Skip to content

H
h2database
提交 f4e7837e authored 作者: Thomas Mueller's avatar Thomas Mueller
浏览文件
操作

LIRS cache: concurrent

上级 cf6fb26c
隐藏空白字符变更
内嵌 并排
正在显示 包含 982 行增加 和 1 行删除
h2/src/main/org/h2/util/Task.java
浏览文件 @ f4e7837e
...@@ -47,7 +47,17 @@ public abstract class Task implements Runnable { ...@@ -47,7 +47,17 @@ public abstract class Task implements Runnable {
* @return this * @return this
*/ */
public Task execute() { public Task execute() {
thread = new Thread(this, getClass().getName()); return execute(getClass().getName());
}
/**
* Start the thread.
*
* @param threadName the name of the thread
* @return this
*/
public Task execute(String threadName) {
thread = new Thread(this, threadName);
thread.setDaemon(true); thread.setDaemon(true);
thread.start(); thread.start();
return this; return this;
......
h2/src/test/org/h2/test/store/TestCacheConcurrentLIRS.java 0 → 100644
浏览文件 @ f4e7837e
/*
* Copyright 2004-2011 H2 Group. Multiple-Licensed under the H2 License, Version
* 1.0, and under the Eclipse Public License, Version 1.0
* (http://h2database.com/html/license.html). Initial Developer: H2 Group
*/
package org.h2.test.store;
import java.util.Random;
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.atomic.AtomicBoolean;
import org.h2.dev.store.btree.CacheConcurrentLIRS;
import org.h2.test.TestBase;
import org.h2.util.Task;
/**
* Tests the cache algorithm.
*/
public class TestCacheConcurrentLIRS extends TestBase {
/**
* Run just this test.
*
* @param a ignored
*/
public static void main(String... a) throws Exception {
TestBase.createCaller().init().test();
}
public void test() throws Exception {
testConcurrent();
}
private static void testConcurrent() {
final CacheConcurrentLIRS<Integer, Integer> test = CacheConcurrentLIRS.newInstance(100, 1);
int threadCount = 8;
final CountDownLatch wait = new CountDownLatch(1);
final AtomicBoolean stopped = new AtomicBoolean();
Task[] tasks = new Task[threadCount];
final int[] getCounts = new int[threadCount];
for (int i = 0; i < threadCount; i++) {
final int x = i;
Task t = new Task() {
@Override
public void call() throws Exception {
Random random = new Random(x);
wait.await();
int i = 0;
for (; !stopped.get(); i++) {
int key;
do {
key = (int) Math.abs(random.nextGaussian() * 50);
} while (key > 100);
test.get(key);
if ((i & 127) == 0) {
test.put(random.nextInt(100), random.nextInt());
}
}
getCounts[x] = i;
}
};
t.execute("t" + i);
tasks[i] = t;
}
wait.countDown();
try {
Thread.sleep(2000);
} catch (InterruptedException e) {
e.printStackTrace();
}
stopped.set(true);
for (Task t : tasks) {
t.get();
}
int totalCount = 0;
for (int x : getCounts) {
totalCount += x;
}
System.out.println("requests: " + totalCount);
}
}
h2/src/tools/org/h2/dev/store/btree/CacheConcurrentLIRS.java 0 → 100644
浏览文件 @ f4e7837e
/*
* Copyright 2004-2011 H2 Group. Multiple-Licensed under the H2 License,
* Version 1.0, and under the Eclipse Public License, Version 1.0
* (http://h2database.com/html/license.html).
* Initial Developer: H2 Group
*/
package org.h2.dev.store.btree;
import java.util.AbstractMap;
import java.util.ArrayList;
import java.util.HashMap;
import java.util.HashSet;
import java.util.List;
import java.util.Map;
import java.util.Set;
/**
* A scan resistant cache. It is meant to cache objects that are relatively
* costly to acquire, for example file content.
* <p>
* This implementation is multi-threading save and supports concurrent access.
* Null keys or null values are not allowed. The map fill factor is at most 75%.
* <p>
* Each entry is assigned a distinct memory size, and the cache will try to use
* at most the specified amount of memory. The memory unit is not relevant,
* however it is suggested to use bytes as the unit.
* <p>
* This class implements an approximation of the the LIRS replacement algorithm
* invented by Xiaodong Zhang and Song Jiang as described in
* http://www.cse.ohio-state.edu/~zhang/lirs-sigmetrics-02.html with a few
* smaller changes: An additional queue for non-resident entries is used, to
* prevent unbound memory usage. The maximum size of this queue is at most the
* size of the rest of the stack. About 6.25% of the mapped entries are cold.
* <p>
* Internally, the cache is split into 16 segments, and each segment is an
* individual LIRS cache. Accessed entries are only moved to the top of the
* stack if at least 20 other entries have been moved to the front. Write access
* and moving entries to the top of the stack is synchronized per segment.
*
* @author Thomas Mueller
* @param <K> the key type
* @param <V> the value type
*/
public class CacheConcurrentLIRS<K, V> extends AbstractMap<K, V> implements Map<K, V> {
/**
* The maximum memory this cache should use.
*/
private long maxMemory;
/**
* The average memory used by one entry.
*/
private int averageMemory;
private Segment<K, V>[] segments;
private int segmentShift;
private int segmentMask;
private CacheConcurrentLIRS(long maxMemory, int averageMemory) {
this.maxMemory = maxMemory;
this.averageMemory = averageMemory;
clear();
}
@SuppressWarnings("unchecked")
public void clear() {
// must be a power of 2
int count = 16;
segmentMask = count - 1;
segments = new Segment[count];
for (int i = 0; i < count; i++) {
segments[i] = new Segment<K, V>(
1 + maxMemory / count, averageMemory);
}
segmentShift = Integer.numberOfTrailingZeros(segments[0].sizeMapArray());
}
private Entry<K, V> find(Object key) {
int hash = getHash(key);
return getSegment(hash).find(key, hash);
}
/**
* Check whether there is a resident entry for the given key. This method
* does not adjusts the internal state of the cache.
*
* @param key the key (may not be null)
* @return true if there is a resident entry
*/
public boolean containsKey(Object key) {
int hash = getHash(key);
return getSegment(hash).containsKey(key, hash);
}
/**
* Get the value for the given key if the entry is cached. This method does
* not modify the internal state.
*
* @param key the key (may not be null)
* @return the value, or null if there is no resident entry
*/
public V peek(K key) {
Entry<K, V> e = find(key);
return e == null ? null : e.value;
}
/**
* Add an entry to the cache. The entry may or may not exist in the cache
* yet. This method will usually mark unknown entries as cold and known
* entries as hot.
*
* @param key the key (may not be null)
* @param value the value (may not be null)
* @param memory the memory used for the given entry
* @return the old value, or null if there is no resident entry
*/
public V put(K key, V value, int memory) {
int hash = getHash(key);
return getSegment(hash).put(key, value, hash, memory);
}
/**
* Add an entry to the cache using the average memory size.
*
* @param key the key (may not be null)
* @param value the value (may not be null)
* @return the old value, or null if there is no resident entry
*/
public V put(K key, V value) {
return put(key, value, averageMemory);
}
/**
* Remove an entry. Both resident and non-resident entries can be removed.
*
* @param key the key (may not be null)
* @return the old value, or null if there is no resident entry
*/
public synchronized V remove(Object key) {
int hash = getHash(key);
return getSegment(hash).remove(key, hash);
}
/**
* Get the memory used for the given key.
*
* @param key the key (may not be null)
* @return the memory, or 0 if there is no resident entry
*/
public int getMemory(K key) {
int hash = getHash(key);
return getSegment(hash).getMemory(key, hash);
}
/**
* Get the value for the given key if the entry is cached. This method
* adjusts the internal state of the cache, to ensure commonly used entries
* stay in the cache.
*
* @param key the key (may not be null)
* @return the value, or null if there is no resident entry
*/
public V get(Object key) {
int hash = getHash(key);
return getSegment(hash).get(key, hash);
}
private Segment<K, V> getSegment(int hash) {
int segmentIndex = (hash >>> segmentShift) & segmentMask;
return segments[segmentIndex];
}
static int getHash(Object key) {
int hash = key.hashCode();
// Doug Lea's supplemental secondaryHash function (inlined)
// to protect against hash codes that don't differ in low order bits
hash ^= (hash >>> 20) ^ (hash >>> 12);
hash ^= (hash >>> 7) ^ (hash >>> 4);
return hash;
}
/**
* Get the currently used memory.
*
* @return the used memory
*/
public long getUsedMemory() {
long x = 0;
for (Segment<K, V> s : segments) {
x += s.getUsedMemory();
}
return x;
}
/**
* Set the maximum memory this cache should use. This will not immediately
* cause entries to get removed however; it will only change the limit. To
* resize the internal array, call the clear method.
*
* @param maxMemory the maximum size (1 or larger)
*/
public void setMaxMemory(long maxMemory) {
if (maxMemory <= 0) {
throw new IllegalArgumentException("Max memory must be larger than 0");
}
this.maxMemory = maxMemory;
for (Segment<K, V> s : segments) {
s.setMaxMemory(1 + maxMemory / segments.length);
}
}
/**
* Get the average memory used per entry.
*
* @return the average memory
*/
public int getAverageMemory() {
return averageMemory;
}
/**
* Get the maximum memory to use.
*
* @return the maximum memory
*/
public long getMaxMemory() {
return maxMemory;
}
/**
* Create a new cache with the given memory size. To just limit the number
* of entries, use the required number as the maximum memory, and an average
* size of 1.
*
* @param maxMemory the maximum memory to use (1 or larger)
* @param averageMemory the average memory (1 or larger)
* @return the cache
*/
public static <K, V> CacheConcurrentLIRS<K, V> newInstance(int maxMemory, int averageMemory) {
return new CacheConcurrentLIRS<K, V>(maxMemory, averageMemory);
}
/**
* Get the entry set for all resident entries.
*
* @return the entry set
*/
public synchronized Set<Map.Entry<K, V>> entrySet() {
HashMap<K, V> map = new HashMap<K, V>();
for (K k : keySet()) {
map.put(k, find(k).value);
}
return map.entrySet();
}
/**
* Get the set of keys for resident entries.
*
* @return the set of keys
*/
public synchronized Set<K> keySet() {
HashSet<K> set = new HashSet<K>();
for (Segment<K, V> s : segments) {
set.addAll(s.keySet());
}
return set;
}
/**
* Get the number of non-resident entries in the cache.
*
* @return the number of non-resident entries
*/
public int sizeNonResident() {
int x = 0;
for (Segment<K, V> s : segments) {
x += s.sizeNonResident();
}
return x;
}
/**
* Get the length of the internal map array.
*
* @return the size of the array
*/
public int sizeMapArray() {
int x = 0;
for (Segment<K, V> s : segments) {
x += s.sizeMapArray();
}
return x;
}
/**
* Get the number of hot entries in the cache.
*
* @return the number of hot entries
*/
public int sizeHot() {
int x = 0;
for (Segment<K, V> s : segments) {
x += s.sizeHot();
}
return x;
}
/**
* Get the number of resident entries.
*
* @return the number of entries
*/
public int size() {
int x = 0;
for (Segment<K, V> s : segments) {
x += s.size();
}
return x;
}
/**
* Get the list of keys. This method allows to view the internal state of
* the cache.
*
* @param cold if true, only keys for the cold entries are returned
* @param nonResident true for non-resident entries
* @return the key list
*/
public synchronized List<K> keys(boolean cold, boolean nonResident) {
ArrayList<K> keys = new ArrayList<K>();
for (Segment<K, V> s : segments) {
keys.addAll(s.keys(cold, nonResident));
}
return keys;
}
/**
* A cache segment
*
* @param <K> the key type
* @param <V> the value type
*/
static class Segment<K, V> {
/**
* How many other item are to be moved to the top of the stack before
* the current item is moved.
*/
private int stackMoveDistance = 20;
/**
* The maximum memory this cache should use.
*/
private long maxMemory;
/**
* The average memory used by one entry.
*/
private int averageMemory;
/**
* The currently used memory.
*/
private long usedMemory;
/**
* The number of (hot, cold, and non-resident) entries in the map.
*/
private int mapSize;
/**
* The bit mask that is applied to the key hash code to get the index in the
* map array. The mask is the length of the array minus one.
*/
private int mask;
/**
* The LIRS stack size.
*/
private int stackSize;
/**
* The size of the LIRS queue for resident cold entries.
*/
private int queueSize;
/**
* The size of the LIRS queue for non-resident cold entries.
*/
private int queue2Size;
/**
* The map array. The size is always a power of 2.
*/
private Entry<K, V>[] entries;
/**
* The stack of recently referenced elements. This includes all hot entries,
* the recently referenced cold entries, and all non-resident cold entries.
*/
private Entry<K, V> stack;
/**
* The queue of resident cold entries.
*/
private Entry<K, V> queue;
/**
* The queue of non-resident cold entries.
*/
private Entry<K, V> queue2;
/**
* The number of times any item was moved to the top of the stack.
*/
private int stackMoveCounter;
/**
* Create a new cache.
*
* @param maxMemory the maximum memory to use
* @param averageMemory the average memory usage of an object
*/
Segment(long maxMemory, int averageMemory) {
setMaxMemory(maxMemory);
setAverageMemory(averageMemory);
clear();
}
synchronized void clear() {
// calculate the size of the map array
// assume a fill factor of at most 80%
long maxLen = (long) (maxMemory / averageMemory / 0.75);
// the size needs to be a power of 2
long l = 8;
while (l < maxLen) {
l += l;
}
// the array size is at most 2^31 elements
int len = (int) Math.min(1L << 31, l);
// the bit mask has all bits set
mask = len - 1;
// initialize the stack and queue heads
stack = new Entry<K, V>();
stack.stackPrev = stack.stackNext = stack;
queue = new Entry<K, V>();
queue.queuePrev = queue.queueNext = queue;
queue2 = new Entry<K, V>();
queue2.queuePrev = queue2.queueNext = queue2;
// first set to null - avoiding out of memory
entries = null;
@SuppressWarnings("unchecked")
Entry<K, V>[] e = new Entry[len];
entries = e;
mapSize = 0;
usedMemory = 0;
stackSize = queueSize = queue2Size = 0;
}
V peek(K key, int hash) {
Entry<K, V> e = find(key, hash);
return e == null ? null : e.value;
}
int getMemory(K key, int hash) {
Entry<K, V> e = find(key, hash);
return e == null ? 0 : e.memory;
}
V get(Object key, int hash) {
Entry<K, V> e = find(key, hash);
if (e == null) {
// either the entry was not found
return null;
}
V value = e.value;
if (value == null) {
// it was a non-resident entry
return null;
}
if (e.isHot()) {
if (e != stack.stackNext) {
if (stackMoveDistance == 0 || stackMoveCounter - e.topMove > stackMoveDistance) {
access(key, hash);
}
}
} else {
access(key, hash);
}
return value;
}
/**
* Access an item, moving the entry to the top of the stack or front of the
* queue if found.
*
* @param key the key
*/
private synchronized void access(Object key, int hash) {
Entry<K, V> e = find(key, hash);
if (e == null || e.value == null) {
return;
}
if (e.isHot()) {
if (e != stack.stackNext) {
if (stackMoveDistance == 0 || stackMoveCounter - e.topMove > stackMoveDistance) {
// move a hot entries to the top of the stack
// unless it is already there
boolean wasEnd = e == stack.stackPrev;
removeFromStack(e);
if (wasEnd) {
// if moving the last entry, the last entry
// could not be cold, which is not allowed
pruneStack();
}
addToStack(e);
}
}
} else {
removeFromQueue(e);
if (e.stackNext != null) {
// resident cold entries become hot
// if they are on the stack
removeFromStack(e);
// which means a hot entry needs to become cold
convertOldestHotToCold();
} else {
// cold entries that are not on the stack
// move to the front of the queue
addToQueue(queue, e);
}
// in any case, the cold entry is moved to the top of the stack
addToStack(e);
}
}
V put(K key, V value, int hash) {
return put(key, value, hash, averageMemory);
}
synchronized V put(K key, V value, int hash, int memory) {
if (value == null) {
throw new NullPointerException();
}
V old;
Entry<K, V> e = find(key, hash);
if (e == null) {
old = null;
} else {
old = e.value;
remove(key, hash);
}
e = new Entry<K, V>();
e.key = key;
e.value = value;
e.memory = memory;
int index = hash & mask;
e.mapNext = entries[index];
entries[index] = e;
usedMemory += memory;
if (usedMemory > maxMemory && mapSize > 0) {
// an old entry needs to be removed
evict(e);
}
mapSize++;
// added entries are always added to the stack
addToStack(e);
return old;
}
synchronized V remove(Object key, int hash) {
int index = hash & mask;
Entry<K, V> e = entries[index];
if (e == null) {
return null;
}
V old;
if (e.key.equals(key)) {
old = e.value;
entries[index] = e.mapNext;
} else {
Entry<K, V> last;
do {
last = e;
e = e.mapNext;
if (e == null) {
return null;
}
} while (!e.key.equals(key));
old = e.value;
last.mapNext = e.mapNext;
}
mapSize--;
usedMemory -= e.memory;
if (e.stackNext != null) {
removeFromStack(e);
}
if (e.isHot()) {
// when removing a hot entry, the newest cold entry gets hot,
// so the number of hot entries does not change
e = queue.queueNext;
if (e != queue) {
removeFromQueue(e);
if (e.stackNext == null) {
addToStackBottom(e);
}
}
} else {
removeFromQueue(e);
}
pruneStack();
return old;
}
/**
* Evict cold entries (resident and non-resident) until the memory limit is
* reached.
*
* @param newCold a new cold entry
*/
private void evict(Entry<K, V> newCold) {
// ensure there are not too many hot entries:
// left shift of 5 is multiplication by 32, that means if there are less
// than 1/32 (3.125%) cold entries, a new hot entry needs to become cold
while ((queueSize << 5) < mapSize) {
convertOldestHotToCold();
}
// the new cold entry is at the top of the queue
addToQueue(queue, newCold);
// the oldest resident cold entries become non-resident
while (usedMemory > maxMemory) {
Entry<K, V> e = queue.queuePrev;
usedMemory -= e.memory;
removeFromQueue(e);
e.value = null;
e.memory = 0;
addToQueue(queue2, e);
// the size of the non-resident-cold entries needs to be limited
while (queue2Size + queue2Size > stackSize) {
e = queue2.queuePrev;
int hash = getHash(e.key);
remove(e.key, hash);
}
}
}
private void convertOldestHotToCold() {
// the last entry of the stack is known to be hot
Entry<K, V> last = stack.stackPrev;
// remove from stack - which is done anyway in the stack pruning, but we
// can do it here as well
removeFromStack(last);
// adding an entry to the queue will make it cold
addToQueue(queue, last);
pruneStack();
}
/**
* Ensure the last entry of the stack is cold.
*/
private void pruneStack() {
while (true) {
Entry<K, V> last = stack.stackPrev;
if (last == stack || last.isHot()) {
break;
}
// the cold entry is still in the queue
removeFromStack(last);
}
}
/**
* Try to find an entry in the map.
*
* @param key the key
* @return the entry (might be a non-resident)
*/
Entry<K, V> find(Object key, int hash) {
int index = hash & mask;
Entry<K, V> e = entries[index];
while (e != null && !e.key.equals(key)) {
e = e.mapNext;
}
return e;
}
private void addToStack(Entry<K, V> e) {
e.stackPrev = stack;
e.stackNext = stack.stackNext;
e.stackNext.stackPrev = e;
stack.stackNext = e;
stackSize++;
e.topMove = stackMoveCounter++;
}
private void addToStackBottom(Entry<K, V> e) {
e.stackNext = stack;
e.stackPrev = stack.stackPrev;
e.stackPrev.stackNext = e;
stack.stackPrev = e;
stackSize++;
}
private void removeFromStack(Entry<K, V> e) {
e.stackPrev.stackNext = e.stackNext;
e.stackNext.stackPrev = e.stackPrev;
e.stackPrev = e.stackNext = null;
stackSize--;
}
private void addToQueue(Entry<K, V> q, Entry<K, V> e) {
e.queuePrev = q;
e.queueNext = q.queueNext;
e.queueNext.queuePrev = e;
q.queueNext = e;
if (e.value != null) {
queueSize++;
} else {
queue2Size++;
}
}
private void removeFromQueue(Entry<K, V> e) {
e.queuePrev.queueNext = e.queueNext;
e.queueNext.queuePrev = e.queuePrev;
e.queuePrev = e.queueNext = null;
if (e.value != null) {
queueSize--;
} else {
queue2Size--;
}
}
synchronized List<K> keys(boolean cold, boolean nonResident) {
ArrayList<K> keys = new ArrayList<K>();
if (cold) {
Entry<K, V> start = nonResident ? queue2 : queue;
for (Entry<K, V> e = start.queueNext; e != start; e = e.queueNext) {
keys.add(e.key);
}
} else {
for (Entry<K, V> e = stack.stackNext; e != stack; e = e.stackNext) {
keys.add(e.key);
}
}
return keys;
}
int size() {
return mapSize - queue2Size;
}
boolean containsKey(Object key, int hash) {
Entry<K, V> e = find(key, hash);
return e != null && e.value != null;
}
synchronized Set<K> keySet() {
HashSet<K> set = new HashSet<K>();
for (Entry<K, V> e = stack.stackNext; e != stack; e = e.stackNext) {
set.add(e.key);
}
for (Entry<K, V> e = queue.queueNext; e != queue; e = e.queueNext) {
set.add(e.key);
}
return set;
}
synchronized Set<Map.Entry<K, V>> entrySet() {
HashMap<K, V> map = new HashMap<K, V>();
for (K k : keySet()) {
int hash = getHash(k);
map.put(k, find(k, hash).value);
}
return map.entrySet();
}
int sizeHot() {
return mapSize - queueSize - queue2Size;
}
int sizeNonResident() {
return queue2Size;
}
int sizeMapArray() {
return entries.length;
}
long getUsedMemory() {
return usedMemory;
}
void setMaxMemory(long maxMemory) {
if (maxMemory <= 0) {
throw new IllegalArgumentException("Max memory must be larger than 0");
}
this.maxMemory = maxMemory;
}
long getMaxMemory() {
return maxMemory;
}
void setAverageMemory(int averageMemory) {
if (averageMemory <= 0) {
throw new IllegalArgumentException("Average memory must be larger than 0");
}
this.averageMemory = averageMemory;
}
}
/**
* A cache entry. Each entry is either hot (low inter-reference recency;
* LIR), cold (high inter-reference recency; HIR), or non-resident-cold. Hot
* entries are in the stack only. Cold entries are in the queue, and may be
* in the stack. Non-resident-cold entries have their value set to null and
* are in the stack and in the non-resident queue.
*
* @param <K> the key type
* @param <V> the value type
*/
static class Entry<K, V> {
/**
* The key.
*/
K key;
/**
* The value. Set to null for non-resident-cold entries.
*/
V value;
/**
* The estimated memory used.
*/
int memory;
/**
* When the item was last moved to the top of the stack.
*/
int topMove;
/**
* The next entry in the stack.
*/
Entry<K, V> stackNext;
/**
* The previous entry in the stack.
*/
Entry<K, V> stackPrev;
/**
* The next entry in the queue (either the resident queue or the
* non-resident queue).
*/
Entry<K, V> queueNext;
/**
* The previous entry in the queue.
*/
Entry<K, V> queuePrev;
/**
* The next entry in the map
*/
Entry<K, V> mapNext;
/**
* Whether this entry is hot. Cold entries are in one of the two queues.
*
* @return whether the entry is hot
*/
boolean isHot() {
return queueNext == null;
}
}
}
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论