Apache SystemML (Incubating)

Clone this repo:
  1. 11a8577 [SYSTEMML-833] Additional cleanup rewrites (unnecess. cast, reorg, agg) by Matthias Boehm · 10 months ago master
  2. 461184a [SYSTEMML-694] Improved binary-unary rewrites (added max(0,X)->selp(X)) by Matthias Boehm · 10 months ago
  3. 382df65 [SYSTEMML-694] Improved wdivmm rewrites (outer-product-like mm only) by Matthias Boehm · 10 months ago
  4. a528f5e [SYSTEMML-824] Performance scalar/unary/binary ops (nnz, sparse, skip) by Matthias Boehm · 10 months ago
  5. 7f33276 [HOTFIX] Update input variables to input parameters for tests on Windows by Deron Eriksson · 10 months ago


Documentation: SystemML Documentation
Mailing List: Dev Mailing List
Build Status: Build Status
Issue Tracker: JIRA
Download: Download SystemML

SystemML is now an Apache Incubator project! Please see the Apache SystemML (incubating) website for more information. The latest project documentation can be found at the SystemML Documentation website on GitHub.

SystemML is a flexible, scalable machine learning system. SystemML's distinguishing characteristics are:

  1. Algorithm customizability via R-like and Python-like languages.
  2. Multiple execution modes, including Standalone, Spark Batch, Spark MLContext, Hadoop Batch, and JMLC.
  3. Automatic optimization based on data and cluster characteristics to ensure both efficiency and scalability.

Algorithm Customizability

ML algorithms in SystemML are specified in a high-level, declarative machine learning (DML) language. Algorithms can be expressed in either an R-like syntax or a Python-like syntax. DML includes linear algebra primitives, statistical functions, and additional constructs.

This high-level language significantly increases the productivity of data scientists as it provides (1) full flexibility in expressing custom analytics and (2) data independence from the underlying input formats and physical data representations.

Multiple Execution Modes

SystemML computations can be executed in a variety of different modes. To begin with, SystemML can be operated in Standalone mode on a single machine, allowing data scientists to develop algorithms locally without need of a distributed cluster. In order to scale up, algorithms can also be distributed across a cluster using Spark or Hadoop. This flexibility allows the utilization of an organization's existing resources and expertise. In addition, SystemML features a Spark MLContext API that allows for programmatic interaction via Scala and Java. SystemML also features an embedded API for scoring models.

Automatic Optimization

Algorithms specified in DML are dynamically compiled and optimized based on data and cluster characteristics using rule-based and cost-based optimization techniques. The optimizer automatically generates hybrid runtime execution plans ranging from in-memory, single-node execution, to distributed computations on Spark or Hadoop. This ensures both efficiency and scalability. Automatic optimization reduces or eliminates the need to hand-tune distributed runtime execution plans and system configurations.

SystemML Assumptions

Before you get started on SystemML, make sure that your environment is set up and ready to go.

  1. If you’re on a mac, you’ll want to install homebrew (http://brew.sh) if you haven’t already.

Copy and paste the following into your terminal.

/usr/bin/ruby -e "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install)"
  1. Now install Java (need Java 8).
brew tap caskroom/cask
brew install Caskroom/cask/java
  1. Now install everything else you need

In order to install something on homebrew all you need to do is type “brew install” followed by what you want to install. See below.

Install Spark 1.6

brew install apache-spark

Install python

brew install python

Install jupyter and matplotlib and numpy.

pip install jupyter matplotlib numpy
  1. Now you can install SystemML.

Go to http://systemml.apache.org/download.html and click on the systemml-0.10.0-incubating zip (should be 2nd).

The next step is optional, but it will make your life a lot easier.

  1. Set SPARK_HOME and SYSTEMML_HOME on your bash profile.

First, use vim to create/edit your bash profile. Not sure what vim is? Check https://www.linux.com/learn/vim-101-beginners-guide-vim.

We are going to insert our file where Spark and SystemML is stored into our bash profile. This will make it easier to access. Type the following to open your bash profile using vim:

vim .bash_profile

Now you are in vim. First, type “i” for insert.


Now insert Spark and SystemML. Note: /Documents is where I saved my Spark and SystemML. Be sure that your file path is accurate.

export SPARK_HOME=/Users/stc/Documents/spark-1.5.1-bin-hadoop2.6

export SYSTEMML_HOME=/Users/stc/Documents/systemml-0.10.0-incubating

Now type :wq to write the file and quit.


Make sure to open a new tab in terminal so that you make sure the changes have been made.

Congrats! You can now run SystemML!

Building SystemML

SystemML is built using Apache Maven. SystemML will build on Linux, MacOS, or Windows, and requires Maven 3 and Java 7 (or higher). To build SystemML, run:

mvn clean package

To build the SystemML distributions (.tar.gz, .zip, etc.), run:

mvn clean package -P distribution

Testing SystemML

SystemML features a comprehensive set of integration tests. To perform these tests, run:

mvn verify

Note: these tests require R to be installed and available as part of the PATH variable on the machine on which you are running these tests.

If required, please install the following packages in R:

install.packages(c("batch", "bitops", "boot", "caTools", "data.table", "doMC", "doSNOW", "ggplot2", "glmnet", "lda", "Matrix", "matrixStats", "moments", "plotrix", "psych", "reshape", "topicmodels", "wordcloud"), dependencies=TRUE)

Importing SystemML into IDE

This section describe how to import SystemML source code into an IDE.

Import SystemML Project to Eclipse

Eclipse IDE include:

File -> Import -> Maven -> Existing Maven Projects

Please see below how to resolve some compilation issues that might occour after importing the SystemML project:

invalid cross-compiled libraries error

Since Scala IDE bundles the latest versions (2.10.5 and 2.11.6 at this point), you need do add one in Eclipse Preferences -> Scala -> Installations by pointing to the lib/ directory of your Scala 2.10.4 distribution. Once this is done, select all Spark projects and right-click, choose Scala -> Set Scala Installation and point to the 2.10.4 installation. This should clear all errors about invalid cross-compiled libraries. A clean build should succeed now.

incompatation scala version error

Change IDE scala version project->propertiest->scala compiler -> scala installation to Fixed scala Installation: 2.10.5

Not found type * error

Run command mvn package, and do project -> refresh

maketplace not found error for Eclipse Luna

Except scala IDE pulgin install, please make sure get update from “http://alchim31.free.fr/m2e-scala/update-site” to update maven connector for scala.

Import SystemML project to IntelliJ

  1. Download IntelliJ and install the Scala plug-in for IntelliJ.
  2. Go to “File -> Import Project”, locate the spark source directory, and select “Maven Project”.
  3. In the Import wizard, it's fine to leave settings at their default. However it is usually useful to enable “Import Maven projects automatically”, since changes to the project structure will automatically update the IntelliJ project.

Running SystemML in Standalone Mode

SystemML can run in distributed mode as well as in local standalone mode. We'll operate in standalone mode in this guide. After you build SystemML from source (mvn clean package), the standalone mode can be executed either on Linux or OS X using the ./bin/systemml script, or on Windows using the .\bin\systemml.bat batch file.

If you run from the script from the project root folder ./ or from the ./bin folder, then the output files from running SystemML will be created inside the ./temp folder to keep them separate from the SystemML source files managed by Git. The output files for all of the examples in this guide will be created under the ./temp folder.

The runtime behavior and logging behavior of SystemML can be customized by editing the files ./conf/SystemML-config.xml and ./conf/log4j.properties. Both files will be created from their corresponding *.template files during the first execution of the SystemML executable script.

When invoking the ./bin/systemml or .\bin\systemml.bat with any of the prepackaged DML scripts you can omit the relative path to the DML script file. The following two commands are equivalent:

./bin/systemml ./scripts/datagen/genLinearRegressionData.dml -nvargs numSamples=1000 numFeatures=50 maxFeatureValue=5 maxWeight=5 addNoise=FALSE b=0 sparsity=0.7 output=linRegData.csv format=csv perc=0.5

./bin/systemml genLinearRegressionData.dml -nvargs numSamples=1000 numFeatures=50 maxFeatureValue=5 maxWeight=5 addNoise=FALSE b=0 sparsity=0.7 output=linRegData.csv format=csv perc=0.5

In this guide we invoke the command with the relative folder to make it easier to look up the source of the DML scripts.

ML Algorithms

SystemML features a suite of algorithms that can be grouped into six broad categories: Descriptive Statistics, Classification, Clustering, Regression, Matrix Factorization, and Survival Analysis. Detailed descriptions of these algorithms can be found in the SystemML Algorithms Reference.

Linear Regression Example

As an example of the capabilities and power of SystemML and DML, let's consider the Linear Regression algorithm. We require sets of data to train and test our model. To obtain this data, we can either use real data or generate data for our algorithm. The UCI Machine Learning Repository Datasets is one location for real data. Use of real data typically involves some degree of data wrangling. In the following example, we will use SystemML to generate random data to train and test our model.

This example consists of the following parts:

SystemML is distributed in several packages, including a standalone package. We'll operate in Standalone mode in this example.

Run DML Script to Generate Random Data

We can execute the genLinearRegressionData.dml script in Standalone mode using either the systemml or systemml.bat file. In this example, we'll generate a matrix of 1000 rows of 50 columns of test data, with sparsity 0.7. In addition to this, a 51st column consisting of labels will be appended to the matrix.

./bin/systemml ./scripts/datagen/genLinearRegressionData.dml -nvargs numSamples=1000 numFeatures=50 maxFeatureValue=5 maxWeight=5 addNoise=FALSE b=0 sparsity=0.7 output=linRegData.csv format=csv perc=0.5

This generates the following files inside the ./temp folder:

linRegData.csv      # 1000 rows of 51 columns of doubles (50 data columns and 1 label column), csv format
linRegData.csv.mtd  # Metadata file
perc.csv            # Used to generate two subsets of the data (for training and testing)
perc.csv.mtd        # Metadata file
scratch_space       # SystemML scratch_space directory

Divide Generated Data into Two Sample Groups

Next, we'll create two subsets of the generated data, each of size ~50%. We can accomplish this using the sample.dml script with the perc.csv file created in the previous step:


The sample.dml script will randomly sample rows from the linRegData.csv file and place them into 2 files based on the percentages specified in perc.csv. This will create two sample groups of roughly 50 percent each.

./bin/systemml ./scripts/utils/sample.dml -nvargs X=linRegData.csv sv=perc.csv O=linRegDataParts ofmt=csv

This script creates two partitions of the original data and places them in a linRegDataParts folder. The files created are as follows:

linRegDataParts/1       # first partition of data, ~50% of rows of linRegData.csv, csv format
linRegDataParts/1.mtd   # metadata
linRegDataParts/2       # second partition of data, ~50% of rows of linRegData.csv, csv format
linRegDataParts/2.mtd   # metadata

The 1 file contains the first partition of data, and the 2 file contains the second partition of data. An associated metadata file describes the nature of each partition of data. If we open 1 and 2 and look at the number of rows, we can see that typically the partitions are not exactly 50% but instead are close to 50%. However, we find that the total number of rows in the original data file equals the sum of the number of rows in 1 and 2.

Split Label Column from First Sample

The next task is to split the label column from the first sample. We can do this using the splitXY.dml script.

./bin/systemml ./scripts/utils/splitXY.dml -nvargs X=linRegDataParts/1 y=51 OX=linRegData.train.data.csv OY=linRegData.train.labels.csv ofmt=csv

This splits column 51, the label column, off from the data. When done, the following files have been created.

linRegData.train.data.csv        # training data of 50 columns, csv format
linRegData.train.data.csv.mtd    # metadata
linRegData.train.labels.csv      # training labels of 1 column, csv format
linRegData.train.labels.csv.mtd  # metadata

Split Label Column from Second Sample

We also need to split the label column from the second sample.

./bin/systemml ./scripts/utils/splitXY.dml -nvargs X=linRegDataParts/2 y=51 OX=linRegData.test.data.csv OY=linRegData.test.labels.csv ofmt=csv

This splits column 51 off the data, resulting in the following files:

linRegData.test.data.csv        # test data of 50 columns, csv format
linRegData.test.data.csv.mtd    # metadata
linRegData.test.labels.csv      # test labels of 1 column, csv format
linRegData.test.labels.csv.mtd  # metadata

Train Model on First Sample

Now, we can train our model based on the first sample. To do this, we utilize the LinearRegDS.dml (Linear Regression Direct Solve) script. Note that SystemML also includes a LinearRegCG.dml (Linear Regression Conjugate Gradient) algorithm for situations where the number of features is large.

./bin/systemml ./scripts/algorithms/LinearRegDS.dml -nvargs X=linRegData.train.data.csv Y=linRegData.train.labels.csv B=betas.csv fmt=csv

This will generate the following files:

betas.csv      # betas, 50 rows of 1 column, csv format
betas.csv.mtd  # metadata

The LinearRegDS.dml script generates statistics to standard output similar to the following.

Reading X and Y...
Calling the Direct Solver...
Computing the statistics...
Writing the output matrix...

Now that we have our betas.csv, we can test our model with our second set of data.

Test Model on Second Sample

To test our model on the second sample, we can use the GLM-predict.dml script. This script can be used for both prediction and scoring. Here, we're using it for scoring since we include the Y named argument. Our betas.csv file is specified as the B named argument.

./bin/systemml ./scripts/algorithms/GLM-predict.dml -nvargs X=linRegData.test.data.csv Y=linRegData.test.labels.csv B=betas.csv fmt=csv

This generates statistics similar to the following to standard output.


We see that the STDEV_RES_Y value of the testing phase is of similar magnitude to the value obtained from the model training phase.

For convenience, we can encapsulate our DML invocations in a single script:


./bin/systemml ./scripts/datagen/genLinearRegressionData.dml -nvargs numSamples=1000 numFeatures=50 maxFeatureValue=5 maxWeight=5 addNoise=FALSE b=0 sparsity=0.7 output=linRegData.csv format=csv perc=0.5

./bin/systemml ./scripts/utils/sample.dml -nvargs X=linRegData.csv sv=perc.csv O=linRegDataParts ofmt=csv

./bin/systemml ./scripts/utils/splitXY.dml -nvargs X=linRegDataParts/1 y=51 OX=linRegData.train.data.csv OY=linRegData.train.labels.csv ofmt=csv

./bin/systemml ./scripts/utils/splitXY.dml -nvargs X=linRegDataParts/2 y=51 OX=linRegData.test.data.csv OY=linRegData.test.labels.csv ofmt=csv

./bin/systemml ./scripts/algorithms/LinearRegDS.dml -nvargs X=linRegData.train.data.csv Y=linRegData.train.labels.csv B=betas.csv fmt=csv

./bin/systemml ./scripts/algorithms/GLM-predict.dml -nvargs X=linRegData.test.data.csv Y=linRegData.test.labels.csv B=betas.csv fmt=csv

Conclusion and Next Steps

In this example, we've seen a small part of the capabilities of SystemML. For more detailed information, please consult the Apache SystemML (incubating) website and the SystemML Documentation website on GitHub.