Towards automatic detection and prioritization of pre-logging overhead: a case study of hadoop ecosystem

摘要

The pre-logging overhead refers to the overhead produced by pre-logging statements (PLS), which are used to construct the parameters of logging calls. In practice, these logging-related statements are usually guarded by conditional statements, known as logging guards, to ensure that they are only executed when necessary (e.g., at the debugging phase). However, developers occasionally forget to add logging guards to costly PLS, resulting in missing logging guards (MLG) issues, which can have a significant impact on performance, particularly for high-throughput software, and lead to critical performance issues. In this paper, (1) we conduct the first empirical study of 137 commits addressing MLG issues in five popular open-source software of the Hadoop ecosystem. Based on the results, we reveal five findings of the current practice of logging guards. (2) We devise an accurate algorithm to detect PLSs (over 95% in precision and recall) and find out 16 problematic partially guarded logging calls (10 of them are confirmed and fixed by developers). (3) We investigate two metric-based ranking approaches using six software metrics to prioritize PLSs based on their impact on performance. We discover that the execution frequency of PLSs achieves the best performance, and combining multiple software metrics can improve performance (7.8% on average).