STRADS-AP: Simplifying Distributed Machine Learning Programming without Introducing a New Programming Model
Talkat USENIX ATC '19 · on Wednesday, 2019-07-10· at 15:00· USENIX ATC Track II: Grand Ballroom VII–IX·
summary
STRADS-AP automatically parallelizes ordinary sequential ML code so that converting published algorithms to a cluster requires few mechanical changes rather than a new programming model. It implements automatic data partitioning, task parallelization, concurrency control and fault tolerance inside a runtime that compiles native C++ kernels and schedules work across nodes, and it reports faster end-to-end training than a popular data-parallel framework while achieving performance comparable to ML-specialized frameworks.
AI-written summary · about
chapters (7)
- 0:00STRADS-AP introduction
- 2:02Distributed ML challenges
- 5:48STRADS-AP programming
- 9:07Performance and debugging
- 12:31Case studies
- 15:53STRADS-AP vs others
- 18:30Q&A
related talks
52mC++ Standard Parallelism
01h 03mScaling up machine learning without tears (and what do programming languages ha...
01h 07mGeneric Parallel Programming
23mA Haskell Auto-Parallelizer for Distributed Computing
01h 07mThe Asynchronous C++ Parallel Programming Model
14mCompressed and Parallelized Structured Tensor Algebra
11mScaling Distributed Machine Learning with In-Network Aggregation
45mAccelerate Model Training with an Easy to Use High-Performance...