Mars-on-Ray: Accelerating Large-scale Tensor and DataFrame Workloads
summary
Mars-on-Ray runs Mars (distributed numpy/pandas/scikit-learn APIs) on Ray with an Oscar-based actor runtime, placement groups, and Ray object store spilling; it tiles arrays into chunks, schedules by data locality and depth-first rules, and supports lazy fusion. A nearest-neighbor pipeline indexes and searches 400 million vectors using an approximate library, finishing end-to-end reads/writes and search in 3.5 hours, achieving 4–5x faster throughput than a pure scikit-learn brute-force baseline on clusters up to 12,000 cores.
AI-written summary · about
chapters (7)
- 0:00Introduction to Mars-on-Ray
- 7:48Mars architecture overview
- 13:30Mars API design principles
- 21:56Execution modes in Mars
- 28:03Cluster setup with Ray
- 32:42Live demo of Mars-on-Ray
- 34:06Conclusion and future work
related talks
24mRay: A System for High-performance, Distributed ML Applications
26mUsing Ray On Large-scale Applications at Ant Group
29mDask-on-Ray: Using Dask and Ray to Analyze Petabytes of Remote Sensing Data
24mBuilding Complex Data Analytics Pipelines with Ray
35mOperating a Production-Scale Physical AI Data Platform with Ray
31mFrom OOMs to 2x Throughput: Scaling Multimodal Data Curation on Ray
32mAnyscale's Ray Data: Revolutionizing Batch Inference
03mRay: A Distributed Execution Framework for Emerging AI Applications