Overview
This paper describes the domain decomposition approach based on the Message Passing Interface (MPI) recently implemented in Environmental Fluid Dynamics Code Plus (EFDC+). With this feature, EFDC+ can run across large cluster systems as well as multi-core desktop computers with greater speed than before — though with increased complexity in specifying the computer configuration for optimal performance.
Approach
The paper describes EFDC+‘s domain decomposition approach and the computer settings needed to achieve faster run times, including guidelines users should follow when preparing and executing an EFDC+ run with MPI. The run time performance was analyzed using a large model of the Chesapeake Bay.
Key Findings
Prior to this version, only multi-threading using Open Multi-Processing (OMP) was available to improve performance, and the previous version of EFDC+ could only achieve a speedup of 2.5 for the Chesapeake Bay model. The current hybrid OMP/MPI version with domain decomposition achieved a speedup of nearly 17 with 32 compute cores. A strong scaling study conducted on an Amazon Web Services cluster reached a speedup of nearly 25 while using 96 compute cores — demonstrating substantially improved performance and the ability to run across a cluster.