How mlnetwork/rostam works
Project type and purpose
• From the file tree and README, “rostam” is a command-line driven simulation framework for multi-GPU training jobs over different interconnect topologies (electrical, ring, fullmesh, optical/OCS).
• It ingests pre-profiled model execution graphs, performs operator placement (optionally via Gurobi-powered ILP), schedules operations across devices, and simulates communication over a chosen network topology to produce performance metrics.
File structure
– Top-level “src” directory splits into cohesive modules:
• compute: GPU device abstraction
• executor: loads graph profiles, drives session execution, tracks op scheduling
• graph_def: protobuf definitions for input profiles
• interconnect: implementations of various network topologies and packet/transport logic
• placement: operator partitioning and device-mapping strategies (DP, MP, base, ILP via Gurobi)
• sipml: command-line frontends (sipml-elect, sipml-ring, etc.)
• utils: shared utilities (config parsing, matrix math, exit codes)
• scripts: profile-generation Python scripts for ResNet, Transformer, etc.
– Build artifacts: autotools files (configure.ac, Makefile.am), autogen.sh.
README for additional insights
– The framework supports optional Gurobi for ILP-based placement; falls back to presets if unavailable.
– Input profiles are protobuf files plus iteration-level stats.
– CLI tools expose flags for number of GPUs, bandwidth, topology, solver method, input profile path.