CAREER: Software Thread Integration for Low-End through High-End Embedded Processors

North Carolina State University, Raleigh NC

Investigators

Abstract

This CAREER project investigates embedded systems architectures through the application of compiler and architectural techniques which use an existing processor more efficiently. Rather than automatically adopt existing personal computer paradigms for embedded systems, the main theme of this research is to rethink architectures and compilers in the context of embedded systems and how they are designed. Software thread integration (STI) is a novel compiler technique for automatically interleaving multiple procedures (which implement multiple threads of control) into one non-speculative implicitly multithreaded procedure. The threads are integrated at the assembly language level; the compiler applies transformations to resolve conflicts in control-flow and resource sharing. This research project uses STI in two ways. The first builds on the extraction of more fine-grain concurrency from a generic processor. STI enables switching from one thread to another with negligible (if any) performance penalty or delay. This means that some of the hardware resources added to compensate for the processor's poor fine-grain concurrency are eliminated and the functions move to software. The research is developing efficient methods to perform this hardware to software migration efficiently and precisely enough to implement network protocol interfaces completely in software, reducing size, weight, cost and development time for low-end embedded systems. The second uses integration's augmentation of a thread's instruction-level parallelism, leading to faster execution on processors which are able to execute multiple instructions simultaneously. Modern digital signal processors used in high-end embedded systems are able to do so, yet they are poorly utilized because single threads do not contain enough independent instructions. STI helps by increasing the number of independent instructions in a thread to keep more functional units busy and finish the computation faster. This enables the more efficient use of lower-performance processors, reducing system costs and development time. When coupled with power management techniques, it also reduces power consumption and increases battery life. The PI is creating a new embedded systems track within the computer engineering curriculum at NC State University. Two new hands-on courses on processor-based embedded systems are being added to supplement the PI's existing embedded system design course. The first is a junior-level course that introduces students to microcontrollers, interfacing them to the physical environment, and using a real-time operating system with modern development tools. The second is a graduate course that covers high-performance embedded systems for media processing. Students learn to program and optimize media applications (e.g audio or video compression and decompression) on various architectures (superscalar, VLIW, chip multi-processor). The PI is also involved in outreach activities at the NC School of Science and Mathematics to promote interest in embedded systems design and programming.