Date of this Version
Profiling deployed software provides valuable insights for quality improvement activities. The probes required for profiling, however, can cause an unacceptable performance overhead for users. In previous work we have shown that significant overhead reduction can be achieved, with limited information loss, through the distribution of probes across deployed instances. However, existing strategies for probe distribution do not account for several relevant factors: acceptable overheads may vary, the distributions to be deployed may be limited, profiled events may have different likelihoods, and the user pool composition may be unknown. This paper evaluates strategies for probe distribution while manipulating these factors through an empirical study. Our findings indicate that for tight overhead bounds: 1) deploying numerous versions with complementary probe distributions can compensate for the reduction of probes per deployed instance, and 2) techniques that balance probe allocation, consider groupings and prior-allocations, can generate distributions that retain significantly more field information.