Transfer Protocol Design And Task Scheduling Optimization On Network Service System

As the technology of computer networks is evolving,more and more high-speed long-distance networks have been deployed into the Internet.The traditional standard TCP has been dominant for a long time,but the additive increase and multiplicative de-crease congestion control strategy used in standard TCP shows no scalability,and has gradually become the bottleneck in data transmission over long fast networks.Mean-while,standard TCP employs Fast Recovery to accelerate packet loss recovery,but the tail loss phenomenon leads to a fact that a large proportion of packet losses are still recovered through the expensive retransmission timeouts,which bring severe latency to network applications.At the same time,the performance of task scheduling on net-work service systems also has a great impact on the response quality of Internet-based applications.How to design transfer protocols that have strong capability of bandwidth utilization,mechanisms that can reduce the network latency and scheduling strategies that make service tasks response fast,has become a hot research topic in computer sci-ence.This dissertation focuses on two core topics:the transfer protocol design methods and the task scheduling optimization strategies on network service systems.The main work and contributions are as follows;1)To support scalable data transfer over high-speed long-distance networks,this dissertation proposes a UDP-based application-layer high-speed transfer protocol named DCUDP.To control packet sending rate,DCUDP proposes a novel rate-control algorith-m.called Double Cubic,which employs cubic-shaped rate growth to improve protocol scalability and transfer stability.To deal with packet loss,DCUDP introduces Random Loss Decrease algorithm,in which the random reduction factors and reduction times can effectively alleviate the impacts of global loss synchronization and continuous loss.Furthermore,to make DCUDP suitable for real networks,the core parameters in the algorithms are set according to deterministic loss model.2)Packet loss can hardly be used to predict congestion trend in standard TCP.This dissertation proposes a delay-based high-speed TCP variant,namely DFTCP,which us-es queueing delay in addition to packet loss,as congestion feedbacks to assist congestion window adjustment.The main idea of this algorithm is to estimate the maximum size of the available buffers in the routers along the path of a network flow,and then calculate a reasonable equilibrium point accordingly for this flow.DFTCP attempts to stabilize the congestion window around the equilibrium point in order to improve bandwidth utilization and reduce packet losses at the same time.3)The interactive performance of Internet-based applications are severely influ-enced by TCP tail losses.To address this issue,this dissertation proposes an efficient universal tail loss recovery algorithm called DTLP.DTLP classifies tail losses into dif-ferent patterns and transmits different types of probes for different loss patterns to trigger different TCP mechanisms,which effectively alleviates the network latency caused by tail losses.4)The spinlocks employed in Linux scheduler lead to severe lock contention on multi-core network service systems.This dissertation proposes three different optimiza-tion strategies on three different levels to decrease lock contention.The first strategy is to shrink the critical sections from code level.The second strategy introduces a new method,which delays the wake-up operations of new processes that are blocked by spin-locks and puts them into lock-free delayed-task queues.The third strategy is tuning the key scheduling parameters according to the process features of top-level applications and takes HTTP service system as an example.5)The kernel-mode scheduler in multi-core Linux network service systems can not fulfill the requirements of real-time scheduling.This dissertation designs and imple-ments a lightweight user-mode real-time multi-task scheduling framework,namely U-Light,based on the multi-core architecture.ULight supports priority-based preemptive lightweight scheduling,high-resolution user-space timing service,and efficient user-mode interrupt handling,so that it provides a new thinking for the design of real-time system and network service system.