MPI并行程序中通信等待问题的诊断方法及其应用

随着并行规模的扩大,现有通信等待问题的诊断方法存在内存开销大、测量时间开销大等问题。通过对现有通信等待问题诊断方法的深入分析,同时考虑测量开销可控的实际需求,建立基于热点函数的通信等待问题诊断模型。基于上述模型,总结出一种更精简、更实用的通信等待问题诊断方法。将该诊断方法分别应用到二维LARED集成、LARED-S、LAP3D等大规模MPI并行程序的通信等待问题诊断过程,应用效果表明本诊断方法可精确定位导致通信等待问题的关键代码段,给出的优化方案及性能提升空间对于后续的程序改进具有参考价值,其中根据诊断结果优化后的LARED-S程序性能提升32%,通信等待时间减少44%。

Diagnostic methods for communication waiting in MPI parallel programs and applications

Communication overhead is an important factor restricting the scalability of MPI parallel programs. The "communication waiting time" is an important part of the communication overhead for large-scale parallel applications. The existing diagnostic methods can identify the root cause of communication waiting phenomenon. However, there still exist some problems such as large measurement cost and memory overhead when running on large scale parallel systems. With the deep analysis on the existing diagnostic methods, and considering the actual demand of controllable measurement, this paper establishes a diagnosis model based on hotspots functions, and presents a tidy and practical diagnostic method based on the above model. The above diagnostic method has been applied to the LARED integration program, the LARED-S program and so on. The application results show that this method can accurately identify the key code segment leading to communication waiting and the proposed optimization solution and performance improvement space have reference value for subsequent program improvement, and the optimized LARED-S program according to the diagnostic result gets an 18% performance increase.