Pushing the Performance Limits of Datacenter Networks With Fine-Grained Priority Assignment

Abstract

Priority plays a crucial role in distinguishing the diverse demands of applications and improving the performance of underlying networks. However, application-provided priority is insufficient to cope with wide variations in traffic characteristics and dynamic network conditions. When assigning priorities to flows, existing proposals are limited to constrained dimensions, rendering them inadequate for accurately and rapidly identifying flow importance. To address this problem, we present Firapam, a novel priority assignment strategy that systematically combines important traffic states to achieve fine-grained and dynamic priority assignments. Firapam employs convex optimization to adaptively respond to changes in requirements and the environment, and implements admission control for flow priority in a distributed manner. Consequently, Firapam significantly reduces the flow completion time and deadline miss rate. We analytically and experimentally demonstrate that Firapam can effectively support existing priority assignments with significant performance improvements. Compared to state-of-the-art priority assignment methods, Firapam exhibits a remarkable decrease in deadline miss rate by 14.5% to 52.3%, across diverse traffic patterns. Moreover, it reduces the flow completion tail by 17.9% and ensures a minimum reduction of 72.3% in deadline miss rate under high network loads.