Off-CPU Profiling and Latency Diagnostics in Java

Explore advanced profiling techniques for diagnosing latency issues in Java applications through this 42-minute conference talk from the 2025 JVM Language Summit. Learn why traditional CPU profiling tools like perf and async-profiler, while excellent for visualizing CPU usage through flame graphs, fall short in explaining transaction latency breakdowns, particularly when off-CPU factors such as I/O operations, scheduling delays, or contention are the primary culprits. Discover how many performance outliers—including slow requests, service invocations, and unexpectedly long GC pauses—stem from time spent blocked, waiting, or delayed by external systems rather than actual CPU processing. Examine a specialized tool designed to address these limitations by providing percentile-based views of tracked transactions, breaking each down into on-CPU time, scheduling delay, I/O wait, and additional metrics. Understand how this tool combines Java instrumentation with Linux uprobe technology to target specific parts of Java or C++ code, and see practical applications including pinpointing latency issues, analyzing virtual thread influences, understanding GC pause behavior, and comparing performance across different architectures. Presented by Yude Lin from the Alibaba Cloud JVM Team, this technical presentation offers valuable insights for Java developers and performance engineers working to optimize application latency beyond traditional CPU-focused profiling approaches.