Studies of Windows NT performance using dynamic execution traces

Abstract

We studied two aspects of the performance of Windows NT: processor bandwidth requirements for memory accesses in a uniprocessor system running benchmark and commercial applications, and locking behavior of a commercial database on a small-scale multiprocessor. Our studies are based on full dynamic execution traces of the systems, which include all instructions executed by the operating system and applications over periods of a few seconds (enough time to allow for significant computation). The traces were obtained on Alpha PCs, using a new software tool called PatchWrx that takes advantage of the Alpha architecture’s PAL-code layer to implement efficient, comprehensive system tracing. Because the Alpha version of Windows NT uses substantially the same code base as other versions, and therefore executes nearly the same sequence of calls, basic blocks, and data structure accesses, we believe our conclusions are relevant for non-Alpha systems as well. This paper describes our performance studies and interesting aspects of PatchWrx. We conclude from our studies that processor bandwidth can be a first-order bottleneck to achieving good performance. This is particularly apparent when studying commercial benchmarks. Operating system code and data structures contribute disproportionately to the memory access load. We also found that operating system software lock contention was a factor preventing the database benchmark from scaling up on the small multiprocessor, and that the cache coherence protocol employed by the machine introduced more cache interference than necessary.