search

agusnt / Berti-Artifact

An artifact for Berti: an Accurate and Timely Local-Delta Data Prefetcher
GNU General Public License v3.0
22 stars 13 forks source link

readme

Berti-Artifact

DOI

Artifact used to evaluate the Berti prefetcher presented in the following paper:

Agustín Navarro-Torres, Biswabandan Panda, Jesús Alastruey-Benedé, Pablo Ibáñez Víctor Viñals-Yúfera, Alberto Ros "Berti: an Accurate and Timely Local-Delta Data Prefetcher". To appear in 55th Proceedings of the International Symposium on Microarchitecture (MICRO-55), October 2022.

Overview

This repository provides an artifact to reproduce the SPEC CPU2017 single-thread experiments. In particular it reproduce the fig 8, 9a and 10 from Berti: an Accurate and Timely Local-Delta Data Prefetcher

A full run of the artifact using the parallel flag takes about 1 hour 40 minutes in an AMD Rome 7702P with 128 threads.

Tested Environment & Dependencies

Python Packages

Python3 packages needed to generate results and graphs:

Directory Structure

This repository is organized as follow:

How to execute the artifact?

The full artifact can be executed typing:

./run.sh

The available options can be consulted with the following command:

./run.sh -h

Sequential vs Parallel execution

To speedup the execution of the artifact we provide an optional flag that runs the simulations in parallel: ./run.sh -p [number of thread].

Native Run

By default our artifact uses the GCC compiler installed on the system. However, GCC higher than version 7.5.0 can generate issues when running our ChampSim version. In order to facilitate the use of our artifact run.sh has two optional options: (1) ./run.sh -g compiles GCC 7.5.0 from scratch, and (2) run.sh -d which uses Docker to compile the artifact.

Expected output

After the execution of the artifact an output like this is expected:

./run.sh -d -p 127                                                                                                        

Building with Docker
Running in Parallel

Download SPEC CPU2017 traces [44/44]
Building Berti... done
Building MLOP... done
Building IPCP... done
Building IP Stride... done
Making everything ready to run... done
Running... done

Results, it requires numpy, scipy, maptlotlib and pprint

Parsing data... done
SPEC CPU2017 Memory Intensive SpeedUp
--------------------------------------
| Prefetch | Speedup | L1D Accuracy |
| IPCP     | 09%     | 64.9%        |
| MLOP     | 08%     | 68.0%        |
| Berti    | 12%     | 88.0%        |
--------------------------------------
Generating Figure 8 SPEC_CPU2017-MemInt... done
Generating Figure 9 (a)... done
Generating Figure 10 SPEC_CPU2017-MemInt... done
Removing Temporal Files... done

Parameters

Options: 
 -h: help
 -v: verbose mode
 -p [num]: run using [num] threads
 -g: build GCC7.5 from scratch
 -d: compile with docker
 -c: clean all generated files (traces and gcc7.5)
 -l: generate a log for debug purpose
 -r: always download SPEC CPU2K17 traces
 -n: no build the simulator
 -f: execute GAP, CloudSuite, and Multi-Level prefetcher
 -m: execute 4-Core

Full Simulation

We have included two extra options that increase the execution time of the artifact, but in return, it reproduce the results for GAP, CloudSuite, Multi-Level prefetchers and MultiCore.

To reproduce the results first of all you have to download all the traces. Since CloudSuite and GAP traces can not be downloaded automatic, so you need to download GAP traces and copy them into traces/gap folder. And download CloudSuite traces and copy them into traces/cs.

After that you have to execute ./run.sh -f -m to run the full abstract. New pdfs will appear, the name of the pdf corresponds to the figure reproduced in the article.

NOTE: The full simulation options (-f -m) takes several days to finish even in a high performance system.