There is a growing demand for a high-performance, high-throughput, on-demand computing infrastructure supporting complex and large-scale modeling, big data and machine learning applications. At Florida International University (FIU), we engage in multi-disciplinary research in environmental science and engineering (such as coastal environmental monitoring, freshwater quality analyses and ecotoxicological studies, everglades restoration), extreme events (such as disaster risk and resiliency analysis, hurricane loss and storm surge modeling), and computational and systems biology (including proteomics, genomics and connectomics). These important science applications are highly complex and dynamic systems (natural or human-built). They all rely on computation to support simulation, modeling, and analyses to enable discovery, facilitate understanding, and make predictions and decisions. However, the computational demand of these important science domains may vary significantly, and can be characterized into one or several of the following categories:
- High-performance computing: Climate models, hurricane storm surge models, surface water and groundwater models, flooding simulations, and agent-based models are all examples of parallel applications. These applications are computationally intensive, and are suitable to run on parallel platforms, ranging from a single multi-processor, multi-core compute node to a high-performance computing cluster with tens or hundreds of processing cores, in order to reduce the computing time and/or achieve the desired problem size or model granularity.
- High-throughput computing: Some applications can generate a large number of compute jobs with varying input data. Each job may only require limited parallelism. An important characteristic of these applications is that their computing efficiency is measured by the time it takes to run all the jobs of the applications rather than an individual job. As such, a job can be preempted and run opportunistically on different nodes, on different platforms, even at different locations.
- Data-intensive computing: Many applications require processing a large amount of data. For example, the Florida rainfall projection application requires continuous access to multiple terabytes of data. A mass-spectrometry data for proteomics study may involve tens of millions of experimental spectra data to be searched against an indexed database of multiple terabytes in size. Such applications need to be run on systems equipped with large memory and data storage.
- Machine learning: A specific type of data-intensive computing is machine learning applications. For example, the deep neural networks (DNN) used in coastal flooding modeling and proteomics studies may include large training datasets (amounting to hundreds of gigabytes for each training batch) and large models (involving billions of parameters). Many machine learning applications are more suitable to run on Graphics Processing Units (GPUs), which are specially designed for handling numerically intensive calculations and neural network model training.
- Real-time on-demand computing: There are applications that come with strict deadlines. For example, the Caribbean coastal inundation and forecast model needs to run simulations using results produced at supercomputers at regular intervals. We anticipate that these applications are becoming more common, especially for coastal resilience and environmental studies, which require real-time data acquisition and data collection. The data needs to be processed and analyzed in real time for decision making or for controlling the experiments. An example of such applications at FIU is steering autonomous surface/underwater vehicles for coastal environmental monitoring.
Many applications are naturally multidisciplinary and require interoperation of multiple computing jobs with different computation demands. For example, the Florida hurricane loss model consists of multiple interacting components, including wind hazard (meteorology), inland flooding (hydrology and hydraulics), coastal flooding and wave (oceanography), vulnerability (engineering), and insured loss cost (actuarial). Some require high-performance computing; others need high-throughput computing or data-intensive computing.