Adopting gRPC for Low Latency, High Bandwidth Scenarios
summary
A custom layer negotiates upgrades over standard gRPC then switches eligible connections to fast transports using a client-side interceptor and a generated server dispatch table, avoiding API or protobuf changes. Shared memory achieves about 2 microsecond latency and RDMA about 20 microsecond RDMA with streaming and unary support, using preallocated messages and zero-copy buffers (well‑known array types) plus safe fallback to plain gRPC for security or unavailable paths.
AI-written summary · about
chapters (11)
- 0:00Introduction to gRPC adoption
- 3:10Current challenges with custom protocols
- 5:56Exploring alternatives to gRPC
- 7:18Performance goals and baseline benchmarks
- 9:45Custom layer for shared memory
- 11:51Optimizing the server-side event loop
- 13:57Achieving zero-copy serialization
- 16:03Challenges in gRPC architecture
- 18:10Resiliency and security in zero-copy
- 21:00Future plans and open source potential
- 24:44Q&A session
related talks
22mA Case Study of Enhancing gRPC Micro-Services for Supporting High Throughput, Low Latency Trading
30mgRPC Meets RDMA
24mHow We Built Protosockets to Go Beyond HTTP and gRPC
28mFrom Backend Powerhouses to Web-Friendly APIs: Unlocking
17mBoost API Performance: gRPC vs REST vs Async
10mConsiderations for a large-scale low-latency system
01h 01mBuilding Microservice Communication with gRPC
31mPatterns of Low Latency