DAG Parallel Of AMRPossion Case In Chombo

Over the past ten years, the computer systems no longer depend on increasing the frequency of CPU to improve the performance. Multi-core or many-core systems composed by CPU with multiple simple cores have been main technology in last decade. How to make full use of multi-core, many-cores systems resources to achieve high performance is critical.Structured grid application is a typical application on high performance computing platform. With the widely use of multi-core and many-core platform,how to transform existing legacy codes onto these new parallel platforms is a great challenge. When we parallel structured grid applications on multi-core and many-core platform, one big data set is divided into many data patches. Communication among adjacent nodes caused by the data partitioning set will affects the whole parallelization system performance significantly. Analyzing communications and implementing the overlap between computing and communication will improve the performance of the whole system.In this thesis, we use sample program AMRPossion of Chombo as target application to study the task graph parallel method on HPC platforms.First of all, analyze the characteristics of neighbor patches communication in structured grid applications.Then, we analyze data dependence between neighbor patches in various computing iteration, through the analysis of dependencies among tasks, distinguish between each patch blocks to determine the relations dependency either it is local or remote data. Local data communicate by shared memory. But remote data must be transformed by network communication among computing nodes. The remote communication is the bottleneck of the system. So, improve data communication tasks by using overlapping techniques to hide the overhead caused by remote data communication, thereby improving the performance of task graph parallelism.Finally, we implement multi-processes DAG parallel of AMRPoisson and test performance on a basic multi-core platform. Testing data shows: compared with MPI+OpenMP parallel model, parallel task graph (MPI+DAG) program can effectively improve the performance of the program.