FastR+Apache Flink

FastR + Apache FlinkEnabling distributed data

processing in R

<juan.fumero@ed.ac.uk>

Disclaimer

This is a work in progress

Outline

1. Introduction and motivation○ Flink and R○ Oracle Truffle + Graal

2. FastR and Flink○ Execution Model○ Memory Management

3. Preliminary Results4. Conclusions and Future Work5. [DEMO] within a distributed configuration

Introduction and motivation

Why R?

http://spectrum.ieee.org

R is one of the most popular languages for data analytics

FunctionalLazyObject OrientedDynamic

R is:

But, R is slow

Benchmark

Slo

wdo

wn

com

pare

d to

C

Problem

It is neither fast nor distributed

Problem

It is neither fast nor distributed

Apache Flink● Framework for distributed stream and batch data processing ● Custom scheduler● Custom optimizer for operations● Hadoop/Amazon plugins ● YARN connector for cluster execution● It runs on laptops, clusters and supercomputers with the same code● High level APIs: Java, Scala and Python

Flink Example, Word Countpublic class LineSplit … { public void flatMap(...) { for (String token : value.split("\\W+")) if (token.length() > 0) { out.collect(new Tuple2<String, Integer>(token, 1); }}… DataSet<String> textInput = env.readTextFile(input); DataSet<Tuple2<String, Integer>> counts= textInput.flatMap(new LineSplit())

.groupBy(0)

.sum(1);

JobManager

TaskManager

TaskManager

Flink Client And Optimizer

We propose a compilation approach built within Truffle/Graal for distributed computing on top of Apache Flink.

Our SolutionHow can we enable distributed data processing in R exploiting Flink?

FastR on Top of Flink

* Flink Material

FastR

Truffle/Graal Overview

How to Implement Your Own Language?

Oracle Labs material

Truffle/Graal Infrastructure

Oracle Labs material

Truffle node specialization

Oracle Labs material

Deoptimization and rewriting

Oracle Labs material

More Details about Truffle/Graalhttps://wiki.openjdk.java.net/display/Graal/Publications+and+Presentations

[Thomas Würthinger, Christian Wimmer, Andreas Wöß, Lukas Stadler, Gilles Duboscq, Christian Humer, Gregor Richards, Doug Simon, Mario Wolczko]

One VM to Rule Them All

In Proceedings of Onward!, 2013.

OpenJDK Graal - New JIT for Java ● Graal Compiler is a JIT compiler for Java which is written in Java● The Graal VM is a modification of the HotSpot VM -> It replaces the client and server compilers with

the Graal compiler.● Open Source: http://openjdk.java.net/projects/graal/

FastR + Flink: implementation

Our Solution: FastR - Flink Compiler● Our goal:

○ Run R data processing applications in a distributed system as easy as possible

● Approach○ Custom FastR-Flink compiler builtins (library) ○ Minimal R/Flink setup and configuration ○ “Offload” hard computation if the cluster is available

FastR-Flink API Example - Word Count # Word Count in R

bigText <- flink.readTextFile("hdfs://192.168.1.10/user/juan/quijote.txt")

createTuples <- function(text) {

words <- strsplit(text, " ")[[1]]

tuples = list()

for (w in words) {

tuples[[length(tuples)+1]] = list(w, 1)

}

return (tuples)

}

splitText <- flink.flatMap(bigText, createTuples)

groupBy <- flink.groupBy(splitText, 0)

count <- flink.sum(groupBy, 1)

hdfsPath <- "hdfs://192.168.1.10/user/juan/newQUIJOTE.txt"

flink.writeAsText(count, hdfsPath)

Supported Operations Operation Description

flink.sapply Local and remove apply (blocking)

flink.execute Execute previous operation (blocking)

flink.collect Execute the operations (blocking)

flink.map Apply local/remove map (non-blocking)

flink.sum Sum arrays (non-blocking)

flink.groupBy Group tuples (non-blocking)

Supported Operations Operation Description

flink.reduce Reduction operation (blocking and non-blocking versions)

flink.readTextFile Read from HDFS (non-blocking)

flink.filter Filter data according to the function (blocking and non-blocking)

flink.arrayMap Map with bigger chunks per Flink thread

flink.connectToJobManager Establish the connection with the main server

flink.setParallelism Set the parallelism degree for future Flink operations

Assumptions ● Distributed code with no side effects allowed → better for parallelism ● R is mostly a functional programming language

alpha <- 0.01652g <- function(x) { …

if (x < 10) {alpha <- 123.46543

} … }

Restrictions● Apply R function which change the data type

f: Integer → (String | Double)

f <- function(x) {if (x < 10) {

return("YES")}else {

return(0.123)}

}

> f(2)[1] “YES” # String> f(100)[1] 0.123 # Double

In FastR + Flink we infered the data type based on the first execution.

We can not change the type with Flink

FastR + Flink : Execution model

Cool, so how does it work?

Where R you?

ipJobManager <- "192.168.1.10"

flink.connectToJobManager(ipJobManager)

flink.setParallelism(2048)

userFunction <- function(x) {

x * x

}

result <- flink.sapply(1:10000, userFunction)

VMs cluster organization

FastR+ GraalVM

TaskManager

TaskManager

FastR + GraalVM

FastR + GraalVM

R Source Code

R Source Code

JobManager

FastR-Flink Execution Workflow

1a) Identify the skeleton

map < − f l i n k . sapply ( input , f )

switch (builtin) { … case "flink.sapply": return new FlinkFastRMap(); … }

R User Program:

FastR-Flink Compiler:

Parallel Map Node

1b ) Create the JAR File

public static ArrayList<String> compileDefaultUserFunctions() {

for (String javaSource : ClassesToCompile) { JavaCompileUtils.compileJavaClass(...) ; String jarFileName = ... String jarPath = FastRFlinkUtils.buildTmpPathForJar(jarFileName); jarFiles.add(jarPath); JavaCompileUtils.createJarFile(...); } jarFiles.add("fastr.jar"); return jarFiles; }

Jar compilation for User Defined Function:

2-3. Data Type Inference and Flink Objects

private static <T> MapOperator<?, ?> remoteApply(Object input, RFunction function, RAbstractVector[] args, String[] scopeVars, String returns) {

Object[] argsPackage = FastRFlinkUtils.getArgsPackage(nArgs, function, (RAbstractVector) input, args, argsName);firstOutput = function.getTarget().call(argsPackage);

returnsType = inferOutput(firstOutput);ArrayList<T> arrayInput = prepareArrayList(returnsType, input, args);

MapOperator<?, ?> mapOperator = computation(arrayInput, codeSource, args, argsName, returnsType, scopeVars, frame);

MetadataPipeline mapPending = new MetadataPipeline(input, firstOutput, returns);mapPending.setIntermediateReturnType(returnsType);FlinkPromises.insertOperation(Operation.REMOTE_MAP, mapPending);return mapOperator;

}

4. Code Distribution

ArrayList<String> compileUDF = FlinkUDFCompilation.compileDefaultUserFunctions();

RemoteEnvironment fastREnv = new RemoteEnvironment();fastREnv.createRemoteEnvironment(FlinkJobManager.jobManagerServer, FLINK_PORT, jarFiles);

Send the JAR files:

The Flink UDF Class receives also the String which represents the R User code:

inputSet.map(new FastRFlinkMap(codeSource)).setParallelism(parallelism)).returns(returns);

5. Data Transformation

RVector Flink DataSets

Flink DataSets preparation

RVector in FastR are not serializable → Boxing and Unboxing are needed

Paper: “Runtime Code Generation and Data Management for Heterogeneous Computing in Java”

Creating a view for the data:

6. FastR-Flink Execution, JIT Compilation

@Override public R map(T t) throws Exception {

if (!setup) { setup(); }

Object[] argsFunction = composeArgsForRFunction(t); Object[] argumentsPackage = RArguments.create(userFunction, null, null, 0, argsFunction, rArguments, null); Object result = target.call(argumentsPackage); return (R) new Tuple2<>(t.getField(0), result);}

6. FastR-Flink Execution, JIT Compilation ● Each Flink execution thread has a copy of

the R code● It builds the AST and start the

interpretation using Truffle● After a while it will reach the compiled

code (JIT) ● The language context (RContext) is

shared among the threads● Scope variables are passed by broadcast

7. Data Transformation - Get the result

RVectorFlink DataSets

Build RVectors:

The R user gets the final results within the right R data representation

Flink Web Interface ● It works with FastR!!!

● Track Jobs● Check history● Check configuration

But, Java trace.

● Maybe not very useful for R users.

Memory Management

And Memory Management on the TaskManager?

70% of the heap for hashing by default for Apache Flink

Solution: taskmanager.memory.fraction:0.5, in the config file.

User Space: ● Graal Compilation Threads● Truffle Compilation Threads● R Interpreter● R memory space

Some Questions ● Is there any heuristic to know the ideal heap distribution? ● Is there any way to format the name of the operations for the Flink Web

Interface? R functions names? Custom names for easy track● Is there any way to attach custom initialization code into TaskManager?

Preliminary Results

Evaluation Setup● Shared memory:

○ Big server machine ■ Xeon(R) 32 real cores■ Heap: 12GB■ FastR compiled with OpenJDK 1.8_60

● 4 Benchmarks:○ PI approximation○ Binary Trees○ Binomial○ Nested-Loops (high computation)

Preliminary Results

Preliminary Results

And Distributed Memory?● It works (I will show you in demo in a bit) ● Some performance issues concerning:

○ Multi-thread context issues. ○ Each Truffle language has a context initialized within the application○ It will need pre-initialization within the Flink TaskManager

● This is work in progress

Conclusions and Future work

Conclusions

● Prototype: R implementation with + some Flink operations included in the compiler

● FastR-Flink in the compiler● Easy installation and easy programmability● Good speedups in shared memory for high computation● Investigating distributed memory to increase speedups

Future Work● A world to explore:

○ Solve distribute performance issues○ More realistic benchmarks (big data)○ Distributed benchmarks○ Support more Flink operations ○ Support Flink for other Truffle languages using same approach○ Define a stable API for high level programming languages○ Optimisations (caching) ○ GPU support based on our previous work [1]

[1] Juan Fumero, Toomas Remmelg, Michel Steuwer and Christophe Dubach. Runtime Code Generation and Data Management for Heterogeneous Computing in Java. 2015 International Conference on Principles and Practices of Programming on the Java Platform

Check it out! It is Open SourceClone and compile:

$ mx sclone https://bitbucket.org/allr/fastr-flink

$ hg update -r rflink-0.1

$ mx build

Run:

$ mx R # start the interpreter with Flink local environment

Thanks for your attention

> flink.ask(questions)

AND DEMO !

Contact: Juan Fumero <juan.fumero@ed.ac.uk>

Thanks to Oracle Labs, TU Berlin and The University of Edinburgh