How Scale Makes Distributed Systems Slower
summary
Asserting that adding machines often slows things down, the talk shows experiments where graph algorithms perform worse when moved from one heavy node to many, attributes the effect to Scale-Originated Coordination Overhead, and identifies transition points (n>1 CPUs, n>1 machines) where coordination cost appears. Demonstrations include a Rust + rayon parallelization that works because addition is commutative, the 50% waiting-at-traffic-lights example from CALM, and a path from Dedalus/Datalog/Bloom/Bud to DDlog/differential dataflow as practical ways to avoid unnecessary coordination.
AI-written summary · about
chapters (10)
- 0:00Intro
- 2:28Scalability! But at what COST?
- 8:07Scale-Originated Coordination Overhead (SOCO)
- 15:12What have we learned so far?
- 15:34Where do we go from here?
- 15:47Keeping CALM: When Distributed Consistency is Easy
- 26:37Dedalus: Datalog in Time & Space
- 30:38From Datalog to Dataflow
- 36:51Conclusion
- 38:06Outro
related talks
32mAn Uncoordinated Approach To Scaling (Almost) Infinitely
43mThe Math of Scalability
22mWhy Are Distributed Systems so Hard?
21mBeyond the Theory: What Actually Breaks When You Scale Your Disaggregat...
27mDistributed Systems, Like It or Not
29mScaleCheck: A Single-Machine Approach for Discovering Scalability Bugs...
25mAvoiding Overload in Distributed Systems
33mWhy Are Distributed Systems So Hard?