CPU Experiment Behavior in Gremlin

As opposed to other resource attacks, CPU experiments in Gremlin consume an additive amount of resources rather than targeting an absolute utilization level.

For example, if you select 50% for a memory attack, Gremlin will consume 50% of the total memory accessible to the Gremlin agent.

Conversely, CPU attacks are additive. If a system is already consuming 75% CPU during normal operation, running a 50% CPU attack will request an additional 50% of CPU capacity, resulting in total demand equal to 125% of available resources.

This difference reflects the nature of the underlying resources and how operating systems handle resource exhaustion.

CPU is a compressible resource: when demand exceeds supply (e.g. our 125% example above), operating systems can throttle or deprioritize processes rather than terminate them outright, and the exact behavior is often configurable.

Using additive CPU consumption more accurately simulates “noisy neighbor” scenarios, where other workloads compete with the target application for CPU time.

Limiting CPU consumption to a fixed absolute percentage would reduce pressure on the target application and would less accurately reflect real-world contention scenarios.