Disaggregation and the Application
summary
Measurements show running unmodified applications on disaggregated data centers causes over an order of magnitude slowdown, so the paper proposes three OS-level primitives to fix that: memory grant to hand pages to another process without copying, memory steal to let a worker take another’s pages for load balancing or straggler mitigation, and failure informer to notify replicas when a memory node dies. Implementation details cover global virtual address namespaces, crash-consistency techniques for stolen pages, and examples speeding dataflow frameworks and Paxos recovery.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:28Traditional data center racks
- 2:04Why? Many benefits for operators
- 4:23Should you run regular applications on D
- 4:56Key issue: Too much data movement Goal: send data from App 1 to App 2
- 5:52Our position
- 7:21We propose three new OS abstractions
- 7:43Memory grant
- 8:47Properties of Grant
- 10:03Memory steal
- 10:51Properties of Steal
- 12:14Failure informers / Spies
- 13:03Some applications
- 15:36Summary
related talks
45mBeyond Machine Boundaries: Experience with Memory Disaggregation
25mAvoiding Overload in Distributed Systems
17mDuhu: Shared Disaggregated Memory for Distributed Data Processing Frameworks
21mBeyond the Theory: What Actually Breaks When You Scale Your Disaggregat...
42mThymesis-P: An Approach to Rack-scale Disaggregation Over OpenCAPI
31mUnshackling Memory!
16mA Tale of Two Paths: Toward a Hybrid Data Plane for Efficient Far-Memory Applications
20mUniMem: Redesigning Disaggregated Memory within A Unified Local-Remote Memory...