Fully Automated Software Logging

点击翻译按钮获取中文摘要

This research addresses a fundamental problem in computer science. Programmers today spend 60% of their time in debugging; software industry spends $153 billion on failure diagnosis annually. When failures bring software down, especially systems software (e.g., operating system) that hosts all applications, no useful work can be produced. This research aims to speed-up the postmortem failure diagnosis process on systems software to minimize software's downtime. When software systems fail in production environments, log data is often the only information available to programmers for postmortem diagnosis. Programmers place log printing statements in software programs (e.g., using printf). At runtime, these statements output diagnostic information to log file. What makes these logs so valuable is their ubiquity and commercial acceptance. It is an industry-standard practice to request logs when a customer reports a failure and, since their data typically focuses narrowly on issues of system health, logs are generally considered far less sensitive than other data sources. Moreover, since logs are typically human readable, they can be inspected by a customer to establish their acceptability. Many software vendors even allow logs to be transmitted automatically and without review. Consequently, the quality of the log data is of critical importance to postmortem debugging. Indeed, if the log is uninformative, or even worse, misleading, postmortem debugging can end up being a wild goose chase. Unfortunately, little attention has been paid to the quality of log. While logging is pervasive it is every programmer's everyday task to decide where to log in the program -- there is very little guideline or established best practices on logging. Worse, programmers often ignore the quality of logging code when they are under time pressure of software release. As a result, it is frequently the case that when a failure occurs, the log does not contain any failure-related information, a situation referred colloquially as “debugging in the dark”.This research proposes to fully automate software logging. It addresses the fundamental problem of where to place logging statements in the program. This research proposes a series of algorithms that are capable of automatically placing logging statements at optimal program locations. It has three specific objectives: (1) logging for failure diagnosis purpose, (2) logging for performance profiling, and (3) logging for security auditing. The key challenge is to measure how informative is logging statements being placed, and to trade-off informativeness with the performance cost of logging. The core idea is to use Information Theory to measure the entropy of the software, and calculate how different logging placement strategies can reduce this entropy.