Argobots: Harnessing Massive Concurrency with Tasks & Fine-grained Control of Scheduling

Parallel computers have become very powerful yet increasingly complex in recent years. Individual nodes comprise heterogeneous components such as CPUs and GPGPU's, and a large number of concurrent actions need to be scheduled efficiently to attain good performance. Task-based systems, and the Argobots tool developed at Argonne National laboratory are critical for this purpose. For broad use by American industry, Argobots needs to be enhanced and commercially supported. We aim to develop "Argobots Pro", as a commercially viable product derived from Argobots. The project will enhance the functionality of Argobots, develop new components within it, and build an ecosystem of libraries and tools around it. Many of the features we plan to develop arise from requests from the Argobots user community. We expect Argobots Pro will be used by developers of task-based runtime systems, parallel programming frameworks, higher-level languages, as well as parallel application developers. It will have a significant positive impact on performance and productivity of applications running on dedicated clusters or in the cloud. The task undertaken during Phase I effort will include support for managing GPGPUs, and improved support for loop parallelism and tasking. We will also develop a prototype of a suite of debugging and profiling tools for Argobots. Our team participated in the original design and implementation of Argobots and have experience in developing task-based systems, including Charm++. We will draw on our experience to shape the capabilities of Argobots Pro, with guidance and feedback from the Argobots team at the Argonne National Laboratory, as well as the Argobots user community. There are no commercially supported tools that can compete with Argobots Pro, although Argobots itself is freely available. Once Argobots offers Profiling/debugging tools, and other proposed improvements, and once it is released as a commercially supported software package, with promise of customer support, assistance in customer code development, and consulting, we expect it to become a commercial success, and help improve competitiveness of American businesses.