NetRPC: Enabling In-Network Computation in Remote Procedure Calls
summary
NetRPC implements an INC-enabled RPC stack that exposes protobuf/grpc interfaces and a JSON net filter to declare primitives like add on map and confirm-and-forward, compiles switch logic, and runs on two programmable switches + eight GPU servers. It uses a flip bit dedup for reliability and a fallback-to-host migration when switch memory is insufficient. Experiments show up to 97% lines-of-code reduction, similar latency/throughput to prior INC solutions, and about 12% higher training throughput than a P4 baseline.
AI-written summary · about
chapters (8)
- 0:00NetRPC introduction and goals
- 3:33In-network computation challenges
- 5:40User-friendly NetRPC interfaces
- 8:14Concurrent in-network computation
- 10:20Failure handling and fallback
- 11:24Switch memory management
- 13:12NetRPC performance evaluation
- 15:21Conclusion and future work
related talks
16mHigh-level Programming for Application Networks
37mgRPC Intro
24mDatacenter RPCs can be General and Fast
14mODRP: On-Demand Remote Paging with Programmable RDMA
21mR2P2: Making RPCs first-class datacenter citizens
16mNetVRM: Virtual Register Memory for Programmable Networks
12mPreview: Networking
13mRakaia: Scalable In-Kernel Scheduling for TCP-Based RPCs