SimAI is a full-stack, high-precision simulator designed for analyzing and optimizing large-scale AI training, particularly for Large Language Models (LLMs). It targets researchers and engineers seeking to understand and improve training performance by modeling various layers of the training stack, from framework parameters to network topology.

How It Works

SimAI integrates four core components: AICB for workload modeling, SimCCL for collective communication analysis, astra-sim-alibabacloud for network simulation, and ns-3-alibabacloud for detailed network communication modeling. This modular design allows for flexible simulation scenarios, ranging from fast analytical estimations using bus bandwidth to high-fidelity, full-stack simulations that capture intricate network behaviors. The project leverages extensions from astra-sim and integrates NCCL algorithms for realistic performance evaluation.

Quick Start & Requirements

• Install: Clone repository, update submodules, and compile using ./scripts/build.sh -c analytical or ./scripts/build.sh -c ns3.
• Prerequisites: Tested on GCC/G++ 9.4.0, Python 3.8.10, Ubuntu 20.04. NGC container images recommended for generation workloads.
• Usage: Analytical mode: ./bin/SimAI_analytical -w example/workload_analytical.txt -g 9216 -g_p_s 8 -r test- -busbw example/busbw.yaml. Simulation mode requires network topology generation and specific environment variables.
• Tutorials: Available for SimAI, aicb, SimCCL, and ns-3-alibabacloud.

Highlighted Details

• Full-stack simulation of LLM training processes.
• Supports analytical (fast estimation) and detailed simulation modes.
• Beta support for physical traffic generation in CPU RDMA cluster environments.
• Accepted paper at NSDI'25 Spring.

Maintenance & Community

• Active community engagement with past and upcoming technical presentations and workshops.
• Contact emails provided for questions. Community chat groups (DingTalk, WeChat) are available.

Licensing & Compatibility

• The README does not explicitly state the license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The "SimAI-Physical" mode is in beta and internal testing. The README does not specify the project's license, which could impact commercial adoption.