org.apache.sysds.runtime.compress.cost

Interface ICostEstimate

All Superinterfaces:

Serializable

All Known Implementing Classes:

ComputationCostEstimator, DistinctCostEstimator, MemoryCostEstimator

public interface ICostEstimate
extends Serializable

A cost estimator interface. Cost here is either memory based or computation based. It is used to estimate the cost of a specific compression layout.

Field Summary

Fields

Modifier and Type	Field and Description
static org.apache.commons.logging.Log	LOG
static long	serialVersionUID

Method Summary

Modifier and Type	Method and Description
double	getCostOfCollectionOfGroups(Collection<CompressedSizeInfoColGroup> gs) Get the computation cost of a collection of column groups information.
double	getCostOfCollectionOfGroups(Collection<CompressedSizeInfoColGroup> gs, CompressedSizeInfoColGroup g)
double	getCostOfColumnGroup(CompressedSizeInfoColGroup g) If the instruction does not care about the inter column group cost, such as in memory cost or in computation cost of right or left matrix multiplication we simply estimate the cost of individual column groups.
double	getCostOfTwoGroups(CompressedSizeInfoColGroup g1, CompressedSizeInfoColGroup g2)
double	getUncompressedCost(int nRows, int nCols, int sparsity) Calculate the cost of execution if uncompressed.
boolean	isCompareAll() Function that returns true if the cost is based an an all compare all column groups false otherwise.
boolean	shouldAnalyze(CompressedSizeInfoColGroup g1, CompressedSizeInfoColGroup g2) Decide if the column groups should be analysed, or the worst case join should be expected.
boolean	shouldTryJoin(CompressedSizeInfoColGroup g1, CompressedSizeInfoColGroup g2) Decide if the column groups should try to join, this is to filter obviously bad joins out.

Field Detail

serialVersionUID

static final long serialVersionUID

See Also:

Constant Field Values

LOG

static final org.apache.commons.logging.Log LOG

Method Detail

getUncompressedCost

double getUncompressedCost(int nRows,
                           int nCols,
                           int sparsity)

Calculate the cost of execution if uncompressed. This is used as an optimization goal to at least achieve, since if the uncompressed cost of calculation is lower than in compressed space we don't want to compress.

Parameters:

nRows - Number of rows in compression

nCols - Number of columns in compression

sparsity - Sparsity of the input

Returns:

The uncompressed cost.

getCostOfColumnGroup

double getCostOfColumnGroup(CompressedSizeInfoColGroup g)

If the instruction does not care about the inter column group cost, such as in memory cost or in computation cost of right or left matrix multiplication we simply estimate the cost of individual column groups.

Parameters:

g - Column group to estimate the cost of

Returns:

The Cost of this column group

getCostOfCollectionOfGroups

double getCostOfCollectionOfGroups(Collection<CompressedSizeInfoColGroup> gs)

Get the computation cost of a collection of column groups information. Since a some operations execution time depends on all compare all calculating the cost needs all the current groups to consider. This method therefore provide a method of comparing cost of all compare all.

Parameters:

gs - All the groups considered for this cost calculation

Returns:

The cost of the calculation

getCostOfCollectionOfGroups

double getCostOfCollectionOfGroups(Collection<CompressedSizeInfoColGroup> gs,
                                   CompressedSizeInfoColGroup g)

getCostOfTwoGroups

double getCostOfTwoGroups(CompressedSizeInfoColGroup g1,
                          CompressedSizeInfoColGroup g2)

isCompareAll

boolean isCompareAll()

Function that returns true if the cost is based an an all compare all column groups false otherwise.

Returns:

boolean

shouldTryJoin

boolean shouldTryJoin(CompressedSizeInfoColGroup g1,
                      CompressedSizeInfoColGroup g2)

Decide if the column groups should try to join, this is to filter obviously bad joins out.

Parameters:

g1 - Group 1

g2 - Group 2

Returns:

Boolean specifying if they should try to join.

shouldAnalyze

boolean shouldAnalyze(CompressedSizeInfoColGroup g1,
                      CompressedSizeInfoColGroup g2)

Decide if the column groups should be analysed, or the worst case join should be expected. This is use full if the column groups are very small and the in practice difference between joining or not is small.

Parameters:

g1 - Group 1

g2 - Group 2

Returns:

If the joining should be analyzed.