When to Hedge in Interactive Services
summary
Compares hedging vs optimal load balancing, shows hedging often gives no benefit above ~60% utilization and can cause congestion, defines an idealized hedge that never idles workers and prioritizes likely-faster replicas, builds LÆDGE that combines load-aware scheduling with work‑conserving hedging, evaluates on AWS with 10 VMs (five shards × two replicas), uses OLDsim and Lucene on English Wikipedia, and reports up to 49% 99th‑percentile latency improvement while avoiding unnecessary congestion.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:44Online Data Intensive Services
- 3:52What can we do
- 4:55Idealizat
- 8:03Late Hatch
- 8:52Ledge
- 9:54Results
- 11:08Benefits
- 11:48Summary
related talks
47mLoad Shedding—Approaches, Principles, Experiences, and Impact in Service Management
25mAvoiding Overload in Distributed Systems
21mHandling Traffic Spikes Through Automated Prioritized Load Shedding
19mLook Before, Before You Leap: Online Vector Load Balancing with Few Reassignments
19mEfficient LLM Serving on Commodity GPU Clusters with Data-Reduced Cross-Instance...
26mEfficient load balancing in distributed systems
16mTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
35mSystems Scale Summer Q&A Week 1