Off-CPU Profiling & Latency Diagnostics in Java #JVMLS

The podcast delves into CPU profiling and latency diagnostics within Java environments, particularly in cloud-based systems. It addresses the shift from GC-related latency issues to off-CPU factors like scheduling and I/O as Java becomes more "postless." The discussion highlights the importance of I/O awareness, using a Spark TPCDS benchmark example where JVM's throughput-focused heuristics negatively impacted performance due to memory allocation. The presenter introduces RT Radar, a tool developed to dissect transactions, revealing dependencies between threads, scheduling costs, and wait times. The tool helps identify bottlenecks and guide optimizations, as demonstrated through a spec.jbb issue, showcasing how a quantitative approach can pinpoint root causes of latency degradation.