Earticle

초록

영어

The processor's Hardware Performance Counter (HPC) is a special purpose register that counts hardware-related events. Although HPC monitoring is recognized to have less overhead than other profiling techniques, but it is still an important issue in using HPC with a short monitoring interval. In this paper, we profile and analyze the overhead of monitoring HPC according to measurement schemes: the number of events in groups, the number of simultaneously active monitoring targets, and the reading method. For the number of events in the group, overhead per event increases by 5-6%. In the case of the number of simultaneous active subjects, overhead increases by 0.8-1% per core and 0.1-0.2% per process. The read method has no consistent relationship with overhead. Therefore, in order to optimize HPC monitoring in consideration of overhead, after considering the number of events and the number of active monitoring targets in order, the reading method is selected according to the situation considering usability.

목차

Abstract
I. INTRODUCTION
II. BACKGROUND AND RELATED WORK
A. Measuring Hardware Performance Counter
B. Configurable options of perf_event_open
III. EXPERIMENT
A. Controlled configuration options
B. Experiment Environment and Setting
C. Results
IV. CONCLUSION
ACKNOWLEDGMENT
REFERENCES

저자

• Minkyu Song [ Department of Computer Science and Engineering Korea University ]
• Taeweon Suh [ Department of Computer Science and Engineering Korea University ]

참고문헌