The Research And Implementation Of Data Transport Protocol For The Reconfigurable Cloud

With the development of emerging services,such as big data and machine learning,the designers of data centers begin to use FPGA-based reconfigurable cloud architecture to accelerate application and network.In order to improve the resource utilization of FPGAs,cloud architecture developers prone to deploy direct end-to-end transport protocols between distributed FPGAs to integrate them into a resource pool.In addition,the FPGA offloads the transport protocol,which effectively reduces the CPU load,to increase the number of tenants for economic revenue.At present,the transmission protocols that can be applied between FPGAs mainly include NDP(New Data-center Protocol architecture),LTL(Light Transport Layer)and traditional TCP protocols.However,offloading the NDP protocol requires customization of all the switches in the network,which makes it hard to deploy the NDP on a large scale data center;Offloading the LTL protocol needs to open complex and risky mechanisms in switches.Once the parameters of the mechanisms are configure wrongly,the performance may seriously drop.To support the mechanisms in switches,the high-level technical teams are needed to maintain it,which is exacting for most companies.Besides,a large number commercial switches do not support the mechanism;Offloading complex TCP protocols will inevitably consume a large amount of hardware resources,and the performance can not meet the requirements of the data centers due to its drop-based congestion control mechanism and slow start.Therefore,in view of the shortcomings of the above protocols,we studie the high performance transmission protocols for reconfigurable clouds.The main work and innovations include:1.Proposing a new reliable end-to-end transport protocol for reconfigurable clouds,STRIDE.STRIDE is a connection-oriented and lightweight transport protocol that provides full-duplex and reliable transport without any modifications to commercial switches.2.Describing the overall structure and detailed design of the implementation model of the STRIDE protocol.In order to improve transport performance,STRIDE adopts a effective congestion control mechanism,which is based single-trip-time,and a splitbandwidth startup mechanism.To simplify the design,STRIDE adopts the Go-back-N mechanism and the transport rate adjustment mechanism,which contributes significantly to reducing hardware resource consumption.3.Deploying STRIDE protocol on real devices and on a large scale simulation platform.The test results of real prototype show that,compared to the traditional TCP,STRIDE can reduce the hardware resource consumption on the FPGA by 36% to 49%.Also,the average flow completion time of the STRIDE increases to 2.2 times.What's more,the simulation results show that STRIDE has a corresponding average transmission performance of 1.8 times and 2.3 times compared to DCTCP and TCP-Vegas.To sum up,in order to solve the shortcomings of existing transmission protocols that are able to be deployed in reconfigurable clouds,we propose a reliable end-to-end transmission protocol,STRIDE.Because the transmission protocol possesses high performance,low resource consumption,and no need for switch,it can effectively solve the problem of insufficient performance and excessive consumption of programmable device resources in the data center,while avoiding the difficulties and risks of opening the complicated mechanism on switches,which has essential theoretical and practical value for the further promotion of reconfigurable cloud architecture.