Effective Straggler Mitigation: Attack of the Clones
summary
Builds a system called Dolly that proactively launches task‑level clones for small interactive jobs, then uses a probabilistic delay assignment to avoid intermediate‑data contention. Evaluation on Facebook and Bing traces on a 150‑node testbed shows a 5% extra resources cloning budget yields 34–46% speedups for small jobs, drops the worst task from ~8× slower than the median to only slightly slower, and beats both exclusive‑copy and contention‑heavy extremes.
AI-written summary · about
chapters (5)
- 0:00Start
- 8:01Task Level Cloning
- 10:02Contention Cloning
- 12:25Delay Assignment
- 14:04Average Completion Time
related talks
16mUnderstanding Stragglers in Large Model Training Using What-if Analysis
39mAsynchronous Processing for Fun and Profity
16mDisaggregation and the Application
25mAvoiding Overload in Distributed Systems
19mLook Before, Before You Leap: Online Vector Load Balancing with Few Reassignments
17mAttack of the Bubbles: Straggler-Resilient Pipeline Parallelism for Large Model Training
38mRecognizing Zombies, Black Holes & Tribbles Before You Get Eaten
24mAverting crisis at multi-petabyte scale