Leveraging Multicore Cluster Nodes by Adding OpenMP to Flow Solvers Parallelized with MPI

Towards an accurate simulation of the crystallisation process in injection moulded plastic components by hybrid parallelisation

“…As a pure (dynamic) MPI parallelisation can suffer from communication overhead and high memory consumption (Rabenseifner and Wellein, 2003; Rabenseifner et al, 2009), our MPI-only version is enhanced with Pthreads to exploit intra-node shared-memory parallelism for a hybrid parallel program. Hybrid parallelisation has gained benefit in several applications, for example Rabenseifner et al (2009), Howison et al (2010) and Iwainsky et al (2010). As a first approach, we added the multi-threading layer on the same parallelisation level as for MPI and let each worker process distribute its work across corresponding threads with a static schedule.…”

“…To utilize these architectures, message passing is employed for inter- and intra-node communication in most scientific applications. Aiming at further scalability, the exploration of hybrid parallelism by adding a multi-threading layer can be advantageous (Balaji et al, 2009; Iwainsky et al, 2010). However, the required high scalability is only achievable with a sufficient load balancing that necessitates a careful analysis of the algorithm and the different workloads.…”