Skip to main content

Linux JVM CPU stress

Last updated on

Linux JVM CPU stress is a chaos fault that uses Byteman to pin CPU cores of busy work inside the target Java process for DURATION, then releases the threads. The target Java process is selected by PID, by STARTUP_COMMAND, or by attaching to a running Byteman agent on PORT. The fault runs through the Linux Chaos Infrastructure (LCI) systemd service installed on the target VM.

Use this fault to test how a Java workload behaves when its own threads pin the CPU: whether GC keeps up, whether request handlers slow down, whether thread-pool exhaustion appears, and whether monitoring detects the in-JVM pressure within the alerting SLA.

Run your first experiment

If you have not installed the Linux Chaos Infrastructure yet, go to Linux Chaos Infrastructure to install the agent and connect the VM to the control plane.

Use cases​

Run this fault when you want to answer concrete questions like:

  • CPU tail behavior: When CPU JVM threads pin cores, does the request handler tail (p99) stay inside its SLA?
  • GC pressure: Does GC absorb the additional CPU pressure, or does it degrade throughput further?
  • Thread-pool starvation: Do worker pools or async executors back up when CPU is starved inside the JVM?
  • Monitoring fidelity: Do alerts on JVM CPU, request latency, and queue depth fire within the alerting SLA?

Prerequisites​

  • Linux Chaos Infrastructure installed: The linux-chaos-infrastructure systemd service is active on the target VM and the infrastructure is in CONNECTED state. Go to Linux Chaos Infrastructure to install it.
  • Target JVM identifiable: Provide one of PID or STARTUP_COMMAND, or ensure a Byteman agent is already listening on PORT.
  • JAVA_HOME reachable: The fault uses Byteman, which requires JAVA_HOME resolvable to the JDK that the target process is running. Set JAVA_HOME if it is not on PATH.
  • Byteman port reachable: PORT (default 9091) must be free or already hosting the Byteman listener for the target JVM.

Supported environments​

The fault has been tested on the following Linux distributions. Go to Linux fault requirements to see the full compatibility matrix.

PlatformSupport status
Ubuntu 16+, Debian 10+Supported
CentOS 7+, RHEL 7+, Fedora 30+Supported
openSUSE LEAP 15.4+ / SUSE Linux Enterprise 15+Supported
JVM versionsOpenJDK 8, 11, 17, 21 (any JVM compatible with Byteman)

Permissions required​

This fault is classified as a Basic Linux fault. It runs with the privileges of the Linux Chaos Infrastructure systemd service (root user and root user group) on the target VM. The LCI service must have permission to attach to the target Java process. No cloud credentials are needed.

Fault tunables​

Configure the following fault parameters when you add Linux JVM CPU stress to an experiment in Chaos Studio. Defaults are shown for reference.

JVM selectors (provide one or rely on PORT)

TunableDescriptionDefault
PIDPID of the target Java process. Set to 0 to fall back to STARTUP_COMMAND or PORT.0
STARTUP_COMMANDSubstring of the Java process command line used to identify the target.""
PORTPort of the Byteman agent. The fault installs Byteman on this port if no listener exists.9091
JAVA_HOMEPath to the JDK used by the target JVM. Required if JAVA_HOME is not on the LCI service environment.""

Chaos parameters

TunableDescriptionDefault
DURATIONTotal duration of the fault. Accepts [hours]h[minutes]m[seconds]s format.30s
CPUNumber of in-JVM threads that pin CPU cores.2
RAMP_TIMEWait period in seconds before and after the fault. Go to ramp time to read how it is applied.0

Tunables that apply to every fault are documented in common tunables for all faults.

Fault execution in brief​

Attaches Byteman to the target JVM on PORT, installs a rule that spawns CPU busy threads for DURATION, then removes the rule and releases the threads.

Expected behavior during fault execution​

  • CPU usage attributed to the target Java process rises sharply.
  • Request latency inside the JVM grows; p99 shifts upward as worker threads compete for CPU.
  • GC may extend pause times if it competes with the busy threads.
  • After the duration ends, the Byteman rule is removed and the busy threads exit.
When the fault ends

The chaos pod removes the Byteman rule and the busy threads exit. CPU returns to baseline within the JVM and on the host.

Signals to watch​

Attach resilience probes to assert each layer:

  • JVM CPU: Use a Prometheus probe on process_cpu_seconds_total{job="<app>"} and assert it rose.
  • GC pause: Use a Prometheus probe on jvm_gc_pause_seconds and assert the maximum stayed inside the SLA.
  • Request latency: Use an HTTP probe on a user-visible endpoint.

Verify the fault execution effect​

  1. Observe live CPU on the target JVM.

    sudo top -bn1 -p <pid> | tail
    sudo jstack <pid> | grep -E "RUNNABLE|stress" | head -20

  2. Inspect Byteman state.

    sudo $JAVA_HOME/bin/bmtool -p <PORT> -l

    The injected rule should be listed during the chaos window and removed afterwards.

  3. Inspect Linux Chaos Infrastructure logs.

    sudo journalctl -u linux-chaos-infrastructure -n 100 --no-pager

Recovery and cleanup​

  • End of duration: The chaos pod removes the Byteman rule when DURATION elapses.
  • Abort the experiment: Stopping the experiment from Chaos Studio also removes the rule.
  • Manual recovery: If the rule survives an abort, remove it manually with sudo $JAVA_HOME/bin/bmtool -p <PORT> -u <rule> (rule name is recorded in the LCI logs). If the Byteman listener itself is orphaned, restart the target JVM.

Limitations​

  • JVM-only: The fault stresses CPU inside the target JVM only; non-JVM processes on the same host are not affected.
  • Byteman dependency: The target JVM must allow Byteman attachment; some hardened JVMs (sealed containers, AppArmor-restricted) reject attachment.
  • Single JVM scope: Each fault run targets one Java process.
  • CPU vs cores: CPU controls the number of busy threads but cannot exceed the JVM's effective CPU limits (cgroups, taskset).

Troubleshooting​

Linux JVM CPU stress fault could not attach to target JVM in Harness Chaos Engineering

Confirm PID or STARTUP_COMMAND resolves to a running Java process and that the LCI service has permission to attach (same user namespace, no AppArmor block). Set JAVA_HOME explicitly to the JDK matching the target JVM.

No measurable CPU rise inside the JVM

Confirm CPU is greater than 0. If the JVM is cgroup-restricted to one core, only one busy thread runs at a time. Increase CPU or run the JVM with more cores.

CPU stayed elevated after the experiment ended

If the Byteman rule was not removed, list rules with sudo $JAVA_HOME/bin/bmtool -p <PORT> -l and remove with -u <rule>. If the rule cannot be removed, restart the target JVM to clear all injected rules.