ATP: In-network Aggregation for Multi-tenant Learning
summary
Claims up to 38–66% training throughput improvement for multi-tenant distributed DNN jobs by moving gradient aggregation into programmable switches. Implements in-network aggregation in P4 on Barefoot Tofino-style hardware, using dynamic per-packet aggregators hashed by gradient index, two-level TOR+aggregation-switch reduction, and a best-effort fallback that forwards missed aggregations to parameter servers. Ensures exactness via sequence-detect retransmit, converts floats to 32-bit scaled integers to avoid switch FP, and evaluates on a cluster against static partitioning and prior systems.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:28Parameter Server Architecture
- 3:06Agenda
- 3:31Aggregation
- 4:41Best Effort
- 7:45Additional Challenges
- 9:46Evaluation
- 10:19Dynamic Allocation
- 10:46Results
- 11:39Summary
related talks
16mAccelerating Distributed Graph Learning by Using Collaborative In-Network...
39mScalable Multi-Node AI Workloads in Multi-Tenant AI Clouds U...
14mBrain-on-Switch: Towards Advanced Intelligent Network Data Plane via NN-Driven Traffic...
19mTowards Domain-Specific Network Transport for Distributed DNN Training
17mZEN: Empowering Distributed Training with Sparsity-driven Data Synchronization
17mTraffic Engineering for AI Training Networks
17mNetHint: White-Box Networking for Multi-Tenant Data Centers
14mMultitenancy for Fast and Programmable Networks in the Cloud
