Counting performance: hardware performance counter and compiler instrumentation.
Jan-Patrick Lehr · TUbilio (Technical University of Darmstadt) · 2016
Analyzing applications for their runtime behavior,e specially in the light of ef®cient resource utilization, involves iterative measurements and the interpretation of the data gathered.For ®ne grained analysis often hardware performance counters are monitored.Since compiler instrumentation augments the program with calls to ameasurement system, it interacts with compiler optimizations.Currently,itisunclear to which degree the instrumentation in¯uences the characteristics of the resulting binary.T odetermine the behavioral change introduced through instrumentation we conduct aseries of measurements on asubset of the SPEC CPU 2006 benchmark suite.We showthat although runtime increases up to acceptable 10% of the original runtime some hardware performance counter are signi®cantly perturbed.However, it is also possible that hardware performance counter deviate only slightly from the values measured in the original binary,e venthough the benchmark's runtime increases substantially.Inparticular,the program 444.namd showed an increase in store instructions by 2x with only 3% runtime overhead, whereas 450.soplex did not showsigni®cant change in mispredicted branches, when exhibiting an increase in runtime by 3x.W einvestigate whether the validity of ah ardware performance counter can be determined via static analysis.Therefore, we outline an ew tool based on the MAQAOb inary analysis framework to compute and compare the static instruction mix of binaries to identify the introduced change in the static instruction mix due to instrumentation.The static instruction mix is comparable to ah istogram, denoting howm any instructions of ac ertain category are found in each function of the binary.W ec onclude that static analysis of the binary'si nstruction mix may describe induced change for hardware performance counters.Finally,weoutline directions of further research in the ®eld of perturbation detection and predictive modelling.