esalcort / phase-aware_pmu_forecasting

Scripts to train phase-aware models that forecast performance counters values

phase-aware_pmu_forecasting

Scripts to train phase-aware models that forecast performance counters values

Data

• We provide sample data in the Data folder, but you can collect your own PMU values
• The Data folder contains CSV files with already collected and formatted PMU data
  • In set104 we provide data collected from running SPEC CPU 2017 (with OpenMP)
  • In set106 we provide data collected from running PARSEC 3

Requirements and Installation

This code has been developed in Python 3. The package-list.txt file contatins a list of Python libraries that are necessary to run the scripts. You can use your preferred Python package manager and virtual environment. Here's an example of how to set up an environment with Anaconda:

conda create -n pmuForecasting python=3.10
conda activate pmuForecasting
conda install --file package-list.txt

User Guide

Running the main scripts should follow the format:

python <script> --dataset <setname> --benchmark <bm> <args>

Where:

• script is classify.py. (We will add support for other phase aware methods later, as indicated in our plan.)
• setname is a folder inside of Data
• bm is a CSV file containing PMU data and located inside setname.
• args can be a list of required arguments specific to the script and/or optional arguments. Use python <script> --help for more information

If you want to generate an output CSV file with the predictions, use the --predictions_csv option and the --results_folder to indicate where to store the CSV file. Note that the default folder is results, and the script assumes that the folder already exists.

Phase classification

To classify phases use the classify.py script. This script has one additional required argument: phase_count, which indicates the number of phases to be classified. The output of this script has a list per-CPU classification metrics.

We support three definitions of phases as defined in [1]: global, local, and local+shared. They can be set with the --multicore_phases input argument. We also support different classification models, including 2kmeans and fgmm, as defined in [1]. The examples below classify the data in Data/set104/644.nab_s0.csv.

• 2kmeans example that classifies data into 6 global phases, and uses a window size of 21 to define the sencond-level phases of 2kmeans.

  python classify.py --dataset set104 --benchmark nab --classifier 2kmeans --phase_count 6 --W 21 --multicore_phases global --input_counters CPI L2_RQSTS.MISS OFFCORE_REQUESTS.DEMAND_DATA_RD FP_ARITH_INST_RETIRED.SCALAR_DOUBLE BR_MISP_RETIRED.ALL_BRANCHES 

• fgmm example that classifies data into 4 phases per CPU, and uses a filter size of 21 before classifying with gaussian mixture models (GMM).

  python classify.py --dataset set104 --benchmark nab --classifier gmm --phase_count 4 --filter_size 21 --multicore_phases local --input_counters CPI L2_RQSTS.MISS OFFCORE_REQUESTS.DEMAND_DATA_RD FP_ARITH_INST_RETIRED.SCALAR_DOUBLE BR_MISP_RETIRED.ALL_BRANCHES 

Single core basic workload forecasting

The forecasting.py script supports training and evaluation of basic forecasting of single-core workloads. Multiple models and different settings are supported. See the supervised_model_args method in input_parser.py or the help option python forecasting.py --help for more details.

Examples:

• Inputs 10 timesteps and output forecast horizon 2, scaled and filter data for preprocessing. The default model is LSTM, and it will be 16 cells wide and run for 100 epuchs.

  python forecasting.py --benchmark 520_0 --dataset x86_homogeneous_3GHz/ --input_counters CPI BR_PI BR_MPI --timesteps 10 --forecast_horizon 2 --scaler minmax --filter median --neurons 16 --epochs 100 --filter_size 5

• The following example shows how to select stacked LSTM (multiple LSTM layers) and enable early stopping

  python forecasting.py --benchmark 520_0 --dataset x86_homogeneous_3GHz/ --input_counters CPI BR_PI BR_MPI --timesteps 10 --forecast_horizon 2 --scaler minmax --filter median --neurons 16 --epochs 100 --early_stopping --patience 5 --model stacked_lstm

• Example with Multi-layer perceptron (MLP)

  python forecasting.py --benchmark 520_0 --dataset x86_homogeneous_3GHz/ --input_counters CPI BR_PI BR_MPI --timesteps 10 --forecast_horizon 2 --scaler minmax --filter median --epochs 100 --model mlp --dense_hidden_layers 50 50

• Example with support vector machine (SVM)

  python forecasting.py --benchmark 520_0 --dataset x86_homogeneous_3GHz/ --input_counters CPI BR_PI BR_MPI --timesteps 10 --forecast_horizon 2 --scaler minmax --filter median --model svm

Work in progress

The code is currently being cleaned and formatted to make it easier to use by people other than the authors. We have outlined a plan with release dates in stages below. Contact Susy at [email protected] if you need a copy of the code in advance.

Feature	Estimated Release Date
Multi-core phase classification	02/10/2023
Multi-core basic forecasting (phase-unaware)	TBD
Multi-core window-based phase prediction	TBD
Multi-core phase-based forecasting	TBD
Multi-core phase-aware forecasting	TBD
Multi-core phase change prediction	TBD
Multi-core phase duration prediction	TBD
Single-core phase classification	TBD
Single-core phase prediction	TBD
Single-core phase-based forecasting	TBD

References

[1] Erika S. Alcorta and Andreas Gerstlauer, "Learning-based Phase-aware Multi-core CPU Workload Forecasting", ACM Transactions on Design Automation of Electronic Systems, 28(2):23:1–23:27, March 2023