Mokka simulates the software contracts around NVIDIA devices rather than the devices themselves. The rule of thumb: reach for Mokka when your system reads hardware state and reacts to it, and for real hardware when it executes work, moves data, or measures performance.
That makes it a good fit for:
- Kubernetes discovery and allocation through the device plugin or DRA — scheduling, ResourceClaims, CDI visibility, topology attributes.
- Software that consumes NVML,
nvidia-smi, DCGM or DCGM Exporter. - Monitoring dashboards, parsers, alerting, and remediation logic that cordons, drains, reschedules and recovers.
- Repeatable fault injection — device loss, Xid errors, ECC errors, temperature, power, utilisation, clocks, and GPU-side NVLink errors.
- Code that interprets declared PCI, NUMA, NVLink, fabric UUID or clique topology.
- IMEX peer readiness and liveness over the pod network.
Use real hardware for CUDA execution, NCCL, GPUDirect and RDMA data paths, any throughput or thermal measurement, driver and firmware lifecycle, switch management planes, physically faithful fault timing, MIG partition lifecycle, and Confidential Computing.
kind create cluster --name mokka
helm install nvml-mock oci://ghcr.io/nvidia/k8s-test-infra/chart/nvml-mock \
--namespace mokka --create-namespaceEvery node now reports four mock GB300 GPUs. Swap in a100, b200, gb200,
h100, l40s or t4 with --set gpu.profile=<name>.
Simulation depth by area breaks this down per surface — GPU and NVML, Kubernetes allocation, metrics, PCI and NUMA, NVLink, NVSwitch and Fabric Manager, and InfiniBand — with what each one does and does not prove.
| Consumer | What works |
|---|---|
| Node Feature Discovery | PCI vendor labels derived from the feature file Mokka writes |
| GPU Feature Discovery | Node labels derived from NVML |
| NVIDIA Device Plugin | Allocatable nvidia.com/gpu matches the profile, and workloads schedule against it |
| NVIDIA DRA Driver | ResourceSlices report the right GPUs, and a ResourceClaimTemplate pod reaches Running |
| NVIDIA GPU Operator | The full operand stack installs and its validator starts |
| DCGM / dcgm-exporter | Telemetry, time-varying power, and injected Xid errors |
| Run:ai fake-gpu-operator | Profile ConfigMaps published in the shape its discovery expects |
Visit our documentation website to find all details about how Mokka works.
| Quick Start | Install and see simulated GPUs |
| Architecture | The moving parts and how the system behaves |
| Guides | Device plugin, DRA, GPU Operator, failure injection, node-wide injection |
| Configuration | Every profile knob |
| FAQ | What Mokka simulates, and what it does not |
See the contributing guide for local development with Tilt, the test suites, and how to submit a change. Substantial changes start with a Mokka Enhancement Proposal.
Report vulnerabilities privately — see SECURITY.md, not the issue tracker.
- Logo designed by Roman Hlushko with the assistance of OpenAI's ChatGPT.
Apache License 2.0 — see LICENSE.