Blackboxing Performance Monitoring Units

Due to the complexity of modern hardware, computer programs often cause complicated interactions when scheduled together. Users therefore frequently turn to specialized tools such as Performance Monitoring Units (PMUs) for diagnosing system issues. However, PMUs themselves are complex, so they require domain knowledge and a good understanding of the underlying hardware to use. Such obstacles prevent PMUs from seeing more widespread use. In this thesis, a survey of PMUs investigates the various intricacies involved in selecting performance events and configuring performance counters. Major problems include the limited number of performance counters relative to the number of measurable performance events, abstruse definitions of performance events, and the processor-specific nature of PMUs. Numerous grouping techniques which associate similar objects are explored with a view of reducing the event space. We also experiment with classification algorithms to automatically detect interference and remove the need for human interpretation of events and counter measurements. Moreover, proposals in this thesis are designed with portability, i.e., the applicability of methods to different machines, in mind. All in all, we show the complexity of PMUs is a manageable problem and select procedures can be automated, which opens up the possibility of abstracting away PMU details from users.