RoCE BALBOA: Service-Enhanced RDMA Offload Engine for Data Center SmartNICs
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
An open-source RDMA offload engine for FPGAs that implements RoCEv2 and sustains 100 Gbps line rate on an Alveo U55C, matching commercial RNICs while keeping resource use similar to TCP stacks. Architecture uses decoupled state and streaming control–data separation to support hundreds of Queue Pairs and per-flow offloads. Demonstrations include an on-path AES-CTR encryptor that saturates wire speed and an ML-based deep packet inspection that runs at line rate with 98% accuracy, plus a GPU-facing preprocessing pipeline.
AI-written summary · about
chapters (10)
- 0:00RoCE BALBOA intro
- 1:02Why RDMA dominates
- 2:27Design pillars
- 4:13Architecture overview
- 7:22Performance evaluation
- 9:31Latency and scalability
- 11:16Programmability options
- 13:31Protocol enhancements
- 15:19ML packet inspection
- 16:43Conclusion
related talks
21mScaling RoCE Networks for AI Training
15mFlor: An Open High Performance RDMA Framework Over Heterogeneous RNICs
22mMetaRoCE: Meta’s RDMA Transport
16mMetaRoCE: From Spec to NIC to Open Source
19mMetaRoCE: From Spec to NIC to Open Source
20mMetaRoCE: Meta’s RDMA Transport
25mRDMA @Scale
20mRDMA at Cloud Scale: The OCI Experience
