[Abstract]

Modern applications face performance bottlenecks both on and off the CPU, yet traditional profiling tools fail to capture these events comprehensively. This seminar introduces blocked samples, a novel sampling-based profiling technique designed to analyze both domains. We will discuss two profilers, bperf and BCOZ, which leverage blocked samples to provide symbol-level insights and causality analysis of performance bottlenecks. Validated with real-world applications, these tools not only identify off-CPU issues like I/O waits and blocking synchronization but also predict optimization outcomes with high accuracy.

[Biography]

Jinkyu Jeong is an Associate Professor in the Department of Computer Science and Engineering at Yonsei University. Prior to joining Yonsei, he was an Assistant and Associate Professor at Sungkyunkwan University. He received his Ph.D. in Computer Science from KAIST and his B.S. from Yonsei University. His research interests span operating systems, systems software, cloud computing, and mobile systems. More recently, his work has focused on systems for AI, performance profiling, and high-performance AI-driven code generation. He has actively contributed to the systems research community through organizing roles in international conferences such as SOSP, APSys and NVMSA, as well as serving on program committees of conferences including FAST, USENIX ATC, HPCA, and MICRO. He has received multiple awards, including Outstanding Lecturer awards from Yonsei University and Best Paper/Presentation awards at various conferences. He has published over 70 papers in premier venues such as OSDI, USENIX ATC, FAST, ISCA, HPCA, ASPLOS, as well as SCI(E)-indexed journals.