Tags: A Framework for Distributed Event Ordering
summary
Proposes Tags, an abstract token-credit model for reconfigurable dataflow accelerators that concretely defines publish/update/ready functions and per-edge token accounts to encode synchronization. Demonstrates a compiler-mappable prototype that implements boundary desynchronization for iterative ML, sending only an iteration integer and reduces communication by factor of Delta, trading statistical for hardware efficiency. Claims Tags can serve as an IR between session types and hardware and enable compiler checks for deadlock and correctness on massively decentralized RDAs.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 0:53Why Memory Modeling
- 1:45Why a New Memory Modeling Framework
- 4:09Examples
- 8:39Future Work
related talks
48mCRDTs: From sequential to concurrent executions
01h 30mImplementing, verifying and debugging distributed event-based systems
43mAn Approach For The Internet Of Things
28mNew insights into distributed and concurrent programming
10mLightning Talk: Not All Tokens Are Equal: Semantic KV-Cache for Agen...
22mEnabling Refinable Cross-Host Attack Investigation with Efficient Data Flow Tagging and Tracking
41mLessons Learned Implementing a Resilient, Distributed, Real-Time Architecture
22mOptimizing network topologies...