Project topic posted by Pat Morin

A Less Intrusive Profiler

A profiler is a system that monitors a program to see how long the that program spends in each of its subroutines. Profilers are used for finding frequently executed sections of code (hot spots) so that they can be optimized. Profilers work either by stopping the program at some random times and recording the current instruction pointer (subroutine) or by inserting code into the program so that each invocation of a subroutine is recorded. In both cases, all this data is stored somewhere and post-processed to come up with the program's "profile". The problem with this approach is that profilers require some place to store all this data. This means that the machine the program is running on must either have a lot of memory or a hard disk. This makes profiling embedded software very difficult. It also means that profiling a program may change its behaviour since the program being profiled has less available memory or is frequently halted while the profiler writes data to disk.

For this project I would like to treat a profiler as a program that processes a data stream consisting of the subroutines (or even individual lines of code) that are executed. Recently, many researchers have proposed algorithms for finding frequently occuring items in data streams and these algorithms require only a small amount of memory that is independent of the length of the data stream. As an example, it is possible to find all items in a data stream that occur more than 10% of the time by using only 10 counters and labels. Using these algorithms we could develop a profiler that does not require much (internal or external) memory or time. P. Bose, E. Kranakis, P. Morin, and Y. Tang. Bounds for frequency estimation of packet streams. In Proceedings of the 10th International Colloquium on Structural Information and Communication Complexity (SIROCCO 2003), 2003.