The Foundation: Scalable Architecture

COMSOL Multiphysics^® features a cluster-aware architecture that scales from local workstations to enterprise HPC environments and clouds. Following a one-time cluster configuration, the software automatically optimizes numerical workloads for a specific hardware. This ability enables engineers to transition from desktop prototyping to large-scale production runs without manual changes or repetitive cluster tuning.

Multicore or Multiprocessor Scaling (Shared-Memory Parallelism)

Take advantage of the full power of a workstation's hardware out of the box. COMSOL Multiphysics^® uses shared-memory parallelism (via OpenMP^®) to automatically distribute computational tasks over all available processor cores. This automatic parallelization means immediate performance gains for large high-fidelity models without the need for any software configuration. For shared-memory parallelism on a single computer, no additional license fees are required to run on an unlimited number of processor cores.

High-Performance Scaling (Distributed-Memory Parallelism)

Scale beyond the physical limits of a single machine by leveraging the pooled resources of a cluster. With a floating network license (FNL), no additional license fees are required to run on a cluster of unlimited size. By using MPI to partition a model’s memory footprint and computational load across multiple nodes, COMSOL Multiphysics^® can perform high-fidelity simulations with hundreds of millions of degrees of freedom (DOFs). This capability enables users to run extensive mesh refinements or solve large, complex multiphysics models, as long as sufficient cluster hardware is available.

Maximizing Hardware Efficiency (Hybrid Memory Parallelism)

Maximize computational throughput by combining shared and distributed memory strategies. By balancing multithreading (via OpenMP^®) within individual nodes and message passing (MPI) across a cluster, COMSOL Multiphysics^® minimizes network communication overhead and optimizes hardware utilization. This hybrid approach ensures that simulations scale efficiently on modern many-core architectures, providing faster turnaround times and predictable performance across massive compute resources.

High-Throughput Computing: Managing Parallel Parametric Sweeps

High-throughput computing enables rapid design exploration by shifting the focus from solving a single massive model to evaluating thousands of parameter variations concurrently. On HPC infrastructure, COMSOL Multiphysics^® offers several distinct task-parallelism capabilities. The Batch Sweep node leverages native scheduler integration, such as SLURM^® or LSF^®, to spawn separate batch jobs for different parameter values, ensuring that a single nonconverging variation does not halt the entire study. Alternatively, a Distributed Parametric Sweep bundles multiple concurrent parameter evaluations into a single large MPI job, which can be highly efficient for bypassing long scheduler queue times when deploying across a distributed system.

This high-throughput capability is not limited to cluster environments. For users operating on local multicore workstations, the Batch Sweep feature brings task parallelism directly to the desktop. As a result, it is possible to solve multiple parameter sets concurrently by maximizing local hardware utilization, providing a method for smaller-scale design mapping before transitioning to a cluster.

A third option is to run completely independent COMSOL Multiphysics processes on a cluster, with each process handling a separate model, parameter set, or simulation scenario. This approach is well suited for loosely coupled workloads where each run can proceed independently, and it gives users full control over job submission, monitoring, and results evaluation through scripts or external tools. With an FNL, this workflow can scale to large numbers of design variations of the same model without additional license fees based on cluster size or core count.

These parallel sweep technologies serve as the engine for machine learning (ML) workflows, specifically synthetic data generation in the Surrogate Model Training study. By rapidly mapping design spaces across local or cluster hardware, engineers can compile massive, ground-truth datasets to train surrogate models such as deep neural networks (DNNs).

Cloud Computing Compatibility

Many users run simulations in the cloud to access scalable, on-demand computing resources without maintaining their own HPC infrastructure. Users with an FNL can use the license on the cloud platform of their choice.

COMSOL models can be deployed on major cloud providers such as AWS^®, Azure^®, and Google Cloud, as well as through HPC cloud platforms and partners such as Rescale and Nimbix. Users also have the option to run the software with Amazon Elastic Compute Cloud™ (Amazon EC2™). More information is available on the HPC partners pages.