Understanding the Impact of X86/nt Computing on Microarchitecture

Many performance evaluation studies in computer architecture rely almost exclusively on simulation of the dynamic instruction stream from a single application. The benchmarks used are often CPU intensive and rely very little on the operating system, such as the SPEC benchmarks. However, a majority of computer systems are subject to a different class of workloads where these common practices may not accurately reflect all performance issues. For example, operating system activity and context switches are ignored because many popular simulators and tracing techniques do not support the additional complexity. The main goal of the research is to understand the effects on the microarchitecture of operating system calls and context switches in a common computing environment. This work analyzes applications running in the ubiquitous Microsoft Windows environment using an x86 processor. Microarchitecture structures such as the instruction and data caches, TLB, and branch predictor are investigated in detail. The behavior of application and operating system code is studied to derive a complete picture of the execution behavior of these applications. In addition, a series of desktop and database applications are presented and compared with the SPEC CPU2000 suite. This analysis is conducted using a hardware tracer capable of tracing all activity including operating system calls and context switches. We observe that the dynamic instruction stream of desktop and database applications contain 19% to 78% operating system activity whereas SPEC2000 applications typically involve less than 1% operating system activity. Not only are there more operating system calls, the average number of instructions exe-