Karma: Resource Allocation for Dynamic Demands
summary
Karma introduces credits that track donated and borrowed slices and prioritizes borrowers with the most credits to balance allocations over time, proving Pareto efficiency and online strategy-proofness for dynamic demands without future knowledge. Built as a prototype on Jiffy and evaluated with Snowflake traces, it handles workloads where 40–70% of users show demand SD >50% of the mean, halves guaranteed fair share as a tunable parameter, and cuts throughput disparity while preserving system-wide throughput.
AI-written summary · about
chapters (6)
- 0:00Resource allocation basics
- 1:14Current allocation methods
- 2:26Dynamic demand challenges
- 6:12Introducing the Karma algorithm
- 11:24Theoretical analysis
- 14:17Evaluation and conclusion
related talks
05mDoughnut Dilemma: A Lesson in Resource Managers
17mApplication of Online Resource Allocation Based on Ray
19mLook Before, Before You Leap: Online Vector Load Balancing with Few Reassignments
21mSvalinn: Overload Control in Large-Scale Servers with Multiple Resource Bottlenecks
41mYARN Resource Management and Dynamic Max
25mAvoiding Overload in Distributed Systems
20mBetter Resilience Through Self Tuning
34mGIFT: A Coupon Based Throttle-and-Reward Mechanism for Fair and Efficient I/O Bandwidth