Skip to content

Latest commit

 

History

History
129 lines (85 loc) · 5.18 KB

README.md

File metadata and controls

129 lines (85 loc) · 5.18 KB
Raw

Erebus

This repository contains the evaluation artifacts of the paper:

Dimitris Palyvos-Giannas, Katerina Tzompanaki, Marina Papatriantafilou, and Vincenzo Gulisano. Erebus: Explaining the Outputs of Data Streaming Queries. PVLDB, 16(2): 230 - 242, 2022. doi:10.14778/3565816.3565825

Full text

In the following, we are providing necessary setup steps as well as instructions on how to run the experiments from our paper and how to visualize them. All steps have been tested under Ubuntu 20.04.

Setup

Step 0: Dependencies

The following dependencies are not managed by our automatic setup and need to be installed beforehand.:

  • git
  • wget
  • maven
  • unzip
  • OpenJDK 1.8
  • python3 (+ pip3)

For Ubuntu 20.04: sudo apt-get install git wget maven unzip openjdk-8-jdk python3-pip python3-yaml libyaml-dev cython

Below we assume the repository has been cloned into the directory repo_dir.

Step 1: Install Python Requirements

  • Using a virtual environment is suggested.
  • Initializing the environment and running the experiments requires: gdown pyyaml tqdm
  • Recreating the figures requires additionally: pandas matplotlib seaborn numpy adjustText xlsxwriter

To install the python requirements automatically, from repo_dir, run one of the following commands:

# Full Dependencies (Running Experiments and Plotting)
pip install -r requirements.txt
# Minimal Dependencies (Running Experiments Only)
pip install -r requirements-minimal.txt

Step 2: Automated Setup

This method will automatically download Apache Flink, Apache Kafka and the input datasets, and compile the framework.

  1. From repo_dir, run ./auto-setup.sh
  2. Run ./init-configs.sh NODE_TYPE where NODE_TYPE is odroid or server to use the Kafka/Flink configurations used in our experiments

(Alternative) Manual Setup

In case of problems with the automatic setup, you can prepare the environment manually:

  1. Download Apache Flink 1.14.0 and decompress in repo_dir
  2. Download Apache Kafka 2.13-3.1.0 and decompress in repo_dir
  3. Get the input datasets, either using ./get-datasets.sh or by manually downloading from here and decompressing in repo_dir/data/input
  4. Compile the two experiment jars, from repo dir: mvn -f helper_pom.xml clean package && mv target/helper*.jar jars; mvn clean package
  5. Run ./init-configs.sh NODE_TYPE where NODE_TYPE is odroid or server to use the Kafka/Flink configurations used in our experiments

Running Experiments

Starting Kafka

In your Kafka node (referred to as kafka_node below), run from repo_dir:

cd kafka_2.13-3.1.0
bin/zookeeper-server-start.sh config/zookeeper.properties

and then, in another terminal session (same directory):

bin/kafka-server-start.sh config/server.properties

Automatic Reproduction of the Paper's Experiments and Plots

The reproduce/ directory contains a script for each evaluation figure of the paper, which will run the experiment automatically, store the results, and create a plot. You need to provide the #repetitions, duration in minutes, and the hostname/port of the kafka_node as arguments. For example, to reproduce Figure 6, run (from repo_dir):

# Reproduce Figure 6 of the paper for 3 repetitions of 10 minutes
./reproduce/figure6.sh 3 10 kafka_node:9092

The only exception is ./reproduce/figure11.sh, which takes as arguments the number of forks, warmup iterations and regular iterations of JMH:

# Reproduce Figure 11 of the paper for 3 forks 10 warmup iterations and 25 regular iterations
./reproduce/figure6.sh 3 10 25

where kafka_node is the host running Kafka (localhost if you started Kafka on the same machine).

Results are stored in the folder data/output. In case of problems with python dependencies (especially on Odroids), it is also possible to plot in a separate device, after running the experiment, by copying the result folder and calling reproduce/plot.py with the arguments of the respective reproduce script.

Some experiment scripts print debugging information that is usually safe to ignore, as long as the figures are generated successfully.

Keeping explanations

By default, explanations are deleted after the experiment finishes because they can consume a lot of disk space. If you want to keep the explanations, pass the argument --keep-outputs to the experiment script.

Appendix

Original Datasets

The original sets from our evaluation are found below:

PyYAML Issues

In case of errors related to YAML CLoader in Odroids, install python3-yaml using apt-get and do not use a python virtual environment. See here for more details: yaml/pyyaml#108 (comment)