Today we went back and discussed the prior papers a bit more, then looked at the ones by Lizzy John (about subsetting SPEC2006) and Xi Yang (about the SHIM tool). We saw that SPEC suffers from a lack of diversity in terms of how it stresses the processor, which enables subsets that capture most of the information that we woudl gain from the full suite. We also saw how running a tool in a separate thread or on another core to access counters can provide reasonable fine grained data about perfromance, but the finer the granularity, the greater the disruption and the more data we have to discard and contaminated by the sampling process. 

Slides are here