Cilantro: Performance-Aware Resource Allocation for General Objectives via Online...
summary
Cilantro is a scheduler that replaces user resource requests with online learning from real-time job feedback to map resources to performance, decouples learners from policies, and applies optimism in the face of uncertainty via confidence bounds. It runs as a cluster manager policy that constrains exploration with a clipped utility model and supports arbitrary objectives. On a 1000-CPU, 20-job tenant cluster it reports 1.2–3.7x utility gains versus nine baselines, and in a 160-CPU, 19-microservice benchmark it cuts end-to-end P99 to 0.57× the next-best baseline.
AI-written summary · about
chapters (12)
- 0:00Cilantro framework intro
- 1:30Demand-based scheduling limits
- 2:35Performance-aware scheduling
- 4:46Cilantro architecture
- 6:52Online learning challenges
- 7:54Utility model for objectives
- 9:18Supported scheduling objectives
- 10:19Uncertainty-aware OFU policies
- 12:22Adapting to uncertainty
- 13:27Microservices evaluation setup
- 14:28Cilantro performance results
- 15:12Summary
related talks
41mEfficient cluster resource management by using Cook and Mesos
39mPractical Container Scheduling: Juggling Optimizations, Guarantees, and Trade-Offs at Netflix
14mPollux: Co-adaptive Cluster Scheduling for Goodput-Optimized Deep Learning
41mEfficient cluster resource management by using Cook and Mesos
14mHeteCCL: Synthesizing Near-Optimal Collective Communication Schedules for Heterogeneous
41mIdle Resources Oversubscription: Improving Cluster Utilization
15mAutoCCL: Automated Collective Communication Tuning for Accelerating Distributed and...
17mApplication of Online Resource Allocation Based on Ray