Embarrassingly Parallel (EP) problem spaces are typified by discrete subparts which can individually be solved efficiently without cross communication, possibly with some pre-processing and post-processing steps to coordinate, but mainly lots of independent members doing their own thing.
EP Grids tend to be simpler and less expensive, and given the problem domains, we have typically aimed at lower memory and faster local disk targets. However, since these are shared resources, higher memory targets are increasing in popularity, and larger local disks as well. In this space, we have stuck with some descendant of Sun Grid Engine, subsequently using Open Grid Scheduler, and now Son of Grid Engine. SGE nee OGS nee SoGE handles job and resource management as well as scheduling.
Our EP Grids mount SEAS Home Directories, and typically do not contain their own separate storage. In our standard configuration, all compute nodes are directly externally accessible.
SEAS HPC Grids
- Architecture and Compiler Group Grid
- Alef-Bet Computation Grid
- Bayes Shared HPC Grid
- Implementation of Computation Grid
- NLP Grid
- NLP GPU Cluster