SPX: Collaborative Research: Intelligent Communication Fabrics to Facilitate Extreme Scale Computing

点击翻译按钮获取中文摘要

Advances in High-Performance Computing over the last several decades have enabled various important applications throughout science and engineering. Semiconductor technology increasingly faces fundamental physical limits. New approaches to hardware/software co-design are now achieving higher performance at extreme scales. This project will explore two new approaches: configurability and integration. Configurability enables hardware to map better to applications. Integration enables system components that have generally been single function to gain additional functionality. An example is a network for the transport of data. Such a network might also operate on that data as it is being transported. Integration enables compute everywhere in the architecture and network. Configurability and integration will lead to much more efficient use of computing resources. Such resources include high-performance computers. Another focus of this project will be in the use of Field Programmable Gate Arrays (FPGAs). FPGAs are a widely used type of integrated circuit. In FPGAs the hardware can be configured to match the application. FPGAs will allow the deployment of the research products into production systems. They will also allow evaluation of changes in system design. This work will advance extreme-scale computing applications. Such applications include medicine and new drug discovery. This project will also include training of graduate and undergraduate students. It will focus on students belonging to groups underrepresented in STEM disciplinesThe central theme of this project is that emerging centrality of hardware configurability and integration means that at least two old systems abstractions must be broken. First, this project will demonstrate that computation and communication should no longer comprise separate silos. Second, it will demonstrate that applications should no longer be mapped to fixed hardware. In their place, new hardware-software abstractions must necessarily be built around the idea of the application-centric system. This project will identify new computing modes -- computation distribution/offload and configurable hardware -- that will be application aware with tight mapping of applications to computer systems. This improved mapping, in turn, will require deploying a number of mechanisms, starting with improved compilers and high-quality application libraries but extending to lighter-weight yet intelligent middleware (for instance, improved Message Passing Interface middleware), automated application modification, dynamic autotuning, and machine learning assisting in all of these components. The project will consider all of these concerns in due course.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.