Research On Prioritized Fine-Grained Flow Scheduling For Lossless Networks

Nowadays,data centers have become a crucial type of infrastructure.The continuous changes in the requirements and patterns of data center applications have brought significant challenges to data center network traffic control.Among them,for the problem of data center traffic control,prioritized bandwidth allocation has received extensive attention from both cloud service providers and network researchers.And most researchers have focused on traditional data center networks and proposed various network traffic scheduling methods and technical improvements.However,for Remote Direct Memory Access(RDMA)lossless networks,prior techniques mainly rely on priority queues;and due to the limitation of queue numbers,it is quite challenging for prioritized bandwidth allocation schemes to meet the bandwidth allocation requirements in scenarios involving a lot of concurrent applications with priority queues.Based on this,in this thesis,we focus on achieving prioritized bandwidth allocation without relying on priority queues in lossless networks.Our main work is as follows:(1)Based on the research progress of detecting and avoiding deadlocks for Prioritybased Flow Control(PFC),we propose a prioritized flow scheduling design based on the idea of delayed resume.For lossless networks driven by the RDMA over Converged Ethernet Version 2(Ro CEv2)protocol,our solution achieves weighted bandwidth allocations for concurrent flows by introducing delays in the response logic of the sender to the PFC resume frame.Specifically,we propose prioritized bandwidth allocation schemes for 2 and N types of flows,respecting the cases that flows’ sending rates would be fixed and dynamically change with priority,respectively.Simulation tests show that the scheme we proposed could ensure fine-grained prioritized bandwidth allocation for long flows and further achieves the goal of optimizing the average flow completion times.(2)Based on the design adopted by the Data Center Quantized Congestion Notification(DCQCN)protocol,we propose the Weighted DCQCN(WDCQCN)algorithm.By introducing configurable weight values into the logic of how DCQCN decelerates its sending rates,our algorithm can achieve prioritized bandwidth allocation for concurrent flows without relying on priority queues.For the selection of weight value,we analyze the stability of the WDCQCN fluid model to obtain the corresponding relationship between the weight value and the obtained transmission rate,along with the range of the weight value,thus determining reasonable settings for the weights.Simulation tests show that this scheme can achieve prioritized bandwidth allocation for concurrent flows,yielding optimized average flow completion times.