Design Of Hetergeneous Computing Model Based On BSP Model

In recent years,the task of large-scale computing and high computational complexity have emerged in various fields,the distributed computing environment has become a research hotspot.Among them,the heterogeneous computing system,which is more close to reality and has more efficient computational performance,gets extensive attention of the researcher.How to improve the existing parallel computing models for making it conform to the requirement of time performance of heterogeneous computing system also becomes a research direction.Basing on the previous research,the paper propose a heterogeneous parallel computing model,named LS-BSP,by analysing the characteristics of the BSP model.The content of the research mainly includes three sections:(1)Make the structure of Superstep better,and propose the thinking of "Logic Superstep" : those tasks that exists data-dependency are put in one Superstep.On the contrary,independent tasks can be put in different Supersteps.Because there is not datadependency between different Supersteps,these steps can run in parallel.Therefore,the parallelism of the model is greatly enhanced.(2)Change the mechanism,which calculate firstly and then communicate,of Superstep,and propose the mechanism of “Compute-Combainer-Send”: set a single synchronized thread for each Superstep,named Combainer,which responsible for data aggregation.On the one hand,computation and communication between different processing nodes can concurrently execute.On the other hand,the result have been aggregated before sending to the server,so that it can reduce the amount of data transferred,and improve the efficiency of transmission.(3)Improve the task allocation mechanism of the model.Usually,the BSP model is applied to the homogeneous computing environments,so its assignment process mainly consider whether the processor is idle or the task data is locally reachable.The DDF-MM algorithm is introduced in the LS-BSP model,which based on the Max-min algorithm.The DDF-MM algorithm use the standard deviation coefficient of processing time and the average processing time as the priority of the task,and simultaneously consider the load condition of processors during the task assignment process.In addition,pi-calculus,a process calculus language,can change network topology and state by transmitting the name of channels.Hence,pi-calculus can briefly describe a heterogeneous computing systems.Therefore,pi-calculus is used to describe the LSBSP model in this article and establish a theoretical model of LS-BSP.Subsequently,relying on the weak bisimulation of pi-calculus,we prove the theoretical correctness.Pict language is a natural parallel language,which is based on the pi-calculus.It can programme in a natural semantics.In this article,we respectively use Pict language and Java languages to describe the LS-BSP model,and describe the BSP model in Java,then compare results to verify the performance improvement in the processor load,the time cost reduction of LS-BSP model,and the optimization of parallel computing model with Pict language.