zaouk / udao

Code and datasets for the paper "Neural-based Modeling for Performance Tuning of Spark Data Analytics"

This repository contains the code as well as the datasets for the paper: "Neural-based Modeling for Performance Tuning of Spark Data Analytics" authored by Khaled Zaouk, Fei Song, Chenghao Lyu and Yanlei Diao.

Requirements

Requirements for running the code in this repository are provided within the file requirements.txt. We have used the code with Python 3.7.6

Apart from that, you just need to add the path of this directory to your PYTHONPATH environment variable, so that calls to imports work fine. (Imports that look like from sparkmodeling.common.etc... won't work until you add the path of this directory to PYTHONPATH)

Datasets

The locations of the workload traces datasets are:

• (1) Spark Streaming trace dataset: datasets/streaming
• (2) TPCx-BB over Spark trace dataset: datasets/tpcx-bb

Each folder contains multiple csv files. Each csv file corresponds to traces for a particular workload. Since the number of files are limited to 1000 per folder on github (if checking from browser), we have made the datasets downloadable at this anonymous address as well: https://drive.google.com/file/d/1k3klCqOWTlijn9wT0--NZ09fOWVTA1RM/view?usp=sharing The compact URL of the dataset is: https://tinyurl.com/yc9z3lfz (MD5 checksum of the zip file: ee9012770882d300b319a174900c0196)

This is because for TPCx-bb dataset we have 1160 workloads, and thus 1160 files.

Preprocessing

Running the code under preprocessing/streaming/make_and_run_all_tests.py (for streaming dataset) and under preprocessing/tpcx-bb/make_and_run_all_tests.py (for tpcx-bb dataset) will produce the datastructures that are under: preprocessing/streaming/output/lods_mul.bin and preprocessing/tpcx-bb/output/lods_mul.bin. These are just serializations of objects from the class sparkmodeling.common.lodatastruct which contain split definitions (which workloads go to training+validation, which other workloads are left as test workloads, which configurations are observed for test workloads and which other configurations are used for evaluation, etc...)

There are no need to run the preprocessing code, as we provided the serialized objects within this repository so that you get exactly the same split definitions.

Training and Evaluation

The training and evaluation code is under experiments/streaming and experiments/tpcx-bb. Each of these directories contain a separate directory per experiment, and these subdirectories have the same order that appears in the paper.

Inside each of these subdirectories (such as experiments/streaming/4-autoencoders/3-vae/) there are usually 3 files:

• config.py: which contains the configuration required for running the training and evaluation including the best hyper-parameters found and which we used to report the results. You should be mainly interested in changing 2 things within this configuration file:
  • 1. N_OBS (put to either 1 or 5)
  • 2. ENCODING_SCHEME (put to either shared (which means configurations are used from the shared pool of traces) or all (which means configurations are selected from arbitrary pool of traces))
• train.py: this file contains instructions for training and evaluating the models by taking the configuration file into account.
• trainingutils.py: this file contains helper functions used within train.py

Running the training & evaluation code has produced the results we have put in Table 2 (within the submission file of the paper):

Hardware information

We trained our models using the provided code on a dedicated server of 20 nodes. Each node has 2x Intel(R) Xeon(R) 65 Gold 6130 CPU @ 2.10GHz processors (with 16 cores per processor) and 754 GB of RAM.

Thus, our code is optimized to run with multiple processes to make advantage of the multiple cores each node had. You may need to change the N_WORKERS within configuration files if you are using less powerful machines.