FAQs

The primary benefit of MPI-based algorithms is they can be run on a cluster; this is in contrast to multi-threaded programs based on something like OMP, which on their own are limited to running on a single node/computer. Increased computational performance now primarily comes from having additional cores available in a given computing environment and writing software with parallelization in mind.

OMP can still be efficient for multiple constituent simulation (e.g. WQ).  When running an EFDC model with a small number of constituents (e.g. only hydrodynamics). MPI is much more efficient, however, with a large number of grid cells where the domain can be split up and handled by separate MPI nodes.

The only requirement to run an EFDC+ model with MPI implementation is to have multiple cores available on the computer running the model. The configuration of that computer is of importance to understand run times and the maximum attainable performance. Most desktop computers have CPUs with multiple cores, which will allow MPI applications to run on that single computer. However, if it desirable to run a problem on more cores than are available on a single computer, this naturally requires several computers to be clustered.

Yes, MPI can run on an PC with more than four cores. However, efficiency gains will only be attained if you can divide your model into a greater number of cores.  For some models running in OMP mode, using only one domain would be more efficient.  To see how to run MPI on your PC or laptop see our guidance here.

Yes.  For example, DSI recently configured a cluster on AWS using their Parallel Cluster service (Services 2020a).  This allows a remote (cloud based) cluster to be configured on demand.  A cluster containing four compute nodes and one head node was set up.  Each compute node contained a 36-core Intel Xeon Platinum 8000-series CPU.  To minimize communication latency between nodes, a high-speed interconnect was set up using the Elastic Fabric Adapter (EFA) interface.  A high-speed file system based on FSx for Lustre was used to minimize time spent writing to the disk during a simulation.  For detailed guidance on how to set up your own system refer here.

Yes.  The recommended approach is to first build your model on a Windows machine running EEMS and then transfer the model input files to the Linux system.  The Linux system will have a different EFDC+ executable than Windows but uses the same input file names and file extensions.  The output generated from Linux can be copied back to the Windows machine and viewed in EFDC_Explorer.

depend to a significance extent on how the PCs are connected.   If the interconnect between the machines is slow then some of the benefit of MPI will be lost.

No, you only need to install EEMS on one machine.  This is the master machine. Other machines, the slave machines, need to have the Intel MPI Library for Windows installed and share the working directory and the model and EEMS installation shared with them.

Recommendations for hardware are provided here.