XPS: EXPL: Hippogriff: Efficient Heterogeneous Servers for Data Centers and Cloud Services

University Of California-San Diego, La Jolla CA

Investigators

Abstract

The growing importance of artificial intelligence, network services, and cloud storage drives the demand of building powerful computer systems that can perform many operation at once. Building computers with different kinds of computing processors (i.e., heterogeneous processing) is an effective way to achieve this goal. However, this approach also creates new problems that can negate some of the benefits it provides. In particular, using different processors for different tasks requires moving data between those processors. This movement takes time and can cancel out saving heterogeneous processing provides. This project is addressing this problem in heterogeneous computing systems by making the movement of data between different processors more efficient. This improved efficiency leads directly to benefits for applications of scientific and commercial importance. Much of the cost of data movement heterogeneous computing systems stems from the entrenched central processing unit (CPU)-centric programming model. This project is revisiting the design of the application interface, system software and hardware components to remove CPUs and main memory from the critical path of moving data. The project provides an efficient programming model that allows the system software stack to automatically and efficiently setup the data movements between heterogeneous processors. We are applying the system to large-scale database systems, massive parallel programming systems like Spark and MapReduce as well as scientific computing that power important daily applications and research projects.