BURST: Seeking High-performance, Interoperability and Scalability in Soft-RDMA
summary
BURST is a user-space RDMA stack that implements the Verbs API, uses a DPDK data plane, Intel DSA for offloaded memcpy, and a kernel-assistant to mmap app memory for true zero-copy; it adds a kernel-bypass connection manager and a multi-producer single-consumer doorbell queue. Claims include 98.7% line-rate on 400G, 3.2-6.3x throughput versus kernel RXE, LLM inference latency at 25.2% of TCP, and a 12x faster connection setup versus native RDMA CM.
AI-written summary · about
chapters (15)
- 0:00BURST introduction
- 1:03Hardware diversity in data centers
- 2:09NR-to-R interoperability problem
- 4:21NR-to-R challenges
- 5:46BURST design goals
- 6:50Limitations of software RDMA
- 7:36BURST architecture
- 8:41Accessing RDMA resources without kernel copies
- 10:11Fast work request notifications
- 11:36Zero-copy data plane design
- 12:39User space connection manager
- 13:42Congestion control in BURST
- 15:08Performance evaluation
- 16:13AI workload deployment
- 17:19Conclusion & Q&A
related talks
14mUnderstanding RDMA Microarchitecture Resources for Performance Isolation
18mRDMA is Turing complete, we just did not know it yet!
17mWhite-Boxing RDMA with Packet-Granular Software Control
15mFlor: An Open High Performance RDMA Framework Over Heterogeneous RNICs
28mPersistent Memory plus RDMA, new age remote device
20mPeRF: Preemption-enabled RDMA Framework
23mMulti-Path Transport for RDMA in Datacenters
12mPreview: Networking
