Pieter Noordhuis

Open-source work

Noordhuis's early open-source work was on Redis: commits from March 2010 include the first implementation of augmented sorted sets and the ZRANK command, and from September 2010 Noordhuis rewrote reply parsing in the hiredis C client library to use a read buffer.⁵⁶

Noordhuis made the initial import of the Gloo collective communications library in February 2017 and in August 2019 created its libuv-based transport. PyTorch's persons-of-interest page lists Noordhuis among the emeritus maintainers of PyTorch's Distributed component.⁷⁸⁹¹

Since September 2022 Noordhuis has contributed to the Databricks CLI repository, with commits continuing into October 2026.¹⁰¹¹

Publications

Noordhuis co-authored "Accurate, Large Minibatch SGD - Training ImageNet in 1 Hour" (2017), which trained ResNet-50 with very large minibatches using a learning-rate scaling rule and warmup, and "PyTorch Distributed - Experiences on Accelerating Data Parallel Training" (VLDB 2020), which describes the design of PyTorch's distributed data parallel module.³²

Work and contributions

• PyTorch Distributed Open-source code

  Listed as an emeritus maintainer of PyTorch's Distributed module, and a co-author of "PyTorch Distributed: Experiences on Accelerating Data Parallel Training" (VLDB 2020), which describes the design of the distributed data parallel module. Noordhuis's GitHub profile pins the pytorch/pytorch repository.¹²¹²

• Gloo Open-source code

  Author of commits to the BSD-licensed Gloo library; those listed from May 2019 to January 2020 include "Create libuv transport" and changes to the TCP transport.⁴⁹

• Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour Published research

  Co-author of the 2017 paper, which trains ResNet-50 on ImageNet with minibatches of up to 8,192 images using a learning-rate scaling rule and warmup.³

Timeline

1. Mar 3, 2010Contributed the initial implementation of augmented sorted sets and the ZRANK command to Redis.⁵
2. Sep 19, 2010Rewrote reply parsing in the hiredis Redis client library to use a read buffer.⁶
3. Feb 9, 2017Made the initial import of the Gloo collective communications library.⁷
4. Jun 8, 2017Co-authored "Accurate, Large Minibatch SGD - Training ImageNet in 1 Hour".³
5. Aug 27, 2019Created the libuv transport for Gloo.⁸
6. Jun 28, 2020Co-authored the PyTorch Distributed data parallel paper, published at VLDB 2020.²
7. Sep 7, 2022Made first commits to the Databricks CLI repository.¹⁰

Documented roles

The sources do not confirm a current role; these are the roles they document.

• Maintainer, Distributed (now listed as emeritus), PyTorch (open-source project)¹

Related people

• Yangqing JiaML frameworks and libraries · Training systems
• William FalconML frameworks and libraries · Training systems
• Brian Van EssenTraining systems · ML frameworks and libraries
• Mu LiML frameworks and libraries · Training systems
• James BradburyML frameworks and libraries · Training systems
• Edward YangML frameworks and libraries · Training systems
• Clément FarabetTraining systems · ML frameworks and libraries
• Adam PaszkeML frameworks and libraries · Training systems
• Carlos GuestrinML frameworks and libraries · Training systems
• Luca AntigaML frameworks and libraries · Training systems