A discrete-event simulator for Kubernetes-native GPU scheduling. Zyvor Janus models clusters, MIG, topology, tenants, quotas, gang scheduling, and AI workloads — so you can develop schedulers, run RL research, and evaluate performance without physical GPUs. It is the digital twin of Zynera, Zyvor's production GPU/Kubernetes control plane.
0 GPUs required · 14 hardware profiles · 4 scheduler policies · Gymnasium + PPO · Apache-2.0
📖 Docs · zyvor.dev/zynera · Blog
| Change | What it does |
|---|---|
| GPU kinds | Hardware profiles for NVIDIA (H100 through B200, L4, A10G, RTX 4090), AMD MI300X/MI250, Intel Gaudi 3 and Apple M-series, plus a cluster (configs/clusters/gpu_kinds.yaml) that places one job on each kind |
| Shadow race | Pick a "Shadow vs" scheduler on Launch simulation and watch a live head-to-head race, then a full metrics comparison with a computed winner |
| Twin library | A /twins page in the web dashboard to browse calibrated GPU/model twins (GET /api/twins) |
| AIPerf overlay | Sim-vs-measured AIPerf overlay on the benchmark page |
| Rename | ZyForgeSim/ForgeSim is now Zyvor Janus: crates zyvor-janus-*, CLI zyvor-janus, Python zyvor_janus, env vars ZYVOR_JANUS_* |
Full history: CHANGELOG.md.
| When this happens… | Janus gives you… |
|---|---|
| Testing a scheduler change means booking scarce GPUs | Full discrete-event simulation of cluster placement, MIG slicing, NVLink/PCIe topology penalties and gang scheduling, with no GPUs |
| You can't tell how a policy would have done on production load | Import real FabricAIJob / FabricGpuNode / FabricQuota CRDs and replay production scheduler traces for oracle-vs-live diffing |
| Comparing policies means rewriting harnesses | fifo, priority, preemptive, bestfit and Zynera's own policy, swappable with one CLI flag |
| RL scheduling research has no realistic environment | A Gymnasium environment and a PPO baseline |
| Results live in log files nobody reads | Rich terminal dashboard, Next.js web UI (runs, benchmark, what-if), and an OpenAI-compatible inference shim for calibrated LLM serving metrics |
| A mixed fleet is hard to reason about | Shipped profiles for NVIDIA, AMD, Intel Gaudi and Apple M-series |
| Janus | kube-scheduler-simulator (kubernetes-sigs) | |
|---|---|---|
| What it simulates | GPU jobs over simulated time: arrival, placement, runtime, preemption, completion | The kube-scheduler's decisions for Pods in a simulated cluster |
| Scheduler | Its own policies: fifo, priority, preemptive, bestfit, zynera |
The real kube-scheduler and its plugins |
| GPU model | MIG slicing, NVLink/PCIe topology penalties, 14 hardware profiles | General Kubernetes resources |
| Workloads | Gang jobs, tenant quotas, synthetic LLM serving workloads, trace replay | Pods and nodes you create |
| Outputs | outputs/metrics.json: jobs completed, GPU utilization, topology penalties; trace diffs; benchmark score |
Per-plugin filter and score results in a web UI |
| Research | Gymnasium environment, PPO baseline, PyO3 bindings | Not its focus |
| Choose kube-scheduler-simulator when | You are tuning the default kube-scheduler and its plugins and want to see its real decisions |
- Rust core — event engine, cluster model, schedulers, metrics, Zynera bundle loader, inference timing model
- Python API — PyO3 bindings, Zynera CRD adapters, Gymnasium env, visualization, FastAPI server, AIPerf adapters
- Web UI — Next.js dashboard (runs, benchmark, what-if) + Rich CLI live dashboard
Design detail: docs/architecture.md.
git clone https://github.com/zyvorai/janus.git
cd janus
cargo run -p zyvor-janus-cli -- run --config configs/clusters/small_h100.yamlA full cluster simulation with no GPU, no Kubernetes, and one YAML file. Requirements: a Rust toolchain (pinned in rust-toolchain.toml); Python only for the bindings, RL and web API.
▸ Zynera export bundle — test Zynera without GPUs
mkdir -p zynera-export/{jobs,cluster,quotas}
kubectl get fabricaijobs -A -o yaml > zynera-export/jobs/all.yaml
kubectl get fabricgpunodes -o yaml > zynera-export/cluster/nodes.yaml
kubectl get fabricquotas -A -o yaml > zynera-export/quotas/all.yaml
cargo run -p zyvor-janus-cli -- run \
--zynera-bundle zynera-export \
--profiles-dir configs/profiles
# Or use the included fixture:
cargo run -p zyvor-janus-cli -- run \
--zynera-bundle tests/fixtures/zynera \
--profiles-dir configs/profiles▸ Scheduler policies — fifo · priority · preemptive · bestfit · zynera
cargo run -p zyvor-janus-cli -- run --config configs/clusters/priority_scheduler.yaml
cargo run -p zyvor-janus-cli -- run --config configs/clusters/preemption_preemptive.yaml
cargo run -p zyvor-janus-cli -- run \
--zynera-bundle tests/fixtures/zynera \
--scheduler zynera▸ Trace replay — compare vs production Zynera
cargo run -p zyvor-janus-cli -- replay \
--trace tests/fixtures/traces/fifo_match.jsonl \
--config configs/clusters/single_gpu.yamlWrites outputs/trace_diff.json with oracle vs FIFO placement diffs.
▸ MIG simulation — fractional GPU slices
cargo run -p zyvor-janus-cli -- run --config configs/clusters/mig_single.yaml▸ Dual-node preemption — placement migrate (not live CUDA)
cargo run -p zyvor-janus-cli -- run --config configs/clusters/dual_node_preempt.yamlThis is a digital-twin placement migrate. Zynera's production live migrate is KubeVirt VMs — see Zynera docs for Path A / Path B.
docker pull ghcr.io/zyvorai/zyvor-janus-api:latest
docker pull ghcr.io/zyvorai/zyvor-janus-web:latest
docker network create zyvor-janus 2>/dev/null || true
docker run -d --name zyvor-janus-api --network zyvor-janus -p 8080:8080 \
ghcr.io/zyvorai/zyvor-janus-api:latest
docker run -d --name zyvor-janus-web --network zyvor-janus -p 3000:3000 \
-e ZYVOR_JANUS_API_URL=http://zyvor-janus-api:8080 \
ghcr.io/zyvorai/zyvor-janus-web:latestOpen http://localhost:3000 (default login Admin / Admin@321 — override via ZYVOR_JANUS_DASHBOARD_USER / ZYVOR_JANUS_DASHBOARD_PASSWORD). Pin a release with :vX.Y.Z.
cd deploy/kubernetes
cp secret.example.yaml secret.yaml # edit credentials
kubectl apply -f secret.yaml
kubectl apply -k .See deploy/kubernetes/README.md.
cargo build --release -p zyvor-janus-cli
# Optional: Python bindings + web API
./scripts/setup_dev.sh
source .venv/bin/activate
pip install -e '.[server]'See CONTRIBUTING.md for viz, rl, and dashboard extras.
crates/ Rust workspace (core, topology, scheduler, simulator, CLI, API, PyO3)
python/ Python package, Gymnasium env, dashboard, baselines
web/ Next.js UI
configs/ Cluster YAMLs + calibrated profiles
tests/fixtures/ Zynera, traces, AIPerf, benchmark goldens
docs/ Architecture, milestones, UI, benchmark platform, deploy
See docs/zynera_input.md for CRD mapping rules, export workflow, and adapter levels.
See docs/milestones.md. M1–M8 complete, including topology runtime inflation, gang timeout, RL (M7), and visualization (M8).
Benchmark platform (MVP shipped): docs/benchmark_platform.md — inference model, serving traces, score vector, /benchmark + /what-if UI, OpenAI shim, AIPerf adapter, twin store API, CI golden script.
Schedulers: fifo, priority, preemptive, zynera (alias for preemptive), bestfit.
Janus is a simulator: the OpenAI-compatible shim returns analytical timing, and dual-node preemption is a placement migrate, not live CUDA migration.
Zyvor Janus is the free digital-twin simulator for Zynera. Janus validates scheduling policy offline; Zynera runs it against real GPUs.
| Zyvor Janus (this repo) | Zynera (zyvor.dev/zynera) | |
|---|---|---|
| What it is | Discrete-event simulator / digital twin | Production GPU/Kubernetes control plane |
| GPUs required | None — fully simulated | Real GPU fleet |
| Use case | Scheduler R&D, RL research, capacity planning, CI gates | Live cluster scheduling, MIG/topology placement, gang scheduling |
| Input | Zynera CRD export bundles, YAML configs, trace replay | Live cluster via Fabric CRDs |
| Support | GitHub Issues | SLA / onboarding — zyvor.dev/contact |
| Product | Role next to Janus |
|---|---|
| Janus | GPU scheduling digital twin: simulate, replay, benchmark |
| Zynera | The production control plane whose CRDs and traces Janus imports |
| Kairo | Pairs with Janus on Kubernetes: previews the blast radius and capacity impact of manifest changes before deploy |
Social assets: docs/social/.
Janus is free and open source under the Apache License, Version 2.0 (see NOTICE). Personal, lab, and commercial production use at no charge, subject to Apache-2.0 (preserve notices / NOTICE where required). That does not change.
Zyvor Enterprise adds what production teams ask for: supported releases, deployment and upgrade guidance, priority incident triage, a named technical contact and 24x7 critical intake. Production support, SLAs, and Zyvor Enterprise products are licensed separately. Plans and terms: docs/SUBSCRIPTION-MODEL.md · Pricing · [email protected].
Contributions: CONTRIBUTING.md.
