OS/architecture interactions and their influence on computer architecture.

Abstract

Good computer architecture design requires a thorough understanding of the interactions between the software and the hardware. Recognizing this fact, architects spend considerable effort studying application code behavior and basing their design decisions on quantitative analysis that considers how the architecture can best support application software. Most studies, however, neglect to consider one of the most important software components--the operating system. This has resulted in a dearth of quantitative data detailing OS/architecture interactions, significantly limiting an architect's ability to make informed design decisions. This dissertation examines the influence operating system software has on architectural design. Using a range of UNIX-based operating systems, each with radically different internal organizations, and a basic RISC workstation architecture as the experimental base, we extend quantitative analysis techniques to the study of OS/architecture interactions. The first phase of our work is measurement and analysis. Using a suite of benchmarks chosen to exercise the operating system and architecture, we compare the performance of each operating system. Our measurements show that operating systems such as OSF/1 or Mach 3.0 suffer significant performance penalties, increasing application run time up to 100% over a more mature operating system, Ultrix. Our analysis reveals that 40% of the performance loss is due directly to unforeseen interactions between the operating system and the architecture. In particular, the decomposition and migration of operating system services radically alters behavior, increasing the amount of stress seen by the architecture's translation lookaside buffer (TLB) and instruction cache. The second phase of our work is the exploration of architecture solutions. Using numerous simulation techniques, we explore TLB and instruction cache design trade-offs that allow the architecture to support advances in operating system technologies. Our TLB studies show that careful design of the architecture's TLB allows the architecture to support service decomposition. Instruction cache results show that caches alone cannot overcome the operating system's increased working set. Instead, architectures should provide intelligent prefetching mechanisms that allow the hardware to support the software's increased working set size.