Build a cluster. Follow the work. Find the fault.
Explore BCM 11 provisioning workflows, eight-rail GPU networking and shared storage in an independent educational simulator. Practice in a hosted session with saved progress, a five-minute trial and optional prepaid hours.
Model, assumptions and supported commands · Switching labs · Back to the lab
Five minutes to explore. More time when you need it.
Try a hosted lab free for five minutes per verified account. After that, buy an hour for $20 USD. No subscription or automatic renewal.
Use your Wirewalk Institute account. Create an account, confirm your email or reset your password, then return here to sign in. No course purchase is required.
Sign in before launching a lab. Your free time starts only when you launch or resume the lab.
Pause stops time and freezes lab controls. Closing or losing the connection stops time within four seconds. One active lab per account. Progress is saved automatically; viewing it and exporting a session use no time. Access and prepaid-hour details.
Cluster configuration Choose your architecture
Your virtual cluster
Node readiness
UP means the modeled node daemon connected. Scheduler readiness also checks images, GPU health, rail mapping and mounts.
Recent model history
Ready nodesMean EW rail use
Event journal
Nodes and readiness gates
| Node | BCM workflow state | Installed / desired | GPU stack | Mount | Gate / allocation |
|---|
Eight rails. Three distinct network roles.
These bars show synthetic aggregate activity across the modeled cluster, not measured port throughput.
Storage
Different redundancy assumptions produce different outcomes here. This is not a performance or reliability comparison between products.
Images, canaries and rolling changes
This lab models preparing the GPU software stack explicitly in an image. The model guide links the BCM 11 image-management sources. Installed node copies change only when reimaged.
Promotion workflow
Prepare gpu-v2 → reimage an idle canary → pass readiness → promote category → roll two idle nodes at a time → check & resume.
Workload and recovery
Each job requests whole nodes: eight GPUs per node. Staging, compute, checkpoints and final result publication have separate dependencies. Work units and timing are illustrative.
State-aware training console
A finite command subset reads or changes the same virtual model as the controls. It is not an OS shell, a vendor CLI emulator or a way to reach a real host.
WIREWALK VIRTUAL CLUSTER — simulated outputs only. Type help to see supported commands. No command is executed on a host.
Break one dependency. Follow the consequences.
Inject or restore a fault
Suggested experiments
- Run a job, interrupt rail3 and inspect the failed collective. Repair it and explicitly requeue from the last committed checkpoint.
- Lose one NS member. Compare LACP capacity and active–backup behavior. Lose both; observe where application I/O waits.
- In Lustre, interrupt the active metadata service and follow modeled HA recovery. Then interrupt an OSS.
- In GPFS, lose one quorum voter, then two. Quorum and data-copy availability are separate gates.
- Try the unprepared gpu-v2 canary. Preparing the image does not modify its installed copy; reimage before promotion.
- Drain a busy node. Existing work may finish, but a new job must wait for suitable ready resources.
“Repair all” restores modeled faults but preserves images, drains and checkpoints. Failed jobs never become completed merely because a fault was cleared.