TCP Optimization Technologies For Data Center Networks

Data Center is the infrastructure of modern Internet,hosting many distributed big data processing tools to provide various Internet services.These tools typically adopt the Partition/Aggregation pattern to process data,generating a mass of TCP flows in Data Center Networks(DCN).Accelerating TCP transmission can effectively improve the response time of applications,increasing revenue.However,TCP in DCN faces serious challenges of throughput and latency.The multiple paths between a pair of hosts are not fully utilized and bandwidth are wasted.The actual flow completion times(FCT)of latency-sensitive short flow can be 1000 times of its theoretical FCT.To address these problems,many TCP optimization technologies have been proposed.In network layer,multipath load balancing schemes improve the utilization rate of multipath bandwidth.In transmission layer,rate control schemes reduce the latency caused by queuing;flow scheduling schemes make TCP flows complete one by one according to their priorities,minimizing FCT by prioritizing short flows.Although existing works improve the performance of DCN,there are still some unsolved problems.The explicit routing control schemes supporting multipath load balancing suffer from low scalability.The TCP schemes designed for single path do not accommodate the underlying multipath load balancing schemes,resulting in low throughput.The switch-based flow scheduling scheme is costly to deploy with hardware modifications on switches.To address these problems and improve TCP's performance on throughput and latency,we research TCP optimization technologies for DCN in this thesis.The main contributions are outlined as follows.1.An ECMP-based scalable explicit routing control schemeExplicit routing control is a basic network function for multipath load balancing.The state-of-the-art approach Xpath does not scale well due to limited in-switch route entries.This thesis presents an ECMP-based scalable explicit routing control scheme.We observe that in an ECMP(Equal Cost Multi-Path)configured network,the in-packet key for hashing determines which path the packet will be routed to,so we can forge a key to route a packet to the specified path pointed by the key.SECMP proactively finds the key-to-path mappings for all paths through a path cover testing,then rewrites packets with these keys at the network edge to explicitly route them to specified paths.Our simulation results show that SECMP outperforms Xpath 8× in scalability.Our testbed implementation and Mininet experiment show that SECMP is ready to be deployed in commodity network or OpenFlow network to improve multipath load balancing.2.A lazy TCP congestion control algorithm based on packet spraying routingPacket spraying is a simple and efficient multipath load balancing scheme in DCN for full utilization of multiple paths.However,existing TCP schemes face serious challenge of throughput when running over packet spraying.They suffer severely from frequent packet reordering and link degradation,resulting in low throughput.To address these issues,in this thesis,we propose Lazy TCP(LTCP),a novel TCP over packet spraying to achieve stable transmission and maintain high throughput.LTCP performs congestion avoidance in a lazy way.First,to mitigate the impact of packet reordering,LTCP delays the time to perform congestion avoidance.As the sender waits more time for the disordered packet,LTCP avoids unnecessary backoff caused by packet reordering.Second,to eliminate the impact of link degradation,LTCP stubbornly maintains data transmission rate unless detecting successive packet loss.Simulations on NS2 show that LTCP is stable and high-throughput,and improves 2.4 times throughput compared to existing TCP in cases of packet reordering and link degradation.3.A server-based flow scheduling schemeMinimizing FCT is a critical issue for low-latency transmission in DCN.The state-ofthe-art approach pFabric is costly to deploy with hardware modification on switches.In this thesis,we present a server-based flow scheduling(SFS)scheme for enabling easy and rapid deployment in servers while almost retaining the same minimal FCT as pFabric.SFS combines traffic control and flow scheduling to keep in-switch queue very short,hence switches do not need to schedule flows and such function can be moved to servers.We propose two novel techniques to achieve SFS.First,the bidirectional flow scheduling technique is used for each server to locally schedule sent or received flows.Second,the most recently seen flow coordination technique is used to coordinate senders and receivers to address the priority disagreement problem.Experimental results on NS2 show that SFS approaches pFabric in minimizing FCT.