High Performance Computing
BREEZE developers are leaders in the development of high-performance computing models for air and explosion modeling. Our parallel and remote modeling versions of AERMOD are the first of their kinds and continue to be the most widely used high-performance versions of these models. In addition to AERMOD, BREEZE had parallelized various internal processing routines incorporated in many BREEZE products (e.g., gridding methods and doublet theory).
Since its first use in 1972 by the U.S. NASA Ames Research Center, parallel processing has been commonly used in high-performance technical computing for engineering applications and scientific research. AERMOD is a scientific model, however it is not considered high-performance. With its voluminous sets of meteorological data and intensive calculations, the progression to parallel processing for AERMOD was natural. Better computing technology and increased demand for smaller, more powerful computers has enabled the evolution. Today, dual-core processors (two CPUs) are standard features on new computers with the availability of a greater number of CPU’s available on a single chip.
While parallel processing dominates computer architecture, none of the U.S. EPA regulatory dispersion models are designed to take advantage of the ubiquitous multi-processor. There are numerous reasons why this may be, including the fact that designing and developing parallel software is a complex and time consuming process if done correctly. The extra efforts that would be required by regulatory software developers are often devoted to improvements to the underlying model science.
To comply with The Guideline on Air Quality Models, BREEZE developers were challenged with leaving the AERMOD mathematical algorithms unchanged while at the same time incorporating sophisticated CPU management routines to handle the simultaneous calculations among multiple processors. BREEZE parallel programs change only the way the computer executes the algorithms by dividing several calculations into smaller problems that are completed simultaneously (i.e., in parallel). Identical results are guaranteed in true parallel applications and can be verified through equivalency demonstrations. Other “parallel” methods include manually separating algorithms to run on multiple individual computers and piecing the results back together.
High Performance Computing Options
There are two deployment options available to users interested in parallel processing— local and remote resources.
A local approach involves purchasing parallel software and determining how many processors are desired to run the software application. If a new octo-core computer is not in the budget, a computer cluster can simulate a single multi-core computer. Clusters are comprised of multiple stand-alone computers connected by a network. Creating a cluster requires in depth technical knowledge of computer networking and requires that each individual computer user designate available processors for the duration of the dispersion model run.
Acquiring local resources can be a costly investment for organizations needing to purchase, maintain, and update computer equipment. Depending on its configuration, it may also require the software application be run during non-office hours when people are not using their computer for daily work, partially offsetting the speed capacity of parallel software. Purchasing or using existing hardware is a viable option for users in organizations that have the ability to maximize local resources.
Alternatively, remote resources are gaining popularity because no proprietary software is required and the parallel application runs on a massively parallel computer cluster owned by someone other than the user. A massively parallel cluster is a single computer with many networked processors; similar to the above mentioned cluster, except it is a much larger cluster with more than 100 processors. Remember, improvements in the parallel application’s performance are based directly on the number of CPUs (or multiprocessors) utilized. The average modeler could not locally simulate the horsepower a massively parallel system can provide.
Non-local resources or remote parallel processing systems are accessed via the Internet. Users upload input files anytime (24/7), receive email notifications when model results are completed, and slash model runtimes. Results are identical to the U.S. EPA’s serial version of AERMOD. Air dispersion modeling is radically streamlined with remote parallel processing.