A tutorial on scalable system simulations for RISC-V architectures and performance analysis for machine learning workloads

This tutorial aims to provide a comprehensive introduction to computer system simulations and performance analysis, focusing on the integration and application of consensual open-source frameworks like Structural Simulation Toolkit (SST), Gem5, and Multi-Level Intermediate Representation (MLIR). Participants will gain hands-on experience in conducting an architectural design exploration with a pragmatic approach where the simulation framework used is chosen based on a trade-off between fidelity and scalability requirements. By the end of the tutorial, participants will be equipped with the skills necessary to conduct in-depth performance analysis and optimize complex systems using state-of-the-art tools.

Contents:

• Application-oriented system modeling and optimization
  • Prerequisites
    • Containerized environment
  • Environment Setup
    • Part 1: Prepare benchmark
    • Part 2: Compile IREE run module
    • Part 3: Compile m5 utility
    • Part 4: Prepare RISC-V disk image
  • Machine Learning Workload Execution
  • Extra: Checkpoints
  • Experimental Studies
    • Part 1: Change CPU model
    • Part 2: Change cache hierarchy
    • Part 3: Vectorization
    • Part 4: New benchmarks
• Installation instructions for scale-out system simulation
  • Install instructions for SST
  • Install instructions for rv64 mpi compiler
    • Scale-out system simulation with SST
• Instruction-level simulation
  • Environment Setup
  • System under exploration
  • Workload under evaluation
  • DEMO
  • References
• Packet-level simulation
  • Environment Setup
  • Parallelism in Large Language Model training
  • DEMO
• About us - CSA, imec