Many scientific computing applications in critical areas of research, such as nanotechnology, astrophysics, climate, bioinformatics, and high-energy physics, are becoming more data intensive than ever before. The volume of the data and the pressure on the runtime system capability of supporting data intensive operations substantially increases over the time. This project introduces and optimizes data processing capabilities within memory. The project provides a fundamental change to data management and program optimization, and brings promising performance and energy benefits. The project will significantly advance simulation capabilities of scientific applications, especially those with intensive data processing.
The goal of the project is to enable high-performance, energy-efficient, and flexible processing-in-memory design, which is adaptive to the irregular, diverse, and changing behaviors among data intensive scientific applications. To achieve the goal, a heterogeneous processing-in-memory design, built up with fixed-function processing-in-memory and general programmable processing-in-memory, is introduced. The project explores a series of critical questions for building emerging processing-in-memory, including heterogeneous processing-in-memory architecture, processing-in-memory programming models, runtime design, and the implications of processing-in-memory on high performance scientific applications. The project will significantly advance the knowledge to build processing-in-memory, and pave the way to integrate it into the existing and future systems.
