Several years ago, IBM added an option to record instruction counts in the SMF 30 data. They almost immediately realized that the instruction counts were skewed by I/O interrupts and so the recommendation was to not use them. More recently, IBM has started to say that maybe you can derive useful information from the instruction counts if you just discard the bad measurements. Does this make sense? Are there insights to be gleaned from those instruction counts? Can you use them to find inefficiencies? Or does the natural variability introduce too much noise to make sense of any signal that may be lurking in the data? In this session, we will explore what we have discovered by looking through data from multiple customers that have enabled the instruction counters in the SMF 30 data.