Nvidia's NCCL library for Nvidia GPUs or AMD's RCCL library for AMD GPUs.
    • HostTransfer: Uses MPI plus the CUDA or HIP runtime to support Nvidia or AMD GPUs without specialized libraries.

Getting Started

For full details, see the Aluminum documentation.

For basic usage examples, see the examples.

Building and Installation

Aluminum is available via Spack or can be installed manually from source.

Source builds need a recent CMake, C++ compiler (with support for C++17), MPI, and hwloc. Accelerator backends need the appropriate runtime libraries.

A basic out-of-source build can be done with mkdir build && cd build cmake /path/to/Aluminum/source

For full details on building, configuration, testing, and benchmarking, see the documentation.

Authors

See also contributors.

License

Aluminum is licensed under the Apache License, Version 2.0. See LICENSE for details.

Owner

  • Name: Lawrence Livermore National Laboratory
  • Login: LLNL
  • Kind: organization
  • Email: github-admin@llnl.gov
  • Location: Livermore, CA, USA

For over 70 years, the Lawrence Livermore National Laboratory has applied science and technology to make the world a safer place.

Citation (CITATION.cff)

cff-version: 1.2.0
title: "Aluminum Communication Library"
message: "If you use Aluminum, please cite it as"
authors:
  - family-names: Dryden
    given-names: Nikoli
  - family-names: Maruyama
    given-names: Naoya
  - family-names: Moon
    given-names: Tim
  - family-names: Benson
    given-names: Tom
  - family-names: Yoo
    given-names: Andy
  - family-names: Van Essen
    given-names: Brian
  - family-names: McNeish
    given-names: Corey
  - family-names: Snir
    given-names: Marc
preferred-citation:
  title: "Aluminum: An Asynchronous, GPU-Aware Communication Library Optimized for Large-Scale Training of Deep Neural Networks on HPC Systems"
  year: "2018"
  type: conference-paper
  collection-title: "Proceedings of the Workshop on Machine Learning in HPC Environments"
  authors:
    - family-names: Dryden
      given-names: Nikoli
    - family-names: Maruyama
      given-names: Naoya
    - family-names: Moon
      given-names: Tim
    - family-names: Benson
      given-names: Tom
    - family-names: Yoo
      given-names: Andy
    - family-names: Snir
      given-names: Marc
    - family-names: Van Essen
      given-names: Brian

GitHub Events

Total
  • Watch event: 2
  • Delete event: 2
  • Issue comment event: 1
  • Push event: 5
  • Pull request review event: 2
  • Pull request event: 4
  • Fork event: 2
  • Create event: 2
Last Year
  • Watch event: 2
  • Delete event: 2
  • Issue comment event: 1
  • Push event: 5
  • Pull request review event: 2
  • Pull request event: 4
  • Fork event: 2
  • Create event: 2

Committers

Last synced: 9 months ago

All Time
  • Total Commits: 602
  • Total Committers: 10
  • Avg Commits per committer: 60.2
  • Development Distribution Score (DDS): 0.354
Past Year
  • Commits: 12
  • Committers: 2
  • Avg Commits per committer: 6.0
  • Development Distribution Score (DDS): 0.167
Top Committers
Name Email Commits
Nikoli Dryden n****n 389
Thomas R. Benson b****1@l****v 124
Naoya Maruyama m****3@l****v 58
Byung Suk Yoo a****o@s****v 19
Brian C. Van Essen v****1@l****v 4
Tim Moon m****3@l****v 4
andy-yoo a****o@l****v 1
NOBLES5E a****n@m****m 1
Byung Suk Yoo a****o@r****v 1
Byung Suk Yoo a****o@r****v 1
Committer Domains (Top 20 + Academic)

Issues and Pull Requests

Last synced: 6 months ago

All Time
  • Total issues: 36
  • Total pull requests: 103
  • Average time to close issues: 9 months
  • Average time to close pull requests: 22 days
  • Total issue authors: 8
  • Total pull request authors: 3
  • Average comments per issue: 0.64
  • Average comments per pull request: 0.18
  • Merged pull requests: 99
  • Bot issues: 0
  • Bot pull requests: 0
Past Year
  • Issues: 0
  • Pull requests: 3
  • Average time to close issues: N/A
  • Average time to close pull requests: about 4 hours
  • Issue authors: 0
  • Pull request authors: 1
  • Average comments per issue: 0
  • Average comments per pull request: 0.67
  • Merged pull requests: 3
  • Bot issues: 0
  • Bot pull requests: 0
Top Authors
Issue Authors
  • ndryden (22)
  • naoyam (3)
  • yurivict (3)
  • benson31 (3)
  • hanfluid (2)
  • lxz12 (1)
  • dmcdougall (1)
  • tbennun (1)
Pull Request Authors
  • ndryden (78)
  • benson31 (32)
  • NOBLES5E (2)
Top Labels
Issue Labels
enhancement (16) bug (9) HIP/ROCm (1) question (1)
Pull Request Labels
enhancement (62) bug (29) HIP/ROCm (16)

Packages

  • Total packages: 1
  • Total downloads: unknown
  • Total dependent packages: 3
  • Total dependent repositories: 0
  • Total versions: 12
  • Total maintainers: 2
spack.io: aluminum

Aluminum provides a generic interface to high-performance communication libraries, with a focus on allreduce algorithms. Blocking and non- blocking algorithms and GPU-aware algorithms are supported. Aluminum also contains custom implementations of select algorithms to optimize for certain situations.

  • Versions: 12
  • Dependent Packages: 3
  • Dependent Repositories: 0
Rankings
Dependent repos count: 0.0%
Average: 13.8%
Dependent packages count: 14.2%
Forks count: 20.5%
Stargazers count: 20.6%
Maintainers (2)
Last synced: 6 months ago

Dependencies

docs/requirements.txt pypi
  • breathe ==4.35.0
  • sphinx ==6.1.3
  • sphinx-rtd-theme ==1.2.0