First at all, you may read book <Intel® 64 and IA-32 Architectures Optimization Reference Manual> if you have time (which can be download from Intel website). For saving your time, you also can read <Tuning Guides and Performance Analysis Papers> for different processor type.
iliyapolak is right that you have to consider stalls from the front end and back end. For example, stall from instruction fetching due to Instruction cache miss, IDQ is buff, branch misprediction, ROB is full, Resource Allocation Table is not usable to Reservation Station, and RS can not dispatch uops to execution unit since Port 1-n are busy. At the back end, maybe Load/Store buffer will cause delay as well as Data Cache Miss (also TLB miss), nonalignment data access, 4K-aliasing, etc . Also, uops retirement (in ROB) might cause delay (piror instruction is retired), etc.
To know which events name supported in your platform will cause stall, you can use "amplxe-runss -event-list | grep STALL" . However some events name without "STALL", for example L2_MISS also will cause penalty - you need to read guideline or optimization user manual.
Regards, Peter
