Partitioning The Operating System: A Step Towards Heterogeneous Chip Multi-Processors

by
Dave Nellans

Advised by
Erik Brunvand

In recent years architectural improvements which focus on exploiting instruction-level parallelism (ILP) have yielded marginal speed-ups while dramatically increasing power consumption and design complexity. These mainstream processors are inherently limited by their attempt to be uniformly good for instruction mixes that are always differing from the optimization set. One promising way to achieve increased performance is to provide custom tailored hardware to the applications which are consistently executing; graphics, I/O, and digital signal processors are examples of application spaces where this hardware provides an undeniable advantage. A similar area that is currently under-studied is the operating system (OS). We propose designing a special-purpose OS processor (OSP) along side a traditional high performance general purpose processor (GPP) to form a heterogeneous chip multiprocessor (HCMP). With just a small increase in total transistor budget a HCMP including a operating system processor can significantly increase total system performance while simultaneously reducing power total consumption.