NexusLab
About
A deterministic C++20 digital twin for AI training clusters. Workloads, networks, scheduling, congestion, and failures in experiments you can run again and get the same answer.
Cluster behavior is expensive to learn from live GPUs. A hang, a bad schedule, or a congestion collapse is hard to replay, and most simulators hide the knobs that actually matter.
A discrete-event twin in modern C++ that models collective communication, datacenter networks, and schedulers as first-class objects. You set the experiment, you keep the seed, you read the trace.
Reproducible cluster experiments without burning a reservation. Public at github.com/abdillahinur/NexusLab.