Create · connect · simulate
Failure & repair experiment
Build CPU/GPU nodes, connect workloads and share cluster resources. Inspect queues, memory and request histories as the system runs.
Default experiment: compute finishes, the result store fails, and an expired lease triggers repeated work. Compare bounded admission and idempotent commits with the same arrivals and faults.
Executable architecture
Links serialize 1 MiB per transfer with shared bandwidth and propagation delay. GPU clusters share a compute budget. No replication or consensus protocol is modeled.
Every request stays visible
Including rejected, dead-letter and unfinished work