Overview
Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore a conference presentation that introduces FLB (Fine-grained Load Balancing), a novel scheme designed specifically for lossless datacenter networks using RDMA over Converged Ethernet (RoCE) with Priority Flow Control (PFC). Learn how researchers from Central South University, Hong Kong University of Science and Technology, and other institutions identified critical issues with existing load balancing mechanisms in lossless environments, where traditional fine-grained rerouting schemes can actually worsen congestion by spreading flows across more paths and aggravating PFC's head-of-line blocking. Discover the innovative two-pronged approach of FLB that employs threshold-free rerouting for effective traffic load balancing during normal conditions and implements timely congested flow isolation to eliminate head-of-line blocking on non-congested flows during congestion events. Examine the comprehensive experimental results demonstrating FLB's significant performance improvements, including up to 96% reduction in PFC PAUSE rate, complete elimination of head-of-line blocking, up to 45% improvement in goodput over CONGA+DCQCN, and substantial reductions in average flow completion time compared to various existing solutions like LetFlow+Swift, MP-RDMA, Proteus+DCQCN, and LetFlow+PCN.
Syllabus
USENIX ATC '25 - FLB: Fine-grained Load Balancing for Lossless Datacenter Networks
Taught by
USENIX