Research And Implementation Of GPU Accelerated Active Storage Based On FastDFS

With the development of network technology,people are more and more dependent on the network.The data are stored in the cloud server side,more and more storage services need to perform analysis,classification,summary and other operations on the data in the storage node.This puts forward higher request to the network bandwidth of the network storage system.The main purpose of this paper is to improve the performance of the distributed file system FastDFS by studying the active storage mechanism.First of all,we designed the active storage(AS)model based on active storage concept,which is suitable for FastDFS.It uses the traditional C/S model,we have implemented the active storage interface function in the client and the server,which is used to realize the active storage mechanism.There are two shortcomings of the above active storage model.The first disadvantage is that it is only suitable for data filtering services,and the accelerating effect of other types of storage services is not obvious.This is because CPU not only needs to handle the command sent by the client,but also need to perform the corresponding calculation task.What's more,CPU is designed to deal with the logic computing tasks,and is not good at data intensive computing tasks.In view of the above two shortcomings,we optimized the original active storage model,and proposed the GPU accelerated active storage(GAS)model.In the model,we added a parallel computing module which uses GPU computing power,as well as the kernel library module to solve the system scalability.By making use of the large scale parallel computing capability of GPU,the computing capability of the server can be enhanced,so that it can be suitable for more types of storage services.In the implementation process,we use the parallel computing OpenCL programming language to utilize the GPU's general computing power.Secondly,after the design of the active storage model is completed,we designed and implemented an active storage communication protocol(AS CP)between the client and the server,and realized the active storage APIs of the clients and servers,respectively.Finally,the experimental platform is set up to test the above two models.First,we carried out the active storage AS model performance improvement test,the experimental results show that the AS model can achieve better performance in the data filtering service.Second,we carried out the GPU accelerated active storage GAS model performance improvement test,the experimental results show that the performance of GAS in the processing of filter storage services greatly improved,more obvious than the AS model.When processing computing intensive services,the performance gains are much higher than that of the AS model.