Key Technologies For Traffic Management Of Programmable Data Plane In Data Center Network

The data center is the infrastructure of application services in the nextgeneration Internet,and improving the performance of data center networks gets the focus from the industrial and academe.Designing efficient and fair mechanisms for load balancing and packet scheduling is the key to optimizing traffic management to improve the transmission performance of data center networks.However,the Partition/Aggregate workflow pattern in data center networks makes the datacenter traffic show the characteristics of high concurrency and strong burst.In addition,data centers are usually built with shallow-buffered switches that only support a small number of queues,and the transmission paths between source and destination switches are prone to exhibit asymmetry.The existing core technologies of traffic management ignore the internal characteristics of data center networks,and expose the problems such as long tailing,high reordering,low efficiency,and unfairness.The emerging programmable technologies have improved the programmability of the data plane,increased the flexibility of packet processing,and provided new research ideas for datacenter traffic management.Based on the new programmable networking technologies,this paper studies the data center network’s load balancing and packet scheduling mechanisms.(1)Aiming at the resource conflict problem when high concurrent heterogeneous flows coexist in the data center network,this paper proposes a load balancing mechanism named Adaptive Packet Spraying(APS)with isolating heterogeneous flows.This mechanism respectively isolates long and short flows on different paths for transmission,and designs fine-grained routing algorithms for long and short flows according to the traffic arrival pattern,to reduce the completion time of short flows on the premise of ensuring the throughput of long flows.In this paper,APS is deployed on the programmable network interface card supporting DPDK for testbed experiments and large-scale simulation tests.The experimental results show that,compared with the classical algorithms,APS can effectively reduce the average completion time of short flows and significantly improve the throughput of long flows.(2)Aiming at the problem that the out-of-order packets are prone to occur under the asymmetric topology and reduce the network utilization in data center network,this paper proposes a load balancing mechanism with Adaptive switching Granularity(AG),which adaptively adjusts the switching granularity according to the asymmetric degree of topology.This mechanism increases switching granularity to alleviate packet reordering under large degrees of topology asymmetry while reducing switching granularity to obtain high link utilization under small degrees of topology asymmetry.The degree of topology asymmetry is measured by using duplicate ACKs to measure the difference of one-way delays of multiple paths between the source and destination switches with low probing overhead.In this paper,AG is deployed on the programmable network interface card supporting DPDK for testbed experiments and large-scale simulation tests.The experimental results show that AG can effectively improve network utilization while reducing the packet reordering ratio,and greatly reducing the average and tail flow completion time.(3)Aiming at the fair bandwidth allocation problem of high concurrent flows on switches with shallow buffer and few queues in the data center,this paper proposes a scheduling mechanism based on Elastic Fair Queueing(EFQ).This mechanism evenly allocates the buffer space of a limited number of priority queues according to the number of active flows,which can achieve fair bandwidth allocation among flows.Secondly,the mechanism uses the traffic attribute to dynamically allocate buffer space in each queue for each flow to reduce unnecessary packet loss,which guarantees flow-level fairness while obtaining high link utilization.Finally,the effectiveness of the proposed mechanism is verified by deploying EFQ on a programmable P4 switch.The results of the testbed experiments and largescale simulation tests show that EFQ can effectively ensure the fairness of the flow,while greatly reducing the average flow completion time.(4)Aiming at the problem that a limited number of queues is hard to guarantee the performance of packet scheduling,this paper proposes a multi-level scheduling mechanism based on Micro-Reordering Queueing(MRQ).This mechanism divides the queues into a micro-reordering queue and an array of priority queues.The arriving packets are firstly buffered in related priority queues for coarse-grained sorting and then obtain further fine-grained sorting with the micro-sorting operation.After two-level sorting,the data packets are sent to the network in an orderly manner.Finally,MRQ is deployed on a programmable P4 switch for testing.The test results show that MRQ can obtain approximately optimal effects according to the requirements of various scheduling algorithms.