I work in an advanced development group at Digital. Our goals are to identify and develop technologies that will enhance the performance of Windows NT applications on Digital hardware platforms.

Developing software that will fully exploit the intrinsic performance capability of any modern microprocessor architecture is a major challenge that faces developers of both application and system software. For simple programs in which standard profiling techniques can be used to identify performance problems, the challenge is to write performance efficient code using algorithms that make good use of a machine s memory hierarchy and other architectural features. For complicated applications, identifying the parts of the application that are likely to cause performance problems can be a formidable challenge in itself. Database applications, for example, that are driven by user queries and dependent on local machine and storage configurations, can be difficult to characterize, much less tune for performance.

Our group is involved, to varying degrees, in the development of several performance tools that are being used at Digital to help identify and debug performance related problems on Alpha systems running Windows NT. Among these is PatchWrks, a tool which can be used to construct full instruction stream traces of the operating system and all of the applications running on a processor or multiprocessor system. PatchWrks, which runs on Windows NT, is useful for looking at snapshots of the overall system performance in minute detail to see the interaction and possible performance conflicts between software components. A second tool is NT-Atom, which runs on Unix as Atom and which our group is developing on Windows NT. NT-Atom is a versatile instrumentation tool that can be used for gathering performance information about dynamic program behavior. Our group has also been instrumental in bringing Digital s continuous profiling infrastructure (DCPI) to the Windows NT platform. DCPI can collect detailed execution profiles of unmodified Unix and Windows NT systems with low overhead. The profiles can be used to understand and identify performance problems for the entire system.

In addition to providing valuable performance information to software developers for applications and systems running on current Alpha platforms, some of our future work will focus on how to use these performance tools to drive optimizations in postlink optimization tools such as SPIKE. Finally, these performance tools are being used to develop representative information of complicated applications for use by the architects of future generations of microprocessors.

Sharon L. Smith
Digital Equipment Corporation
529 Bryant Street
Palo Alto, CA 94301
(415) 617-3475 (voice)
(415) 617-3654 (fax)
sharon@pa.dec.com