Workshop: Profiling JVM Applications in Production

Profiling JVM applications it not very easy in the first place. Many profilers, such as JVisualVM and jstack, will simply lie to your face about which call stacks are hottest and where your bottlenecks lie. Profiling in production environments has even more challenges, because you have to carefully manage overhead, account for other processes running on the system, and choose non-invasive tools that don't require an application restart. This is the workshop for you -- we will learn, understand, and experiment first-hand with JVM profiling tools designed for production use.

We will build a simple checklist for verifying JVM application performance, and finding the area to focus on in a closer investigation. Then, we will experiment with two approaches for CPU profiling on Linux: the perf multi-tool, combined with perf-map-agent, and the async-profiler project, an innovative tool that brings perf together with traditional JVM profiling techniques. We will visualize stack traces using flame graphs, and understand where the CPU bottlenecks lie, through a series of hands-on labs. In the second half of this workshop, we will talk about more complicated scenarios: diagnosing errors when opening files, tracing database queries, monitoring system I/O load, understanding the reasons for excessive garbage collection, figuring out why threads are blocked off-CPU, and more.

Outline:

• Linux and JVM performance information sources: tracepoints, {k,u}probes, USDT, JMX, JVMTI, Java SA, JFR, GC/JIT logs
• Linux JVM performance checklist: top, sar, free, iostat, pidstat, vmstat, jstack, jstat (based on hsperfdata), jcmd, jattach
• Linux perf, CPU sampling, getting stack reports, problem with symbols
• Generating perf maps with perf-map-agent, inline frames, source info
• Visualizing JVM profiles with flame graphs
• JVMTI profiling with an agent & jattach: GetAllStackTraces, exceptions, class loads, monitor contention [link] etc.
• Full-stack JVM profiling with async-profiler, how it works, pros and cons
• Introduction to BPF
• BPF scenarios and the BCC toolkit, Java-specific tools, JVM USDT probes, -XX:+ExtendedDTraceProbes
• Dedicated BCC tools: fileslower, opensnoop, gethostlatency, runqlat, cpudist
• General-purpose BCC tools: trace, argdist, funccount, stackcount, funclatency
• Heap allocation profiling (based on TLAB sampling) with async-profiler vs. funccount/stackcount vs. grav