Svalinn: Overload Control in Large-Scale Servers with Multiple Resource Bottlenecks
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
Svalinn uses a credit-based admission controller that runs online micro-experiments to find the credit pool maximizing a user-specified convex utility (throughput, drop rate, latency). Per-bottleneck AQM enforces latency targets for explicit queues, while a new msemaphore primitive uses a multi-arm bandit to pick the minimal concurrency that saturates memory bandwidth without wasting CPU. Implemented in four apps and two runtimes, it raises goodput up to 6.51× while keeping tail latency bounded.
AI-written summary · about
chapters (8)
- 0:00Svalinn introduction
- 3:09Traditional overload limits
- 9:51Svalinn design overview
- 13:19Credit-based admission control
- 14:20Memory bandwidth control
- 17:33Evaluation setup
- 19:41Performance results
- 20:22Summary
related talks
35mSystems Scale Summer Q&A Week 1
40mEverything Will Flow
25mAvoiding Overload in Distributed Systems
26mEfficient load balancing in distributed systems
16mTowards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload...
20mBetter Resilience Through Self Tuning
44mDeep Dive talk | Resilience in Microservices with Congestion Control
01h 01mLittle things and big things: lessons for building large systems.