Extending OpenMP for the Optimization of Parallel Component Applications

Yunfeng Peng,Hai Liu

doi:10.1109/access.2020.2996669

Yunfeng Peng, Hai Liu

Open Access

https://doi.org/10.1109/access.2020.2996669

Copy DOI

Abstract

Motivated by the promotion of development efficiency and performance, we research on the common performance optimization technology for parallel component applications. We define OpenMP pragmas to describe the parallel component instance calls in three types of complex structure code. By data flow analysis, we can find the parallelism potential of these codes. We extends the OpenMP pragma to support virtual computing resources and the deployment of complex structure codes on them. We provide three reference rules for the component instance tasks to be scheduled on certain resources to get better performance. To better utilize resources, we provide mapping of virtual computing resources to real resources. We extends the OpenMP programming model and execution model to accommodate the extended pragmas. The extended Babel compiler is responsible for data flow analysis. The extended CCAFFEINE framework is in charge of generating the task schedule policy. Experiments show that our performance optimization platform Bomp can get better scalability and performance than other methods existed.

Highlights

The traditional method of parallel software development is code written by scientists in specific fields, which greatly affects the development efficiency of parallel software
After using the common strategy of performance optimization of parallel component program, most of the execution time of parallel component program is spent on some complex structure codes which can not be optimized by common means of
For the parallel component instance call in the complex structure codes proposed in this paper, we provides extended OpenMP pragmas to support the scheduling and specific scheduling strategy of the parallel component to virtual

Summary

Introduction

The traditional method of parallel software development is code written by scientists in specific fields, which greatly affects the development efficiency of parallel software. There are corresponding extended OpenMP pragmas for pointer tracing and parallel component instance calls accessing irregular data structures. A. DEPLOYMENT ON VIRTUAL RESOURCES To better describe the platform on which the component runs and facilitate the scheduled execution of the component, this paper extends the OpenMP pragma to support virtual computing resources.

Results

Conclusion