| Description |
Computer programs have complex interactions with their underlying hardware, exhibiting complex behaviors as a result. It is critical to understand these programs, as they serve an importantrole: researchers use them to express new ideas in computer science, while many others derive production value from them. In both cases, program understanding leads to mastery over these functions, adding value to human endeavors. Memory behavior is one of the hallmarks of general program behavior: it represents the critical function of retrieving data for the program to work on; it often reflects the overall actions taken by the program, providing a signature of program behavior; and it is often an important performance bottleneck, as the the memory subsystem is typically much slower than the processor. These reasons justify an investigation into the memory behavior of programs. A memory reference trace is a list of memory transactions performed by a program at runtime, a rich data source capturing the whole of a program's interaction with the memory subsystem, and a clear starting point for investigating program memory behavior. However, such a trace is extremely difficult to interpret by mere inspection, as it consists solely of many, many addresses and operation codes, without any more structure or context. This dissertation proposes to use visualization to construct images and animations of the data within a reference trace, thereby visually transmitting structures and events as encoded in the trace. These visualization approaches are designed with different focuses, meant to expose various aspects of the trace. For instance, the time dimension of the reference traces can be handled either with animation, showing events as they occur, or by laying time out in a spatial dimension, giving a view of the entire history of the trace at once. The approaches also vary in their level of abstraction from the hardware: some are concretely connected to representations of the memory itself, while others are more free-form, using more abstract metaphors to highlight general behaviors and patterns, which in turn characterize the program behavior. Each approach delivers its own set of insights, as demonstrated in this dissertation. |
