IN-NETWORK TECHNIQUES FOR HIGHLY RELIABLE DATACENTER NETWORKS

Abstract

This thesis proposes three in-network techniques to mitigate the increase in tail flow completion times (FCTs) in datacenter networks that are caused by unpredictable congestion (microbursts) and link failures. The proposed techniques are BurstRadar for detecting and characterizing microbursts, Shared Queue Ring (SQR) for managing fail-stop link failures by seamlessly diverting the affected flows to alternative paths, and LinkGuardian for mitigating packet loss due to gray (corruption) link failures. The evaluation results on multi-gigabit hardware testbeds demonstrate that BurstRadar can accurately detect microbursts while incurring 10 times fewer data collection and processing overhead than existing solutions, SQR can completely mask link failures and reduce tail FCT by up to 4 orders of magnitude for latency-sensitive flows, and LinkGuardian can improve the 99.9th percentile FCT for short TCP and RDMA flows by 51x and 66x, respectively.