Efficient Core Utilization in a Hybrid Parallel Delaunay Meshing Algorithm on Distributed-Memory Cluster

Most of the current supercomputer architectures consist of clusters of nodes that are used by many clients (users). A user wants his/her job submitted in the job queue to be scheduled promptly. However, the resource sharing and job scheduling policies that are used in the scheduling system to manage the jobs are usually beyond the control of users. Therefore, in order to reduce the waiting time of their jobs, it is becoming more and more crucial for the users to consider how to implement the algorithms that are suitable to the system scheduling policies and are able to effectively and efficiently utilize the available resources of the supercomputers. We proposed a hybrid MPI+Threads parallel mesh generation algorithm on distributed memory clusters with efficient core utilization. The algorithm takes the system scheduling information into account and is able to utilize the nodes that have been partially occupied by the jobs of other users. The experimental results demonstrated that the algorithm is effective and efficient to utilize available cores, which reduces the waiting time of the algorithm in the system job scheduling queue. It is up to 12.74 times faster than the traditional implementation without efficient core utilization when a mesh with 2.58 billion elements is created for 400 cores