Prefetching And Data Percolation Technique For Applications On Many-core Architecture

In recent years,the applications on many-core have been a great success in many fields of science and engineering.Compared to the traditional multi-core architecture,the many-core processor can provide a more powerful computing power and more efficiency memory access bandwidth.However,Because of these new features,the optimization of large-scale data parallel applications based on the new many-core architecture is also facing great challenges.In order to solve the problem of low utilization of computing resources which caused by the impact of the access latency,this paper studies prefetching and data percolation technique for applications on many-core architecture.In this paper,to solve problem of data access latency in many-core applications,we analysis of the status of the current many-core platform data prefetching techniques,based on many-core architecture execution mechanism and storage characteristics,summed up the optimization for many-core principles;using many-core application warp execution rules and shared storage resources,we design the inter-thread prefetching method to hence the thread level parallelism;using data exchange between local and global memory of many-core architecture,increase the rate of application data reuse,and implement in benchmark program.Experiment results show that the prefetching method can effectively hide the latency between threads and improve the computing performance of the applications.To solve computing dense in thread memory inefficiency and instruction level parallel degree low for many-core application,we redefined the data percolation operation;to solve the low computational resources utilization rate under static scheduling,we design the dynamic scheduling algorithm;by making use of many-core processor architecture features,we design optimization model of percolation operation.Finally,we implement the percolation operation in the benchmark program.Experiment results show that data percolation can effectively hide the latency of memory access and improve the instruction level parallelism.