Linhai Song - Research Statement

My main research interests are software system, program analysis, and software engineering. The goal of my research is to help programmers build more efficient software systems. Everyone wants software to run faster. Slow and inefficient software can easily frustrate end users and has already caused several highly publicized failures. Although researchers devoted decades to improving software performance transparently, performance bugs continue to pervasively degrade performance and waste computation resources in the field. Performance bugs are software implementation mistakes, where relatively simple source code changes can speed up software, while preserving functionalities. The current tool support for performance bug is preliminary due to the poor understanding of performance bugs. There are three trends making performance-bug problems even more urgent in the future. Firstly, in the multi-core era, when each core is unlikely to become faster, performance bugs are particularly harmful. Secondly, the increasing complexity of software systems, rapidly changing workloads and frequently updated hardware provide new opportunities for performance waste and new challenges in diagnosis. Thirdly, increasing energy costs provide a powerful economic argument for avoiding performance bugs. Combating performance bugs is critical. My Ph.D. research has mainly focused on performance bugs. My experience spans different stages of combating software bugs: real-world bug understanding, bug detection, compiler optimization, failure diagnosis, and patch generation. In particular, I conduct the first characteristics study on 110 real-world performance bugs from 5 representative open-source software suites [6]. I build a static performance bug detection tool suite, based on extracted efficiency rules from fixed performance bugs [6], and help to build a dynamic performance bug detection tool for inefficient nested loops [5]. These two techniques find hundreds of previously unknown performance bugs, many of which have already been confirmed and fixed by developers. I design three source-to-source compiler optimizations to transparently eleminate performance bugs in applications offloading computation to Intel Xeon Phi Coprocessor [3], and this work won MICRO’2014 best paper nomination. For production-run performance failure diagnosis, I provide a statistical debugging solution to conduct fault localization [4], and design staticdynamic hybrid analysis to further diagnose identified inefficient loops [2]. Besides combating performance bugs, I also work on a concurrent bug fixing system [7], which won a SIGPLAN CACM research highlight nomination [1]. My future research will continue to focus on improving the understanding of performance problems and providing better tool support for developers to build efficient software, including an on-line algorithmic profiling system, a tool chain to better balance precision and performance for floating-point applications, and an empirical study for performance problems in web applications.