cluster-scheduling

Official artifact for MoSim, a trace-driven GPU-cluster simulator with dynamic network contention modeling (IEEE MASCOTS 2026).

Generate reproducible deep-learning workload traces by mapping production GPU-cluster traces to profiled executable workloads.

A learned wait-time score cannot rank a queue by anything except requested size — a negative result with zero counterexamples over 45,432 dispatch instants on the reference platform (45,268 and zero on Linux). Every number regenerates from one command.

Sanitized cluster resource forecasting and placement simulation lab with synthetic demo