SLearn - A Case for Task Sampling based Learning for Cluster Job Scheduling
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
SLearn proposes learning in space by running a small, high-priority sample of tasks per job to estimate average task duration and total job size, then scheduling remaining tasks via a generic GS scheduler. Experiments use 3 production traces (Two Sigma, two Google traces) plus a 150-node Azure testbed and show JCT speedups of 1.28x, 1.56x, and 1.32x over the history-based Three Sigma predictor; Three Sigma also reports ≥100% error on ~23% of jobs. An adaptive sampling controller typically settles near 2% sampled tasks; a DAG extension yields ~1.26x improvement.
AI-written summary · about
chapters (16)
- 0:00Authors Introduction
- 0:35Challenges in Cluster Scheduling
- 1:12Learning Runtime Properties for Cluster Scheduling
- 2:07History-based Learning: Assumptions and Reality
- 2:50Poor Performance of the State-of-the-Art History-based Predictor
- 3:10SLearn: A Novel Approach for Learning Runtime Properties
- 5:24Learning in Time (History) vs Learning in Space (SLearn)
- 6:28Comparing Prediction Accuracy: Large Scale Trace-based Analysis
- 6:54Comparing Coefficients of Variations (CoVs) across Space and Time
- 8:03Varying the History Length in CoV comparison
- 8:52Comparing Prediction Overhead: Simulation and Testbed Experiments Using GS
- 9:45SLearn's Implementation and Design
- 11:57Baselines and Experimental Setup
- 12:43Simulation and Testbed Experimental Results
- 14:35SLearn for DAG and Future Work
- 15:06SLearn Summary
related talks
16mCASSINI: Network-Aware Job Scheduling in Machine Learning Clusters
14mLightning Talk: Cognitive and SelfAdaptive System for Distributed Tracing
25mGreen AI with JobRunr: Sustainable Training and Inference on the JVM
19mAn Experimentation and Analytics Framework for Large-Scale AI Operations Platforms
14mPollux: Co-adaptive Cluster Scheduling for Goodput-Optimized Deep Learning
18mShockwave: Fair and Efficient Cluster Scheduling for Dynamic Adaptation in Machine Learning
56mAutomatically Scalable Computation
34mScaling Scikit Learn