org.apache.sysds.runtime.controlprogram.context

Class SparkExecutionContext

java.lang.Object

org.apache.sysds.runtime.controlprogram.context.ExecutionContext

org.apache.sysds.runtime.controlprogram.context.SparkExecutionContext

public class SparkExecutionContext
extends ExecutionContext

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	SparkExecutionContext.SparkClusterConfig Captures relevant spark cluster configuration properties, e.g., memory budgets and degree of parallelism.

Field Summary

Fields

Modifier and Type	Field and Description
static boolean	FAIR_SCHEDULER_MODE

Method Summary

Modifier and Type	Method and Description
void	addLineage(String varParent, String varChild, boolean broadcast)
void	addLineageBroadcast(String varParent, String varChild) Adds a child broadcast object to the lineage of a parent rdd.
void	addLineageRDD(String varParent, String varChild) Adds a child rdd object to the lineage of a parent rdd.
org.apache.spark.broadcast.Broadcast<CacheBlock>	broadcastVariable(CacheableData<CacheBlock> cd)
void	cacheMatrixObject(String var)
static void	cleanupBroadcastVariable(org.apache.spark.broadcast.Broadcast<?> bvar) This call destroys a broadcast variable at all executors and the driver.
void	cleanupCacheableData(CacheableData<?> mo)
static void	cleanupRDDVariable(org.apache.spark.api.java.JavaPairRDD<?,?> rvar) This call removes an rdd variable from executor memory and disk if required.
void	cleanupThreadLocalSchedulerPool(int pool)
void	close()
static org.apache.spark.SparkConf	createSystemDSSparkConf() Sets up a SystemDS-preferred Spark configuration based on the implicit default configuration (as passed via configurations from outside).
org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock>	getBinaryMatrixBlockRDDHandleForVariable(String varname) Spark instructions should call this for all matrix inputs except broadcast variables.
org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock>	getBinaryMatrixBlockRDDHandleForVariable(String varname, int numParts, boolean inclEmpty)
org.apache.spark.api.java.JavaPairRDD<TensorIndexes,TensorBlock>	getBinaryTensorBlockRDDHandleForVariable(String varname) Spark instructions should call this for all tensor inputs except broadcast variables.
org.apache.spark.api.java.JavaPairRDD<TensorIndexes,TensorBlock>	getBinaryTensorBlockRDDHandleForVariable(String varname, int numParts, boolean inclEmpty)
PartitionedBroadcast<FrameBlock>	getBroadcastForFrameVariable(String varname)
PartitionedBroadcast<MatrixBlock>	getBroadcastForMatrixObject(MatrixObject mo)
PartitionedBroadcast<TensorBlock>	getBroadcastForTensorObject(TensorObject to)
PartitionedBroadcast<TensorBlock>	getBroadcastForTensorVariable(String varname)
PartitionedBroadcast<MatrixBlock>	getBroadcastForVariable(String varname)
static double	getBroadcastMemoryBudget() Obtains the available memory budget for broadcast variables in bytes.
static double	getDataMemoryBudget(boolean min, boolean refresh) Obtain the available memory budget for data storage in bytes.
static int	getDefaultParallelism(boolean refresh) Obtain the default degree of parallelism (cores in the cluster).
org.apache.spark.api.java.JavaPairRDD<Long,FrameBlock>	getFrameBinaryBlockRDDHandleForVariable(String varname) Spark instructions should call this for all frame inputs except broadcast variables.
static int	getNumExecutors() Obtain the number of executors in the cluster (excluding the driver).
org.apache.spark.api.java.JavaPairRDD<?,?>	getRDDHandleForFrameObject(FrameObject fo, Types.FileFormat fmt) FIXME: currently this implementation assumes matrix representations but frame signature in order to support the old transform implementation.
org.apache.spark.api.java.JavaPairRDD<?,?>	getRDDHandleForMatrixObject(MatrixObject mo, Types.FileFormat fmt)
org.apache.spark.api.java.JavaPairRDD<?,?>	getRDDHandleForMatrixObject(MatrixObject mo, Types.FileFormat fmt, int numParts, boolean inclEmpty)
org.apache.spark.api.java.JavaPairRDD<?,?>	getRDDHandleForTensorObject(TensorObject to, Types.FileFormat fmt, int numParts, boolean inclEmpty)
org.apache.spark.api.java.JavaPairRDD<?,?>	getRDDHandleForVariable(String varname, Types.FileFormat fmt, int numParts, boolean inclEmpty)
static SparkExecutionContext.SparkClusterConfig	getSparkClusterConfig() Obtains the lazily analyzed spark cluster configuration.
org.apache.spark.api.java.JavaSparkContext	getSparkContext() Returns the used singleton spark context.
static org.apache.spark.api.java.JavaSparkContext	getSparkContextStatic()
static boolean	isLazySparkContextCreation()
static boolean	isLocalMaster()
boolean	isRDDCached(int rddID)
static boolean	isSparkContextCreated() Indicates if the spark context has been created or has been passed in from outside.
void	repartitionAndCacheMatrixObject(String var)
static void	resetSparkContextStatic()
void	setRDDHandleForVariable(String varname, org.apache.spark.api.java.JavaPairRDD<?,?> rdd) Keep the output rdd of spark rdd operations as meta data of matrix/frame objects in the symbol table.
int	setThreadLocalSchedulerPool()
static FrameBlock	toFrameBlock(org.apache.spark.api.java.JavaPairRDD<Long,FrameBlock> rdd, Types.ValueType[] schema, int rlen, int clen)
static FrameBlock	toFrameBlock(RDDObject rdd, Types.ValueType[] schema, int rlen, int clen)
static org.apache.spark.api.java.JavaPairRDD<Long,FrameBlock>	toFrameJavaPairRDD(org.apache.spark.api.java.JavaSparkContext sc, FrameBlock src)
static MatrixBlock	toMatrixBlock(org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock> rdd, int rlen, int clen, int blen, long nnz) Utility method for creating a single matrix block out of a binary block RDD.
static MatrixBlock	toMatrixBlock(org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixCell> rdd, int rlen, int clen, long nnz) Utility method for creating a single matrix block out of a binary cell RDD.
static MatrixBlock	toMatrixBlock(RDDObject rdd, int rlen, int clen, int blen, long nnz) This method is a generic abstraction for calls from the buffer pool.
static MatrixBlock	toMatrixBlock(RDDObject rdd, int rlen, int clen, long nnz)
static org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock>	toMatrixJavaPairRDD(org.apache.spark.api.java.JavaSparkContext sc, MatrixBlock src, int blen)
static org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock>	toMatrixJavaPairRDD(org.apache.spark.api.java.JavaSparkContext sc, MatrixBlock src, int blen, int numParts, boolean inclEmpty)
static PartitionedBlock<MatrixBlock>	toPartitionedMatrixBlock(org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock> rdd, int rlen, int clen, int blen, long nnz)
static TensorBlock	toTensorBlock(org.apache.spark.api.java.JavaPairRDD<TensorIndexes,TensorBlock> rdd, DataCharacteristics dc)
static org.apache.spark.api.java.JavaPairRDD<TensorIndexes,TensorBlock>	toTensorJavaPairRDD(org.apache.spark.api.java.JavaSparkContext sc, TensorBlock src, int blen)
static org.apache.spark.api.java.JavaPairRDD<TensorIndexes,TensorBlock>	toTensorJavaPairRDD(org.apache.spark.api.java.JavaSparkContext sc, TensorBlock src, int blen, int numParts, boolean inclEmpty)
static void	writeFrameRDDtoHDFS(RDDObject rdd, String path, Types.FileFormat fmt)
static long	writeMatrixRDDtoHDFS(RDDObject rdd, String path, Types.FileFormat fmt)

Methods inherited from class org.apache.sysds.runtime.controlprogram.context.ExecutionContext

allocateGPUMatrixObject, cleanupDataObject, containsVariable, containsVariable, createCacheableData, createFrameObject, createMatrixObject, getCacheableData, getCacheableData, getDataCharacteristics, getDenseMatrixOutputForGPUInstruction, getFrameInput, getFrameObject, getFrameObject, getGPUContext, getGPUContexts, getLineage, getLineageItem, getListObject, getListObject, getMatrixInput, getMatrixInputForGPUInstruction, getMatrixInputs, getMatrixInputs, getMatrixObject, getMatrixObject, getMetaData, getNumGPUContexts, getOrCreateLineageItem, getProgram, getScalarInput, getScalarInput, getScalarInputs, getSparseMatrixOutputForGPUInstruction, getTensorInput, getTensorObject, getTID, getVariable, getVariable, getVariables, getVarList, getVarListPartitioned, isAutoCreateVars, isFrameObject, isMatrixObject, pinVariables, releaseCacheableData, releaseFrameInput, releaseMatrixInput, releaseMatrixInput, releaseMatrixInputForGPUInstruction, releaseMatrixInputs, releaseMatrixInputs, releaseMatrixOutputForGPUInstruction, releaseTensorInput, releaseTensorInput, removeVariable, setAutoCreateVars, setFrameOutput, setGPUContexts, setLineage, setMatrixOutput, setMatrixOutput, setMetaData, setMetaData, setProgram, setScalarOutput, setTensorOutput, setTID, setVariable, setVariables, toString, traceLineage, unpinVariables

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, wait, wait, wait

Field Detail

FAIR_SCHEDULER_MODE

public static final boolean FAIR_SCHEDULER_MODE

See Also:

Constant Field Values

Method Detail

getSparkContext

public org.apache.spark.api.java.JavaSparkContext getSparkContext()

Returns the used singleton spark context. In case of lazy spark context creation, this methods blocks until the spark context is created.

Returns:

java spark context

getSparkContextStatic

public static org.apache.spark.api.java.JavaSparkContext getSparkContextStatic()

isSparkContextCreated

public static boolean isSparkContextCreated()

Indicates if the spark context has been created or has been passed in from outside.

Returns:

true if spark context created

resetSparkContextStatic

public static void resetSparkContextStatic()

close

public void close()

isLazySparkContextCreation

public static boolean isLazySparkContextCreation()

createSystemDSSparkConf

public static org.apache.spark.SparkConf createSystemDSSparkConf()

Sets up a SystemDS-preferred Spark configuration based on the implicit default configuration (as passed via configurations from outside).

Returns:

spark configuration

isLocalMaster

public static boolean isLocalMaster()

getBinaryMatrixBlockRDDHandleForVariable

public org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock> getBinaryMatrixBlockRDDHandleForVariable(String varname)

Spark instructions should call this for all matrix inputs except broadcast variables.

Parameters:

varname - variable name

Returns:

JavaPairRDD of MatrixIndexes-MatrixBlocks

getBinaryMatrixBlockRDDHandleForVariable

public org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock> getBinaryMatrixBlockRDDHandleForVariable(String varname,
                                                                                                                 int numParts,
                                                                                                                 boolean inclEmpty)

getBinaryTensorBlockRDDHandleForVariable

public org.apache.spark.api.java.JavaPairRDD<TensorIndexes,TensorBlock> getBinaryTensorBlockRDDHandleForVariable(String varname)

Spark instructions should call this for all tensor inputs except broadcast variables.

Parameters:

varname - variable name

Returns:

JavaPairRDD of TensorIndexes-HomogTensors

getBinaryTensorBlockRDDHandleForVariable

public org.apache.spark.api.java.JavaPairRDD<TensorIndexes,TensorBlock> getBinaryTensorBlockRDDHandleForVariable(String varname,
                                                                                                                 int numParts,
                                                                                                                 boolean inclEmpty)

getFrameBinaryBlockRDDHandleForVariable

public org.apache.spark.api.java.JavaPairRDD<Long,FrameBlock> getFrameBinaryBlockRDDHandleForVariable(String varname)

Spark instructions should call this for all frame inputs except broadcast variables.

Parameters:

varname - variable name

Returns:

JavaPairRDD of Longs-FrameBlocks

getRDDHandleForVariable

public org.apache.spark.api.java.JavaPairRDD<?,?> getRDDHandleForVariable(String varname,
                                                                          Types.FileFormat fmt,
                                                                          int numParts,
                                                                          boolean inclEmpty)

getRDDHandleForMatrixObject

public org.apache.spark.api.java.JavaPairRDD<?,?> getRDDHandleForMatrixObject(MatrixObject mo,
                                                                              Types.FileFormat fmt)

getRDDHandleForMatrixObject

public org.apache.spark.api.java.JavaPairRDD<?,?> getRDDHandleForMatrixObject(MatrixObject mo,
                                                                              Types.FileFormat fmt,
                                                                              int numParts,
                                                                              boolean inclEmpty)

getRDDHandleForTensorObject

public org.apache.spark.api.java.JavaPairRDD<?,?> getRDDHandleForTensorObject(TensorObject to,
                                                                              Types.FileFormat fmt,
                                                                              int numParts,
                                                                              boolean inclEmpty)

getRDDHandleForFrameObject

public org.apache.spark.api.java.JavaPairRDD<?,?> getRDDHandleForFrameObject(FrameObject fo,
                                                                             Types.FileFormat fmt)

FIXME: currently this implementation assumes matrix representations but frame signature in order to support the old transform implementation.

Parameters:

fo - frame object

fmt - file format type

Returns:

JavaPairRDD handle for a frame object

broadcastVariable

public org.apache.spark.broadcast.Broadcast<CacheBlock> broadcastVariable(CacheableData<CacheBlock> cd)

getBroadcastForMatrixObject

public PartitionedBroadcast<MatrixBlock> getBroadcastForMatrixObject(MatrixObject mo)

getBroadcastForTensorObject

public PartitionedBroadcast<TensorBlock> getBroadcastForTensorObject(TensorObject to)

getBroadcastForVariable

public PartitionedBroadcast<MatrixBlock> getBroadcastForVariable(String varname)

getBroadcastForTensorVariable

public PartitionedBroadcast<TensorBlock> getBroadcastForTensorVariable(String varname)

getBroadcastForFrameVariable

public PartitionedBroadcast<FrameBlock> getBroadcastForFrameVariable(String varname)

setRDDHandleForVariable

public void setRDDHandleForVariable(String varname,
                                    org.apache.spark.api.java.JavaPairRDD<?,?> rdd)

Keep the output rdd of spark rdd operations as meta data of matrix/frame objects in the symbol table.

Parameters:

varname - variable name

rdd - JavaPairRDD handle for variable

toMatrixJavaPairRDD

public static org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock> toMatrixJavaPairRDD(org.apache.spark.api.java.JavaSparkContext sc,
                                                                                                   MatrixBlock src,
                                                                                                   int blen)

toMatrixJavaPairRDD

public static org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock> toMatrixJavaPairRDD(org.apache.spark.api.java.JavaSparkContext sc,
                                                                                                   MatrixBlock src,
                                                                                                   int blen,
                                                                                                   int numParts,
                                                                                                   boolean inclEmpty)

toTensorJavaPairRDD

public static org.apache.spark.api.java.JavaPairRDD<TensorIndexes,TensorBlock> toTensorJavaPairRDD(org.apache.spark.api.java.JavaSparkContext sc,
                                                                                                   TensorBlock src,
                                                                                                   int blen)

toTensorJavaPairRDD

public static org.apache.spark.api.java.JavaPairRDD<TensorIndexes,TensorBlock> toTensorJavaPairRDD(org.apache.spark.api.java.JavaSparkContext sc,
                                                                                                   TensorBlock src,
                                                                                                   int blen,
                                                                                                   int numParts,
                                                                                                   boolean inclEmpty)

toFrameJavaPairRDD

public static org.apache.spark.api.java.JavaPairRDD<Long,FrameBlock> toFrameJavaPairRDD(org.apache.spark.api.java.JavaSparkContext sc,
                                                                                        FrameBlock src)

toMatrixBlock

public static MatrixBlock toMatrixBlock(RDDObject rdd,
                                        int rlen,
                                        int clen,
                                        int blen,
                                        long nnz)

This method is a generic abstraction for calls from the buffer pool.

Parameters:

rdd - rdd object

rlen - number of rows

clen - number of columns

blen - block length

nnz - number of non-zeros

Returns:

matrix block

toMatrixBlock

public static MatrixBlock toMatrixBlock(org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock> rdd,
                                        int rlen,
                                        int clen,
                                        int blen,
                                        long nnz)

Utility method for creating a single matrix block out of a binary block RDD. Note that this collect call might trigger execution of any pending transformations. NOTE: This is an unguarded utility function, which requires memory for both the output matrix and its collected, blocked representation.

Parameters:

rdd - JavaPairRDD for matrix block

rlen - number of rows

clen - number of columns

blen - block length

nnz - number of non-zeros

Returns:

matrix block

toMatrixBlock

public static MatrixBlock toMatrixBlock(RDDObject rdd,
                                        int rlen,
                                        int clen,
                                        long nnz)

toMatrixBlock

public static MatrixBlock toMatrixBlock(org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixCell> rdd,
                                        int rlen,
                                        int clen,
                                        long nnz)

Utility method for creating a single matrix block out of a binary cell RDD. Note that this collect call might trigger execution of any pending transformations.

Parameters:

rdd - JavaPairRDD for matrix block

rlen - number of rows

clen - number of columns

nnz - number of non-zeros

Returns:

matrix block

toTensorBlock

public static TensorBlock toTensorBlock(org.apache.spark.api.java.JavaPairRDD<TensorIndexes,TensorBlock> rdd,
                                        DataCharacteristics dc)

toPartitionedMatrixBlock

public static PartitionedBlock<MatrixBlock> toPartitionedMatrixBlock(org.apache.spark.api.java.JavaPairRDD<MatrixIndexes,MatrixBlock> rdd,
                                                                     int rlen,
                                                                     int clen,
                                                                     int blen,
                                                                     long nnz)

toFrameBlock

public static FrameBlock toFrameBlock(RDDObject rdd,
                                      Types.ValueType[] schema,
                                      int rlen,
                                      int clen)

toFrameBlock

public static FrameBlock toFrameBlock(org.apache.spark.api.java.JavaPairRDD<Long,FrameBlock> rdd,
                                      Types.ValueType[] schema,
                                      int rlen,
                                      int clen)

writeMatrixRDDtoHDFS

public static long writeMatrixRDDtoHDFS(RDDObject rdd,
                                        String path,
                                        Types.FileFormat fmt)

writeFrameRDDtoHDFS

public static void writeFrameRDDtoHDFS(RDDObject rdd,
                                       String path,
                                       Types.FileFormat fmt)

addLineageRDD

public void addLineageRDD(String varParent,
                          String varChild)

Adds a child rdd object to the lineage of a parent rdd.

Parameters:

varParent - parent variable

varChild - child variable

addLineageBroadcast

public void addLineageBroadcast(String varParent,
                                String varChild)

Adds a child broadcast object to the lineage of a parent rdd.

Parameters:

varParent - parent variable

varChild - child variable

addLineage

public void addLineage(String varParent,
                       String varChild,
                       boolean broadcast)

cleanupCacheableData

public void cleanupCacheableData(CacheableData<?> mo)

Overrides:

cleanupCacheableData in class ExecutionContext

cleanupBroadcastVariable

public static void cleanupBroadcastVariable(org.apache.spark.broadcast.Broadcast<?> bvar)

This call destroys a broadcast variable at all executors and the driver. Hence, it is intended to be used on rmvar only. Depending on the ASYNCHRONOUS_VAR_DESTROY configuration, this is asynchronous or not.

Parameters:

bvar - broadcast variable

cleanupRDDVariable

public static void cleanupRDDVariable(org.apache.spark.api.java.JavaPairRDD<?,?> rvar)

This call removes an rdd variable from executor memory and disk if required. Hence, it is intended to be used on rmvar only. Depending on the ASYNCHRONOUS_VAR_DESTROY configuration, this is asynchronous or not.

Parameters:

rvar - rdd variable to remove

repartitionAndCacheMatrixObject

public void repartitionAndCacheMatrixObject(String var)

cacheMatrixObject

public void cacheMatrixObject(String var)

setThreadLocalSchedulerPool

public int setThreadLocalSchedulerPool()

cleanupThreadLocalSchedulerPool

public void cleanupThreadLocalSchedulerPool(int pool)

isRDDCached

public boolean isRDDCached(int rddID)

getSparkClusterConfig

public static SparkExecutionContext.SparkClusterConfig getSparkClusterConfig()

Obtains the lazily analyzed spark cluster configuration.

Returns:

spark cluster configuration

getBroadcastMemoryBudget

public static double getBroadcastMemoryBudget()

Obtains the available memory budget for broadcast variables in bytes.

Returns:

broadcast memory budget

getDataMemoryBudget

public static double getDataMemoryBudget(boolean min,
                                         boolean refresh)

Obtain the available memory budget for data storage in bytes.

Parameters:

min - flag for minimum data budget

refresh - flag for refresh with spark context

Returns:

data memory budget

getNumExecutors

public static int getNumExecutors()

Obtain the number of executors in the cluster (excluding the driver).

Returns:

number of executors

getDefaultParallelism

public static int getDefaultParallelism(boolean refresh)

Obtain the default degree of parallelism (cores in the cluster).

Parameters:

refresh - flag for refresh with spark context

Returns:

default degree of parallelism