edu.ucla.sspace.matrix.AtomicGrowingSparseHashMatrix Maven / Gradle / Ivy
Show all versions of sspace-wordsi Show documentation
/*
* Copyright 2010 David Jurgens
*
* This file is part of the S-Space package and is covered under the terms and
* conditions therein.
*
* The S-Space package is free software: you can redistribute it and/or modify
* it under the terms of the GNU General Public License version 2 as published
* by the Free Software Foundation and distributed hereunder to you.
*
* THIS SOFTWARE IS PROVIDED "AS IS" AND NO REPRESENTATIONS OR WARRANTIES,
* EXPRESS OR IMPLIED ARE MADE. BY WAY OF EXAMPLE, BUT NOT LIMITATION, WE MAKE
* NO REPRESENTATIONS OR WARRANTIES OF MERCHANT- ABILITY OR FITNESS FOR ANY
* PARTICULAR PURPOSE OR THAT THE USE OF THE LICENSED SOFTWARE OR DOCUMENTATION
* WILL NOT INFRINGE ANY THIRD PARTY PATENTS, COPYRIGHTS, TRADEMARKS OR OTHER
* RIGHTS.
*
* You should have received a copy of the GNU General Public License
* along with this program. If not, see
*
* This class offers different performance trade-offs than the {@link
* AtomicGrowingSparseMatrix}. Specifically, this class supports concurrent
* write access to the same row. Atomicity is localized to the specific matrix
* entry, which ensures high concurrency for workloads consisting primarily of
* {@code get}, {@code addAndGet}, {@code getAndAdd} and {@code set}. This
* performance comes at a cost for the full row- or column-related operations.
* Each time these operations are used after a modification to the matrix, a
* full {@code O(m * n)} computation must be done to determine the structure of
* the matrix. An subsequent calls until the next modification will not incur
* this penalty and will operate in {@code O(k)} where {@code k} is the number
* of non-zero entries in the row or column. It is therefore highly recommended
* that vector method be used only during periods when the matrix is not likely
* to be modified.
*
* This class also provides support for access the matrix data without ensuring
* data coherency (i.e. non-atomic) through the {@link #getRowVectorUnsafe(int)}
* and {@link #getColumnVectorUnsafe(int)} methods. These methods are designed
* to be used only in circumstances where the matrix is guaranteed to only be
* accessed by one thread. In such circumstances, the overhead of locking row
* or column data can be avoided with no side-effects. Use these methods with
* great caution.
*
* @author David Jurgens
*
* @see AtomicGrowingSparseMatrix
* @see Matrices#synchronizedMatrix(Matrix)
*/
public class AtomicGrowingSparseHashMatrix
implements AtomicMatrix, SparseMatrix {
/**
* The matrix entries that are currently being write-locked by some thread.
* The value of this map is not meaningful.
*/
private final ConcurrentMap lockedEntries;
/**
* A mapping from row, column to the value at that entry in the matrix.
*/
private final ConcurrentMap matrixEntries;
/**
* The number of rows represented in this {@code AtomicGrowingSparseMatrix}.
*/
private final AtomicInteger rows;
/**
* The number of columns represented in this {@code
* AtomicGrowingSparseMatrix}.
*/
private final AtomicInteger cols;
/**
* A counter of the number of modications to this matrix. This value is
* compared with the {@link #lastVectorCacheUpdate} to determine whether the
* vector cache is out of date.
*/
private final AtomicInteger modifications;
/**
* The value of {@link #modifications} when {@link #updateVectorCache()} was
* last called.
*/
private final AtomicInteger lastVectorCacheUpdate;
/**
* A mapping from row to the columns that contain non-zero values. This
* mapping is only valid when the vector-cache is valid.
*
* @see #updateVectorCache()
*/
private int[][] rowToColsCache;
/**
* A mapping from column to the rows that contain non-zero values. This
* mapping is only valid when the vector-cache is valid.
*
* @see #updateVectorCache()
*/
private int[][] colToRowsCache;
/**
* Create an {@code AtomicGrowingSparseMatrix} with 0 rows and 0 columns.
*/
public AtomicGrowingSparseHashMatrix() {
this.rows = new AtomicInteger(0);
this.cols = new AtomicInteger(0);
modifications = new AtomicInteger(0);
lastVectorCacheUpdate = new AtomicInteger(0);
rowToColsCache = null;
colToRowsCache = null;
// Base the concurrency of the maps on the number of avaible processors,
// which assumes that all use cases use this value as a hint for
// concurrency
int threads = Runtime.getRuntime().availableProcessors();
lockedEntries = new ConcurrentHashMap(
1000, .75f, threads * 16);
matrixEntries = new ConcurrentHashMap(
10000, 4f, threads * 16);
}
/**
* {@inheritDoc}
*/
public double add(int row, int col, double delta) {
checkIndices(row, col, true);
Entry e = new Entry(row, col);
// Spin waiting for the entry to be unlocked
while (lockedEntries.putIfAbsent(e, new Object()) != null)
;
Double val = matrixEntries.get(e);
double newVal = (val == null) ? delta : delta + val;
if (newVal != 0) {
matrixEntries.put(e, newVal);
// Only invalidate the cache if the number of rows or columns
// containing data has changed
if (val == null)
modifications.incrementAndGet();
}
else {
matrixEntries.remove(e);
modifications.incrementAndGet();
}
lockedEntries.remove(e);
return (val == null) ? 0 : val;
}
/**
* {@inheritDoc}
*/
public double addAndGet(int row, int col, double delta) {
checkIndices(row, col, true);
Entry e = new Entry(row, col);
// Spin waiting for the entry to be unlocked
while (lockedEntries.putIfAbsent(e, new Object()) != null)
;
Double val = matrixEntries.get(e);
double newVal = (val == null) ? delta : delta + val;
if (newVal != 0) {
matrixEntries.put(e, newVal);
// Only invalidate the cache if the number of rows or columns
// containing data has changed
if (val == null)
modifications.incrementAndGet();
}
else {
matrixEntries.remove(e);
modifications.incrementAndGet();
}
lockedEntries.remove(e);
return newVal;
}
/**
* Verify that the given row and column value is non-negative, and
* optionally expand the size of the matrix if the row or column are outside
* the current bounds.
*
* @param row the row index to check.
* @param the the column index to check.
* @param expand {@code true} if the current dimensions of the matrix should
* be updated if either parameter exceeds the current values
*/
private void checkIndices(int row, int col, boolean expand) {
if (row < 0 || col < 0) {
throw new ArrayIndexOutOfBoundsException();
}
if (expand) {
int r = row + 1;
int cur = 0;
while (r > (cur = rows.get()) && !rows.compareAndSet(cur, r))
;
int c = col + 1;
cur = 0;
while (c > (cur = cols.get()) && !cols.compareAndSet(cur, c))
;
}
}
/**
* {@inheritDoc}
*/
public int columns() {
return cols.get();
}
/**
* {@inheritDoc}
*/
public double get(int row, int col) {
checkIndices(row, col, false);
Double val = matrixEntries.get(new Entry(row, col));
return (val == null) ? 0 : val;
}
/**
* {@inheritDoc} The length of the returned column reflects the size of
* matrix at the time of the call, which may be different from earlier calls
* to {@link #rows()}
*/
public double[] getColumn(int column) {
return getColumnVector(column).toArray();
}
/**
* {@inheritDoc} The length of the returned row vector reflects the size of
* matrix at the time of the call, which may be different from earlier calls
* to {@link #rows()}
*/
public SparseDoubleVector getColumnVector(int column) {
return getColumnVector(column, true);
}
/**
* Provides non-atomic access to the data at the specified column, which may
* present an inconsistent view of the data if this matrix is being
* concurrently modified. This method should only be used in special cases
* where the vector is being accessed at a time when the matrix (or this
* particular column) will not be modified.
*/
public SparseDoubleVector getColumnVectorUnsafe(int column) {
return getColumnVector(column, false);
}
/**
* Returns the column vector, locking the data if {@code shouldLock} is
* {@code true}.
*/
private SparseDoubleVector getColumnVector(int column, boolean shouldLock) {
int r = rows.get();
if (shouldLock)
lockColumn(column, r);
// Ensure that the column data is up to date
while (lastVectorCacheUpdate.get() != modifications.get())
updateVectorCache();
int[] rowArr = colToRowsCache[column];
SparseDoubleVector colVec = new SparseHashDoubleVector(r);
for (int row : rowArr)
colVec.set(row, matrixEntries.get(new Entry(row, column)));
if (shouldLock)
unlockColumn(column, r);
return colVec;
}
/**
* {@inheritDoc} The length of the returned row reflects the size of matrix
* at the time of the call, which may be different from earlier calls to
* {@link #columns()}.
*/
public double[] getRow(int row) {
return getRowVector(row).toArray();
}
/**
* {@inheritDoc} The length of the returned row vector reflects the size of
* matrix at the time of the call, which may be different from earlier calls
* to {@link #columns()}.
*/
public SparseDoubleVector getRowVector(int row) {
return getRowVector(row, true);
}
/**
* Provides non-atomic access to the data at the specified row, which may
* present an inconsistent view of the data if this matrix is being
* concurrently modified. This method should only be used in special cases
* where the vector is being accessed at a time when the matrix (or this
* particular row) will not be modified.
*/
public SparseDoubleVector getRowVectorUnsafe(int row) {
return getRowVector(row, false);
}
/**
* Returns the row vector, locking the data if {@code shouldLock} is {@code
* true}.
*/
private SparseDoubleVector getRowVector(int row, boolean shouldLock) {
int c = cols.get();
if (shouldLock)
lockRow(row, c);
// Ensure that the column data is up to date
while (lastVectorCacheUpdate.get() != modifications.get())
updateVectorCache();
int[] colArr = rowToColsCache[row];
SparseDoubleVector rowVec = new SparseHashDoubleVector(c);
for (int column : colArr)
rowVec.set(column, matrixEntries.get(new Entry(row, column)));
if (shouldLock)
unlockRow(row, c);
return rowVec;
}
/**
* Locks all the column entries for this row, thereby preventing write or
* read access to the values. Note that the number of columns to lock
* must be the same value used with {@link #unlockRow(int,int)},
* otherwise the unlock may potentially unlock matrix entries associated
* with this lock call.
*
* @param row the row to lock
* @param colsToLock the number of rows to lock. This value should be the
* number of {@link #cols} at the time of the call.
*/
private void lockRow(int row, int colsToLock) {
// Put in an entry for all the row's columns
for (int col = 0; col < colsToLock; ++col) {
Entry e = new Entry(row, col);
// Spin waiting for the entry to be unlocked
while (lockedEntries.putIfAbsent(e, new Object()) != null)
;
}
}
/**
* Locks all the row entries for this column, thereby preventing write or
* read access to the values. Note that the number of rows to lock
* must be the same value used with {@link #unlockColumn(int,int)},
* otherwise the unlock may potentially unlock matrix entries associated
* with this lock call.
*
* @param col the column to lock
* @param rowsToLock the number of rows to lock. This value should be the
* number of {@link #rows} at the time of the call.
*/
private void lockColumn(int col, int rowsToLock) {
// Put in an entry for all the columns's rows
for (int row = 0; row < rowsToLock; ++row) {
Entry e = new Entry(row, col);
// Spin waiting for the entry to be unlocked
while (lockedEntries.putIfAbsent(e, new Object()) != null)
;
}
}
/**
* {@inheritDoc}
*/
public int rows() {
return rows.get();
}
/**
* {@inheritDoc}
*/
public void set(int row, int col, double val) {
checkIndices(row, col, true);
Entry e = new Entry(row, col);
// Spin waiting for the entry to be unlocked
while (lockedEntries.putIfAbsent(e, new Object()) != null)
;
boolean present = matrixEntries.containsKey(e);
if (val != 0) {
matrixEntries.put(e, val);
// Only invalidate the cache if the number of rows or columns
// containing data has changed
if (!present)
modifications.incrementAndGet();
}
else if (present) {
matrixEntries.remove(e);
modifications.incrementAndGet();
}
lockedEntries.remove(e);
}
/**
* {@inheritDoc}
*/
public void setColumn(int column, double[] values) {
setColumn(column, Vectors.asVector(values));
}
/**
* {@inheritDoc}
*/
public void setColumn(int column, DoubleVector rowValues) {
checkIndices(rowValues.length(), column, true);
int r = rows.get();
lockColumn(column, r);
boolean modified = false;
for (int row = 0; row < r; ++row) {
double val = rowValues.get(row);
Entry e = new Entry(row, column);
boolean present = matrixEntries.containsKey(e);
if (val != 0) {
matrixEntries.put(e, val);
// Only invalidate the cache if the number of rows or columns
// containing data has changed
modified = modified || !present;
}
else if (present) {
matrixEntries.remove(e);
modified = true;
}
}
if (modified)
modifications.incrementAndGet();
unlockColumn(column, r);
}
/**
* {@inheritDoc}
*/
public void setRow(int row, double[] columns) {
setRow(row, Vectors.asVector(columns));
}
/**
* {@inheritDoc}
*/
public void setRow(int row, DoubleVector colValues) {
checkIndices(row, colValues.length(), true);
int c = cols.get();
lockRow(row, c);
boolean modified = false;
for (int col = 0; col < c; ++col) {
double val = colValues.get(col);
Entry e = new Entry(row, col);
boolean present = matrixEntries.containsKey(e);
if (val != 0) {
matrixEntries.put(e, val);
// Only invalidate the cache if the number of rows or columns
// containing data has changed
modified = modified || !present;
}
else if (present) {
matrixEntries.remove(e);
modified = true;
}
}
if (modified)
modifications.incrementAndGet();
unlockRow(row, c);
}
/**
* {@inheritDoc}
*/
public double[][] toDenseArray() {
int r = rows.get();
int c = cols.get();
for (int i = 0; i < r; ++i)
lockRow(i, c);
double[][] m = new double[r][0];
for (int i = 0; i < r; ++i) {
DoubleVector row = getRowVector(i);
// Ensure that we see a consistent length for all the rows
if (row.length() != c)
row = Vectors.subview(row, 0, c);
m[i] = row.toArray();
}
for (int i = 0; i < r; ++i)
unlockRow(i, c);
return m;
}
/**
* Unlocks the column for write access based on the number of rows that were
* initially locked
*
* @param col the column to unlock
* @param rowsToUnlock the number of rows locked by the previous {@link
* #lockColumn(int, int) call}
*/
private void unlockColumn(int col, int rowsToUnlock) {
// Remove the locks on all of the columns's columns
for (int row = 0; row < rowsToUnlock; ++row)
lockedEntries.remove(new Entry(row, col));
}
/**
* Unlocks the row for write access based on the number of columns that were
* initially locked
*
* @param row the row to unlock
* @param colsToUnlock the number of cols locked by the previous {@link
* #lockRow(int, int) call}
*/
private void unlockRow(int row, int colsToUnlock) {
// Remove the locks on all of the row's columns
for (int col = 0; col < colsToUnlock; ++col)
lockedEntries.remove(new Entry(row, col));
}
/**
* Updates the {@link Vector} cache data so that any call to {@link
* #getRowVector(int)} or {@link #getColumnVector(int)} has an accurate
* mapping of that row or column's non-zero values. Should the matrix be
* modifed during this call, the cache will be repeatedly computed until it
* is up-to-date.
*/
private synchronized void updateVectorCache() {
// NOTE: this method is synchronized to prevent having mulitple threads
// potentiall recomputing the cache at the same time.
while (lastVectorCacheUpdate.get() != modifications.get()) {
lastVectorCacheUpdate.set(modifications.get());
Map> rowVals =
new HashMap>();
Map> colVals =
new HashMap>();
for (Entry e : matrixEntries.keySet()) {
int row = e.row;
int col = e.col;
List c = rowVals.get(row);
if (c == null) {
c = new ArrayList();
rowVals.put(row, c);
}
c.add(col);
List r = colVals.get(col);
if (r == null) {
r = new ArrayList();
colVals.put(col, r);
}
r.add(row);
}
// Convert the object version to integer lists to save space and
// improve access
rowToColsCache = new int[rows.get()][0];
for (Map.Entry> e : rowVals.entrySet()) {
int row = e.getKey();
List cols = e.getValue();
int[] colArr = new int[cols.size()];
for (int i = 0; i < cols.size(); ++i)
colArr[i] = cols.get(i);
rowToColsCache[row] = colArr;
// null out the list for GC to try to cancel out the overhead of
// having both the array and the list in memory at the same time
e.setValue(null);
}
rowVals = null;
colToRowsCache = new int[cols.get()][0];
for (Map.Entry> e : colVals.entrySet()) {
int col = e.getKey();
List rows = e.getValue();
int[] rowArr = new int[rows.size()];
for (int i = 0; i < rows.size(); ++i)
rowArr[i] = rows.get(i);
colToRowsCache[col] = rowArr;
// null out the list for GC to try to cancel out the overhead of
// having both the array and the list in memory at the same time
e.setValue(null);
}
}
}
/**
* A utility class for holding the row-value pair
*/
private static class Entry {
final int row;
final int col;
public Entry(int row, int col) {
this.row = row;
this.col = col;
}
public boolean equals(Object o) {
if (o instanceof Entry) {
Entry e = (Entry)o;
return e.row == row && e.col == col;
}
return false;
}
public int hashCode() {
return (row << 16) ^ col;
}
}
}