unisa-hpc / synergy-sc-ae

SYnergy SC23 Artifact Evaluation

SYnergy SC23 Artifact Evaluation

This repository provides the details for reproducing the results of the SC23 paper SYnergy: Fine-grained Energy-Efficient Heterogeneous Computing for Scalable Energy Saving.

OS Requirements

Our experiments have been tested on Ubuntu 20.04, 22.04, and RHEL 8.1.

Hardware Requirements

• Single-node experiments: at least one NVIDIA GPU is required.
• Multi-node experiments: a cluster with NVIDIA GPUs is required, equipped with the provided NVGPUFREQ SLURM plugin.

Software Requirements

• DPC++ (Intel/LLVM) 2022-09
  • Install using the Getting Started Guide
• Clang and LLVM 15
  • Install using the LLVM automatic installation script
  • The opt-15 tool must be available in the system PATH
• CUDA Toolkit (tested with CUDA 11.8)
• Python 3
  • Install with sudo apt install python3
  • Required packages: scikit-learn>=0.24, pandas, numpy, matplotlib, paretoset
    • Install with pip install scikit-learn>=0.24 pandas numpy matplotlib paretoset
• cmake 3.17 or later
  • Install with sudo apt install cmake
  • Alternatively, download the latest stable release
  • Check that cmake version is >= 3.17 using cmake --version

Required for multi-node experiments:

• NVGPUFREQ SLURM plugin
  • Follow the instruction in the readme file of the repository
• Application libraries
  • Install with sudo apt install libpnetcdf-dev (for MiniWeather)
• MPI Implementation (tested with Spectrum MPI)
  • Install with sudo apt install openmpi-bin libopenmpi-dev

How to use this repository

This repository is divided in four directories:

• training-dataset, it contains all the scripts required to generate the data on which the models are trained
• testing-dataset, that contains the scripts to run the SYCL-Bench suite to gather validation dataset and multi-objective characterization plots
• models-validation, provides the scripts for models training and inference based on the previous datasets and the scripts to reproduce the validation results
• energy-scaling, that provides the scripts to launch the MiniWeather and CloverLeaf applications to reproduce the energy scaling results

Each subdirectory has its own README.md file that provides additional information.

Using pre-generated data

As running the tests may take some time and some specific requirements, we provide our data to obtain the exact same results of the paper.

This workflow does not run any application, but uses the data obtained during our experimental analysis. To use this workflow visit the testing-dataset, models-validation and energy-scaling folders and follow the readme files.

Reproduce the results of the paper

In order to reproduce the results without the pre-generated data, make sure that all the requirements are fulfilled. This workflow requires to visit the folders in the following order:

1. training-dataset
2. testing-dataset
3. models-validation
4. energy-scaling

Follow the steps defined in the respective readme files.

Additional information

This section contains some useful information that you may need during the reproduction of the experiments.

Specifying the CUDA architecture

Some scripts will require to specify the CUDA architecture (or Compute Capability), this table provides a reference of the format and code to be used to specify the CUDA architecture.

Fermi	Kepler	Maxwell	Pascal	Volta	Turing	Ampere	Ada (Lovelace)	Hopper
sm_20	sm_30	sm_50	sm_60	sm_70	sm_75	sm_80	sm_89	sm_90
	sm_35	sm_52	sm_61	sm_72 (Xavier)		sm_86		sm_90a (Thor)

Obtain the GPU frequencies

When running the micro-benchmarks to generate the training datasets, the scripts will test the available frequencies of the GPU. You can reduce the number of tested frequencies through the --freq_sampling command-line argument, that allows sampling the frequencies.

If you do not know how many core frequencies your GPU has, you can run the following command.

# All frequencies
nvidia-smi -i 0 --query-supported-clocks=gr --format=csv

# Number of frequencies
nvidia-smi -i 0 --query-supported-clocks=gr --format=csv,noheader | wc -l

If your GPU has a lot of frequencies, then it may be a good idea to sample some frequencies to reduce the execution time (this may change the models' accuracy).

--gres:nvgpufreq for SLURM batch jobs

In order to run the SLURM jobs with the NVGPUFREQ plugin, the --gres:nvgpufreq and --exclusive options must be specified in the batch job.

The provided scripts already specify these options.

Compilation problems with SYCL

Sometimes when compiling SYCL programs, if more than one gcc version is installed on the system, the SYCL compiler may have troubles finding the correct gcc toolchain that must be used. In these cases, the --cxx_flags command-line argument can be used to give more information to the compiler about the location of the correct gcc toolchain. Specifying --gcc-toolchain=<gcc_toolchain_path> (LLVM < 16) or --gcc-install-dir=<gcc_install_path> (LLVM >= 16) in the --cxx_flags will allow the compiler to locate the correct toolchain version.