FLB: Fine-grained Load Balancing for Lossless Datacenter Networks
summary
FLB claims to balance RoCE traffic by combining threshold-free rerouting for uncongested flows with congested-flow isolation when queues near PFC thresholds, implemented on a P4/DPDK programmable-switch testbed. It measures one-way path delay via packet timestamps to avoid reordering and marks isolated flows before PFC fires (isolation threshold < PFC threshold). Prototype shows up to 96% reduction in PFC PAUSEs, up to 45% goodput gain over CONGA+DCQCN, and 18–40+% average FCT reductions vs several baselines.
AI-written summary · about
chapters (9)
- 0:00FLB introduction
- 2:52Existing load balancing limits
- 7:13FLB design goals
- 8:18Routing in FLB
- 11:28Measuring one-way delay
- 13:20Isolating congestive flows
- 14:02Implementation setup
- 15:08Performance results
- 15:51Q&A
related talks
20mFLAIR: Accelerating Reads with Consistency Aware Network Routing
17mRe architecting Congestion Management in Lossless Ethernet
16mBackpressure Flow Control
15mFlor: An Open High Performance RDMA Framework Over Heterogeneous RNICs
18mRoCE BALBOA: Service-Enhanced RDMA Offload Engine for Data Center SmartNICs
16mFlow Scheduling with Imprecise Knowledge
19mTackling DC Congestion and Bursts
25mRDMA @Scale
