TETRIS: A Dataset for FPGA-Based Heterogeneous Multi-Core Optimization

The design of heterogeneous SoCs has shown a growing trend in the latest years, since the end of Moore's Law and Dennard Scaling made hardware acceleration the only viable solution too keep the pace with increasing performance demands in the HPC industry. However, the state of the art lacks any general methodology for accurately predicting the throughput of a hardware accelerator in a system where resources like buses and memory are contended between many computing elements. A performance prediction is mandatory for SoC optimization, because both the simulation and the FPGA implementation of complex SoCs require too much time to explore a significative part of the design space.

The main idea behind TETRIS is quite simple: unleashing the power of machine learning to predict the throughput of applications in different configurations. TETRIS is a collection of area and performance results of almost 5000 different SoC configurations, that differ for size and chosen applications. This dataset aims to help researchers develop new methodologies for effectively predicting the throughput of accelerators in environment with high traffic and resource contention, boosting the research of new optimization solutions for heterogeneous SoCs.

Repository organization

The repository is organized as follows:

• data contains all the area and performance results related to the SoCs that have been implemented and tested. This folder is further divided in UC and TC.

• models includes the three models used to validate the TETRIS dataset, namely the Gaussian Process, the K-Nearest Neighbors, and the Random Forest.

• scripts contains two Python scripts, regression_main.py and optimization_main.py, that compute the regression error, and identify the optimal configuration with an iterative procedure given a set of constraints on the throughput, respectively.

Getting started

The repository contains two Python scripts, regression_main.py and optimization_main.py, intended to showcase two exemplifying usages of the TETRIS dataset.

In order to use the scripts, the user must ensure the following Python dependencies:

pip install scikit_learn

pip install matplotlib

Both the scripts can be launched by simply specifying the JSON file corresponding to the scenario for which we would like to perform the regression or the optimization (note that the optimization script is intended to be used with exhaustive scenarios, i.e. the ones in the UC folder). For example, if we want to find the regression error for the TC1 random scenario, from the main folder we run:

python3 scripts/regression_main.py data/TC/tc1_rand.json

The scripts can be tuned to ajust the experiments as needed. Regarding the regression_main.py script, the user can change the number of tests and the explored training set sizes. For the optimization_main.py script, the user can modify the number of tests, the maximum number of iteration for the optimization process, and the throughput threshold of each application (refer to the tiles.json file to get an idea of the throughput of the various applications).

Intended use

The goal of the TETRIS dataset is to allow researchers to develop methodologies for performance prediction and optimization of complex SoCs without having to generate the necessary hardware results by themselves. In this sense, the scripts included in this repository can be viewed as templates for the use and manipulation of the dataset. A researcher wanting to design a novel methodology for performance prediction could start by modifying the source files available in the models folder, developing new approaches that adopt customized regression models and/or make use of a more extensive set of accelerators and SoC features.