Package org.apache.wayang.profiler.spark

Class SparkOperatorProfiler

java.lang.Object

org.apache.wayang.profiler.spark.SparkOperatorProfiler

Direct Known Subclasses:

BinaryOperatorProfiler, SinkProfiler, SparkSourceProfiler, SparkUnaryOperatorProfiler

public abstract class SparkOperatorProfiler
extends Object

Allows to instrument an SparkExecutionOperator.

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static class	SparkOperatorProfiler.Result	The result of a single profiling run.

Field Summary

Fields

Modifier and Type	Field	Description
int	cpuMhz
protected final List<Supplier<?>>	dataQuantumGenerators
protected final long	executionPaddingTime
protected final FunctionCompiler	functionCompiler
protected List<Long>	inputCardinalities
protected final org.apache.logging.log4j.Logger	logger
int	numCoresPerMachine
int	numMachines
int	numPartitions
protected SparkExecutionOperator	operator
protected Supplier<SparkExecutionOperator>	operatorGenerator
protected SparkExecutor	sparkExecutor

Constructor Summary

Constructors

Constructor	Description
SparkOperatorProfiler(Supplier<SparkExecutionOperator> operatorGenerator, Configuration configuration, Supplier<?>... dataQuantumGenerators)

Method Summary

Modifier and Type	Method	Description
void	cleanUp()	Override this method to implement any clean-up logic.
protected static RddChannel.Instance	createChannelInstance(org.apache.spark.api.java.JavaRDD<?> rdd, SparkExecutor sparkExecutor)	Creates a ChannelInstance that carries the given rdd.
protected static RddChannel.Instance	createChannelInstance(SparkExecutor sparkExecutor)	Creates an empty ChannelInstance.
protected void	evaluate(SparkExecutionOperator operator, ChannelInstance[] inputs, ChannelInstance[] outputs)	Utility method to invoke SparkExecutionOperator.evaluate(ChannelInstance[], ChannelInstance[], SparkExecutor, OptimizationContext.OperatorContext).
protected abstract SparkOperatorProfiler.Result	executeOperator()	Executes the profiling task.
protected <T> org.apache.spark.api.java.JavaRDD<T>	partition(org.apache.spark.api.java.JavaRDD<T> rdd)	If a desired number of partitions for the input JavaRDDs is requested, enforce this.
void	prepare(long... inputCardinalities)	Call this method before run() to prepare the profiling run
protected abstract void	prepareInput(int inputIndex, long inputCardinality)
protected <T> org.apache.spark.api.java.JavaRDD<T>	prepareInputRdd(long cardinality, int inputIndex)	Helper method to generate data quanta and provide them as a cached JavaRDD.
protected <T> org.apache.spark.api.java.JavaRDD<T>	prepareInputRddInDriver(long cardinality, int inputIndex)	Helper method to generate data quanta and provide them as a cached JavaRDD.
protected <T> org.apache.spark.api.java.JavaRDD<T>	prepareInputRddInWorker(long cardinality, int inputIndex)	Helper method to generate data quanta and provide them as a cached JavaRDD.
protected long	provideCpuCycles(long startTime, long endTime)	Estimates the disk bytes occurred in the cluster during the given time span by waiting for Ganglia to provide the respective information in its RRD files.
protected long	provideDiskBytes(long startTime, long endTime)	Estimates the disk bytes occurred in the cluster during the given time span by waiting for Ganglia to provide the respective information in its RRD files.
protected long	provideNetworkBytes(long startTime, long endTime)	Estimates the network bytes occurred in the cluster during the given time span by waiting for Ganglia to provide the respective information in its RRD files.
SparkOperatorProfiler.Result	run()	Executes and profiles the profiling task.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Details

logger

protected final org.apache.logging.log4j.Logger logger

operatorGenerator

protected Supplier<SparkExecutionOperator> operatorGenerator

dataQuantumGenerators

protected final List<Supplier<?>> dataQuantumGenerators

cpuMhz

public int cpuMhz

numMachines

public int numMachines

numCoresPerMachine

public int numCoresPerMachine

numPartitions

public int numPartitions

executionPaddingTime

protected final long executionPaddingTime

operator

protected SparkExecutionOperator operator

sparkExecutor

protected SparkExecutor sparkExecutor

functionCompiler

protected final FunctionCompiler functionCompiler

inputCardinalities

protected List<Long> inputCardinalities

Constructor Details

SparkOperatorProfiler

public SparkOperatorProfiler(Supplier<SparkExecutionOperator> operatorGenerator,
 Configuration configuration,
 Supplier<?>... dataQuantumGenerators)

Method Details

prepare

public void prepare(long... inputCardinalities)

Call this method before run() to prepare the profiling run

Parameters:

inputCardinalities - number of input elements for each input of the profiled operator

prepareInput

protected abstract void prepareInput(int inputIndex,
 long inputCardinality)

prepareInputRdd

protected <T> org.apache.spark.api.java.JavaRDD<T> prepareInputRdd(long cardinality,
 int inputIndex)

Helper method to generate data quanta and provide them as a cached JavaRDD. Uses an implementation based on the wayang.profiler.datagen.location property.

prepareInputRddInDriver

protected <T> org.apache.spark.api.java.JavaRDD<T> prepareInputRddInDriver(long cardinality,
 int inputIndex)

Helper method to generate data quanta and provide them as a cached JavaRDD.

prepareInputRddInWorker

protected <T> org.apache.spark.api.java.JavaRDD<T> prepareInputRddInWorker(long cardinality,
 int inputIndex)

Helper method to generate data quanta and provide them as a cached JavaRDD.

partition

protected <T> org.apache.spark.api.java.JavaRDD<T> partition(org.apache.spark.api.java.JavaRDD<T> rdd)

If a desired number of partitions for the input JavaRDDs is requested, enforce this.

run

public SparkOperatorProfiler.Result run()

Executes and profiles the profiling task. Requires that this instance is prepared.

provideCpuCycles

protected long provideCpuCycles(long startTime,
 long endTime)

Estimates the disk bytes occurred in the cluster during the given time span by waiting for Ganglia to provide the respective information in its RRD files.

provideNetworkBytes

protected long provideNetworkBytes(long startTime,
 long endTime)

Estimates the network bytes occurred in the cluster during the given time span by waiting for Ganglia to provide the respective information in its RRD files.

provideDiskBytes

protected long provideDiskBytes(long startTime,
 long endTime)

Estimates the disk bytes occurred in the cluster during the given time span by waiting for Ganglia to provide the respective information in its RRD files.

executeOperator

protected abstract SparkOperatorProfiler.Result executeOperator()

Executes the profiling task. Requires that this instance is prepared.

evaluate

protected void evaluate(SparkExecutionOperator operator,
 ChannelInstance[] inputs,
 ChannelInstance[] outputs)

Utility method to invoke SparkExecutionOperator.evaluate(ChannelInstance[], ChannelInstance[], SparkExecutor, OptimizationContext.OperatorContext).

createChannelInstance

protected static RddChannel.Instance createChannelInstance(org.apache.spark.api.java.JavaRDD<?> rdd,
 SparkExecutor sparkExecutor)

Creates a ChannelInstance that carries the given rdd.

createChannelInstance

protected static RddChannel.Instance createChannelInstance(SparkExecutor sparkExecutor)

Creates an empty ChannelInstance.

cleanUp

public void cleanUp()

Override this method to implement any clean-up logic.