Skip to content
This repository has been archived by the owner on May 29, 2021. It is now read-only.

Latest commit

 

History

History

benchmark

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 
 
 

A first benchmark

In this first benchmark, the goal is to evaluate the performance of the skale data processing engine.

For that, we use an iterative machine learning application, the logistic regression.

Overview

The same algorithm is implemented both in NodeJS/Skale (skaleLR.js), and Python/Spark (sparkLR.py). Note that we do not use the spark-ml library or the skale-ml library, because our focus here is on core engine performances, and we also want this test to remain standalone and stable accross multiple versions of Skale and Spark.

We use the exact same input file for both, stored on the local file system, and run the test using the same host. We randomly generated data using gen_data.js program.

Results

Tests are performed on an AWS EC2 instance, called m4.4xlarge, with 16 cores and 64 GB RAM. The spark environment is spark-1.6.1, java-1.8.0_31, Python-3.4.2 (from https://hub.docker.com/r/gettyimages/spark/) The skale environment is skale-0.4.5, nodejs-4.4.3.

The spark test is run using the following command:

$ spark-submit --executor-memory 60G sparkLR.py <file> <iteration>

The skale test is run using:

$ skale-run --worker 16 --memory 4000 -- <file> <iteration>

The first set of results is for a 1 GB input file, with 3.829 millions entries of 16 features.

iterations skale (sec) spark (sec) speedup
1 4.6 19.8 4.3
4 5 34.8 7.0
10 6.5 65.9 10.1
20 8.7 116 13.3
70 20 369 18.5
100 26 522 20.1

logreg1

The second set of results is for a 10 GB input file, with 38.29 millions entries of 16 features.

iterations skale (sec) spark (sec) speedup
1 43.9 164.9 3.8
4 48.1 455 9.5
10 58.8 1038 17.7
20 82.2 2010 24.5
50 170 4927 29
100 224 9772 43.6

logreg1

Call for contribution

It would be nice to have a spark/scala version of this benchmark which could possibly perform better than the spark/python version.