Bruce Jenkins, Ora Research
This test project for high-performance computing (HPC) in the cloud was designed to explore how cloud HPC resources can help to speed up and enable high-performance finite element simulations carried out with COMSOL Multiphysics and COMSOL Server. The objective was to find out how HPC cloud providers can augment engineering organizations’ on-premise hardware to allow for more detailed and faster simulations.
COMSOL Multiphysics is a general-purpose software platform based on advanced numerical methods for modeling and simulating physics-based problems. Its multiphysics capabilities make it especially well suited for coupled and multiphysics simulations, facilitating the modeling and analysis of real-world multiphysics phenomena and systems.
COMSOL Server was created specifically for running and distributing applications built with COMSOL’s Application Builder. To help users spread the advantages of simulation throughout their organization and networks, COMSOL Server is a platform for deploying applications created by an organization’s simulation experts to its design teams, manufacturing departments, test laboratories, customers and clients anywhere in the world. Colleagues can access and run an expert’s apps on COMSOL Server through web browsers or a desktop-installed client. The COMSOL Server web interface lets experts and their colleagues manage access to the variety of apps they have created, and manage hardware settings and preferences for running the apps.
For this project, two COMSOL applications were used: a 2D model and a 3D model of an industrial fragrance extraction boiler. Both models used specific modeling for mass and heat transfer of fluid flows and immobile solids built from the COMSOL Multiphysics Heat Transfer Module.
These two models were chosen because they represent two frequently used options for thermal and fluid flow models. The 2D model enabled exploration of a wide range of operating conditions, while the 3D model was suitable for detailed analysis and optimization. Only one form of parallelization of multiphysics simulations was tested: shared-memory parallelism (SMP).
Challenges
Each model brought its own challenges. In the 2D model, phase change was added to allow simulation of water vaporization. This slows down the computation so much that such simulations take days to execute on powerful engineering computing laptops or desktop PCs.
For the 3D model with its large geometry, a large amount of memory is needed to be able to compute the model at all. For many users in small or medium enterprises who lack the required computational capacity on premise, the cloud provides an ideal hardware extension for handling these kinds of models sporadically.
Work process and benchmark results
Computations were performed on cloud resources from CPU 24/7, a leading provider of Simulation-as-a-Service solutions for all application areas of industrial and academic research and development. Headquartered in Potsdam, Germany, CPU 24/7 develops and operates on-demand services for HPC based on latest industry standards for hardware, software and applications. All computations were performed on a single node equipped with dual-socket Intel Xeon E5-2690 v3 and 256 GB of RAM, giving a total count of 24 cores. This hardware setup was chosen because it is equivalent to the latest Intel desktop workstations suitable for engineering computations. The hardware (bare-metal, non-shared infrastructure) was supplied by CPU 24/7, with COMSOL Multiphysics pre-installed on UberCloud’s application software container.
As the COMSOL Floating Network License (FNL) and COMSOL Server License (CSL) allow for remote and cloud computing out of the box, the project team only needed to use the built-in COMSOL Multiphysics functionality to launch jobs on remote clouds and clusters, and to forward access to the on-premise license manager. Once this functionality has been configured, any model can be sent to the cloud cluster in a matter of seconds.
The simulations were performed both in the COMSOL Multiphysics environment and on the COMSOL Server.
The remote COMSOL Multiphysics cloud environment proved to be very responsive, and the graphics quality on a large screen was excellent, after setting the proper parameter in the network connection software. This enabled handling of the large 3D model without any clumsy delays between mouse inputs and GUI response. Furthermore, operations such as meshing and post-processing of the large model took only a few seconds—tasks that push the limits of computational modeling in precision and accuracy by an order of magnitude compared with regularly available hardware.
On the other hand, the Application Builder is not available from within the Linux container, so the COMSOL Apps had to be built on the local desktop. That was no problem, as building and testing an app in the Application Builder requires less memory and computing power than meshing, solving or post-processing a model. Further, the application can be easily uploaded to the COMSOL Server and tested there in the final stage of the build/test cycle.
Benefits
For end users, the use of COMSOL Multiphysics from the UberCloud container together with the HPC cloud services of CPU 24/7 proved to bring many advantages:
- Increased number of cores and memory channels yield higher throughput for time-critical projects.
- The increased amount of memory and faster memory access available from an increased number of cores allows simulation of more realistic and detailed models, which could be impractical or impossible to execute on local
- Users can gain fast access to a powerful “workstation in the cloud” thanks to the fast and capable support from CPU 24/7. With COMSOL Multiphysics pre-installed in an UberCloud software container, and access to the CPU 24/7 cloud resource enabled via both remote desktop solutions and SSH, there proves to be no reduction in usability compared with an on-premise
- The easy access to the cloud provided by the COMSOL Multiphysics GUI, and the possibility for an organization to forward its license, makes it easy to extend its on-premise hardware with cloud HPC resources when needed.
Conclusions
- The CPU 24/7 HPC bare-metal cloud solution is a beneficial solution for COMSOL users who want to obtain higher throughput and more realistic results in their simulations.
- Use of cloud computing removes the need for investment in high-end in-house workstations or servers that would only be used occasionally. Furthermore, future simulation projects will benefit from continually updated hardware and software, thanks to CPU 24/7 offering permanently available and tailored HPC bare-bone cloud
- Once validated, the containerized environments provided by UberCloud were set up entirely by the UberCloud team, requiring no end-user involvement except for a handful of remote computing settings to adapt screen resolution to the COMSOL GUI
This article is based on a case study authored by Stephan Savarese of ForCES, a provider of multiphysics simulation, training and consulting services for climate, energy and science using COMSOL Multiphysics and COMSOL Server. The COMSOL software container from UberCloud was running on CPU 24/7 cloud resources.
More on The UberCloud Experiment here:
http://www.theubercloud.com/
For end users wishing to participate in the free and voluntary UberCloud Experiment:
http://www.theubercloud.com/hpc-experiment/
For service providers interested in promoting their services on the UberCloud Marketplace:
https://www.theubercloud.com/help/