Tracking Memory Accesses to determine Working Set Size

Fast memory is a valuable resource that has to be managed carefully. Specifically, sharing caches efficiently is not a trivial task. To do so, the amount of memory that is actively used by an application, its so-called working set size, must be computed regularly. However, this number cannot generally be determined from established hardware counters.
In this thesis a mechanism that could eventually be used in hardware to determine the size of the working set, is designed. In order to minimize the amount of memory needed for that mechanism, bloom filters are used. They will act as the storage for encountered memory elements and should make a hardware implementation feasible through their space-efficiency.
To remain independent from specific hardware, the mechanism is implemented as a tool for the dynamic, binary instrumentation framework Valgrind. The tool plugin will simulate the potential dedicated hardware and give insight into important characteristics, such as the length of the interval the working set is based on, or the size of the used bloom filter.

(Masterverteidigung)