--- jsr166/src/jsr166e/StripedAdder.java 2011/07/20 16:06:19 1.2
+++ jsr166/src/jsr166e/StripedAdder.java 2011/07/30 16:26:34 1.12
@@ -5,7 +5,6 @@
*/
package jsr166e;
-import java.util.Arrays;
import java.util.Random;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.atomic.AtomicLong;
@@ -15,21 +14,33 @@ import java.io.ObjectInputStream;
import java.io.ObjectOutputStream;
/**
- * A set of variables that together maintain a sum. When updates
- * (method {@link #add}) are contended across threads, the set of
- * adders may grow to reduce contention. Method {@link #sum} returns
- * the current combined total across these adders. This value is
- * NOT an atomic snapshot (concurrent updates may occur while
- * the sum is being calculated), and so cannot be used alone for
- * fine-grained synchronization control.
+ * One or more variables that together maintain an initially zero sum.
+ * When updates (method {@link #add}) are contended across threads,
+ * the set of variables may grow dynamically to reduce contention.
*
- *
This class may be applicable when many threads frequently
- * update a common sum that is used for purposes such as collecting
- * statistics. In this case, performance may be significantly faster
- * than using a shared {@link AtomicLong}, at the expense of using
- * significantly more space. On the other hand, if it is known that
- * only one thread can ever update the sum, performance may be
- * significantly slower than just updating a local variable.
+ *
This class is usually preferable to {@link AtomicLong} when
+ * multiple threads update a common sum that is used for purposes such
+ * as collecting statistics, not for fine-grained synchronization
+ * control. Under high update contention, throughput of this class is
+ * expected to be significantly higher, at the expense of higher space
+ * consumption. Under low contention, this class imposes very little
+ * time and space overhead compared to AtomicLong. On the other hand,
+ * in contexts where it is statically known that only one thread can
+ * ever update a sum, time and space overhead is noticeably greater
+ * than just updating a local variable.
+ *
+ *
Method {@link #sum} returns the current combined total across
+ * the variables maintaining the sum. This value is NOT an
+ * atomic snapshot: Concurrent updates may occur while the sum is
+ * being calculated. However, updates cannot be "lost", so invocation
+ * of sum
in the absence of concurrent updates always
+ * returns an accurate result. The sum may also be reset
+ * to zero, as an alternative to creating a new adder. However,
+ * method {@link #reset} is intrinsically racy, so should only be used
+ * when it is known that no threads are concurrently updating the sum.
+ *
+ *
jsr166e note: This class is targeted to be placed in
+ * java.util.concurrent.atomic
*
* @author Doug Lea
*/
@@ -37,90 +48,128 @@ public class StripedAdder implements Ser
private static final long serialVersionUID = 7249069246863182397L;
/*
- * Overview: We maintain a table of AtomicLongs (padded to reduce
- * false sharing). The table is indexed by per-thread hash codes
- * that are initialized as random values. The table doubles in
- * size upon contention (as indicated by failed CASes when
- * performing add()), but is capped at the nearest power of two >=
- * #cpus: At that point, contention should be infrequent if each
- * thread has a unique index; so we instead adjust hash codes to
- * new random values upon contention rather than expanding. A
- * single spinlock is used for resizing the table as well as
- * populating slots with new Adders. Upon lock contention, threads
- * just try other slots rather than blocking. We guarantee that at
- * least one slot exists, so retries will eventually find a
- * candidate Adder.
+ * A StripedAdder maintains a lazily-initialized table of
+ * atomically updated variables, plus an extra "base" field. The
+ * table size is a power of two. Indexing uses masked per-thread
+ * hash codes
+ *
+ * Table entries are of class Cell; a variant of AtomicLong padded
+ * to reduce cache contention on most processors. Padding is
+ * overkill for most Atomics because they are usually irregularly
+ * scattered in memory and thus don't interfere much with each
+ * other. But Atomic objects residing in arrays will tend to be
+ * placed adjacent to each other, and so will most often share
+ * cache lines (with a huge negative performance impact) without
+ * this precaution.
+ *
+ * In part because Cells are relatively large, we avoid creating
+ * them until they are needed. When there is no contention, all
+ * updates are made to the base field. Upon first contention (a
+ * failed CAS on base update), the table is initialized to size 2.
+ * The table size is doubled upon further contention until
+ * reaching the nearest power of two greater than or equal to the
+ * number of CPUS.
+ *
+ * Per-thread hash codes are initialized to random values.
+ * Contention and/or table collisions are indicated by failed
+ * CASes when performing an add operation (see method
+ * retryAdd). Upon a collision, if the table size is less than the
+ * capacity, it is doubled in size unless some other thread holds
+ * the lock. If a hashed slot is empty, and lock is available, a
+ * new Cell is created. Otherwise, if the slot exists, a CAS is
+ * tried. Retries proceed by "double hashing", using a secondary
+ * hash (Marsaglia XorShift) to try to find a free slot.
+ *
+ * The table size is capped because, when there are more threads
+ * than CPUs, supposing that each thread were bound to a CPU,
+ * there would exist a perfect hash function mapping threads to
+ * slots that eliminates collisions. When we reach capacity, we
+ * search for this mapping by randomly varying the hash codes of
+ * colliding threads. Because search is random, and collisions
+ * only become known via CAS failures, convergence can be slow,
+ * and because threads are typically not bound to CPUS forever,
+ * may not occur at all. However, despite these limitations,
+ * observed contention rates are typically low in these cases.
+ *
+ * A single spinlock is used for initializing and resizing the
+ * table, as well as populating slots with new Cells. There is no
+ * need for a blocking lock: Upon lock contention, threads try
+ * other slots (or the base) rather than blocking. During these
+ * retries, there is increased contention and reduced locality,
+ * which is still better than alternatives.
+ *
+ * It is possible for a Cell to become unused when threads that
+ * once hashed to it terminate, as well as in the case where
+ * doubling the table causes no thread to hash to it under
+ * expanded mask. We do not try to detect or remove such cells,
+ * under the assumption that for long-running adders, observed
+ * contention levels will recur, so the cells will eventually be
+ * needed again; and for short-lived ones, it does not matter.
+ *
*/
- /**
- * Number of processors, to place a cap on table growth.
- */
- static final int NCPU = Runtime.getRuntime().availableProcessors();
+ private static final int NCPU = Runtime.getRuntime().availableProcessors();
/**
- * Version of AtomicLong padded to avoid sharing cache
- * lines on most processors
+ * Padded variant of AtomicLong. The value field is placed
+ * between pads, hoping that the JVM doesn't reorder them.
+ * Updates are via inlined CAS in methods add and retryAdd.
*/
- static final class Adder extends AtomicLong {
- long p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, pa, pb, pc, pd;
- Adder(long x) { super(x); }
+ static final class Cell {
+ volatile long p0, p1, p2, p3, p4, p5, p6;
+ volatile long value;
+ volatile long q0, q1, q2, q3, q4, q5, q6;
+ Cell(long x) { value = x; }
}
/**
- * Holder for the thread-local hash code.
+ * Holder for the thread-local hash code. The code is initially
+ * random, but may be set to a different value upon collisions.
*/
static final class HashCode {
+ static final Random rng = new Random();
int code;
- HashCode(int h) { code = h; }
+ HashCode() {
+ int h = rng.nextInt(); // Avoid zero to allow xorShift rehash
+ code = (h == 0) ? 1 : h;
+ }
}
/**
* The corresponding ThreadLocal class
*/
static final class ThreadHashCode extends ThreadLocal {
- static final Random rng = new Random();
- public HashCode initialValue() {
- int h = rng.nextInt();
- return new HashCode((h == 0) ? 1 : h); // ensure nonzero
- }
+ public HashCode initialValue() { return new HashCode(); }
}
/**
* Static per-thread hash codes. Shared across all StripedAdders
- * because adjustments due to collisions in one table are likely
- * to be appropriate for others.
+ * to reduce ThreadLocal pollution and because adjustments due to
+ * collisions in one table are likely to be appropriate for
+ * others.
*/
static final ThreadHashCode threadHashCode = new ThreadHashCode();
/**
- * Table of adders. Initially of size 2; grows to be at most NCPU.
+ * Table of cells. When non-null, size is a power of 2.
*/
- private transient volatile Adder[] adders;
+ private transient volatile Cell[] cells;
/**
- * Serves as a lock when resizing and/or creating Adders. There
- * is no need for a blocking lock: When busy, other threads try
- * other slots.
+ * Base sum, used mainly when there is no contention, but also as
+ * a fallback during table initializion races. Updated via CAS.
*/
- private final AtomicInteger mutex;
+ private transient volatile long base;
/**
- * Marsaglia XorShift for rehashing on collisions
+ * Spinlock (locked via CAS) used when resizing and/or creating Cells.
*/
- private static int xorShift(int r) {
- r ^= r << 13;
- r ^= r >>> 17;
- return r ^ (r << 5);
- }
+ private transient volatile int busy;
/**
- * Creates a new adder with initially zero sum.
+ * Creates a new adder with initial sum of zero.
*/
public StripedAdder() {
- Adder[] as = new Adder[2];
- as[0] = new Adder(0); // ensure at least one available adder
- this.adders = as;
- this.mutex = new AtomicInteger();
}
/**
@@ -129,107 +178,196 @@ public class StripedAdder implements Ser
* @param x the value to add
*/
public void add(long x) {
- HashCode hc = threadHashCode.get();
- for (int h = hc.code;;) {
- Adder[] as = adders;
- int n = as.length;
- Adder a = as[h & (n - 1)];
- if (a != null) {
- long v = a.get();
- if (a.compareAndSet(v, v + x))
- break;
- if (n >= NCPU) { // Collision when table at max
- h = hc.code = xorShift(h); // change code
- continue;
+ Cell[] as; long v; HashCode hc; Cell a; int n; boolean contended;
+ if ((as = cells) != null ||
+ !UNSAFE.compareAndSwapLong(this, baseOffset, v = base, v + x)) {
+ int h = (hc = threadHashCode.get()).code;
+ if (as != null && (n = as.length) > 0 &&
+ (a = as[(n - 1) & h]) != null) {
+ if (UNSAFE.compareAndSwapLong(a, valueOffset,
+ v = a.value, v + x))
+ return;
+ contended = true;
+ }
+ else
+ contended = false;
+ retryAdd(x, hc, contended);
+ }
+ }
+
+ /**
+ * Handle cases of add involving initialization, resizing,
+ * creating new Cells, and/or contention. See above for
+ * explanation. This method suffers the usual non-modularity
+ * problems of optimistic retry code, relying on rechecked sets of
+ * reads.
+ *
+ * @param x the value to add
+ * @param hc the hash code holder
+ * @param precontended true if CAS failed before call
+ */
+ private void retryAdd(long x, HashCode hc, boolean precontended) {
+ int h = hc.code;
+ boolean collide = false; // true if last slot nonempty
+ for (;;) {
+ Cell[] as; Cell a; int n;
+ if ((as = cells) != null && (n = as.length) > 0) {
+ if ((a = as[(n - 1) & h]) == null) {
+ if (busy == 0) { // Try to attach new Cell
+ Cell r = new Cell(x); // Optimistically create
+ if (busy == 0 &&
+ UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
+ boolean created = false;
+ try { // Recheck under lock
+ Cell[] rs; int m, j;
+ if ((rs = cells) != null &&
+ (m = rs.length) > 0 &&
+ rs[j = (m - 1) & h] == null) {
+ rs[j] = r;
+ created = true;
+ }
+ } finally {
+ busy = 0;
+ }
+ if (created)
+ break;
+ continue; // Slot is now non-empty
+ }
+ }
+ collide = false;
}
+ else if (precontended) // CAS already known to fail
+ precontended = false; // Continue after rehash
+ else {
+ long v = a.value;
+ if (UNSAFE.compareAndSwapLong(a, valueOffset, v, v + x))
+ break;
+ if (!collide)
+ collide = true;
+ else if (n >= NCPU || cells != as)
+ collide = false; // Can't expand
+ else if (busy == 0 &&
+ UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
+ collide = false;
+ try {
+ if (cells == as) { // Expand table
+ Cell[] rs = new Cell[n << 1];
+ for (int i = 0; i < n; ++i)
+ rs[i] = as[i];
+ cells = rs;
+ }
+ } finally {
+ busy = 0;
+ }
+ continue;
+ }
+ }
+ h ^= h << 13; // Rehash
+ h ^= h >>> 17;
+ h ^= h << 5;
}
- final AtomicInteger mutex = this.mutex;
- if (mutex.get() != 0)
- h = xorShift(h); // Try elsewhere
- else if (mutex.compareAndSet(0, 1)) {
- boolean created = false;
- try {
- Adder[] rs = adders;
- if (a != null && rs == as) // Resize table
- rs = adders = Arrays.copyOf(as, as.length << 1);
- int j = h & (rs.length - 1);
- if (rs[j] == null) { // Create adder
- rs[j] = new Adder(x);
- created = true;
+ else if (busy == 0 && cells == as &&
+ UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
+ boolean init = false;
+ try { // Initialize
+ if (cells == as) {
+ Cell r = new Cell(x);
+ Cell[] rs = new Cell[2];
+ rs[h & 1] = r;
+ cells = rs;
+ init = true;
}
} finally {
- mutex.set(0);
+ busy = 0;
}
- if (created) {
- hc.code = h; // Use this adder next time
+ if (init)
+ break;
+ }
+ else { // Lost initialization race
+ long b = base; // Fall back on using base
+ if (UNSAFE.compareAndSwapLong(this, baseOffset, b, b + x))
break;
- }
}
}
+ hc.code = h; // Record index for next time
}
/**
- * Returns an estimate of the current sum. The result is
- * calculated by summing multiple variables, so may not be
- * accurate if updates occur concurrently with this method.
- *
- * @return the estimated sum
+ * Equivalent to {@code add(1)}.
*/
- public long sum() {
- long sum = 0;
- Adder[] as = adders;
- int n = as.length;
- for (int i = 0; i < n; ++i) {
- Adder a = as[i];
- if (a != null)
- sum += a.get();
- }
- return sum;
+ public void increment() {
+ add(1L);
}
/**
- * Resets each of the variables to zero. This is effective in
- * fully resetting the sum only if there are no concurrent
- * updates.
+ * Equivalent to {@code add(-1)}.
*/
- public void reset() {
- Adder[] as = adders;
- int n = as.length;
- for (int i = 0; i < n; ++i) {
- Adder a = as[i];
- if (a != null)
- a.set(0L);
- }
+ public void decrement() {
+ add(-1L);
}
/**
- * Equivalent to {@code add(1)}.
+ * Returns the current sum. The result is only guaranteed to be
+ * accurate in the absence of concurrent updates. Otherwise, it
+ * may fail to reflect one or more updates occuring while
+ * calculating the result.
+ *
+ * @return the sum
*/
- public void increment() {
- add(1L);
+ public long sum() {
+ Cell[] as = cells;
+ long sum = base;
+ if (as != null) {
+ int n = as.length;
+ for (int i = 0; i < n; ++i) {
+ Cell a = as[i];
+ if (a != null)
+ sum += a.value;
+ }
+ }
+ return sum;
}
/**
- * Equivalent to {@code add(-1)}.
+ * Resets variables maintaining the sum to zero. This is
+ * effective in setting the sum to zero only if there are no
+ * concurrent updates.
*/
- public void decrement() {
- add(-1L);
+ public void reset() {
+ Cell[] as = cells;
+ base = 0L;
+ if (as != null) {
+ int n = as.length;
+ for (int i = 0; i < n; ++i) {
+ Cell a = as[i];
+ if (a != null)
+ a.value = 0L;
+ }
+ }
}
/**
- * Equivalent to {@link #sum} followed by {@link #reset}.
+ * Equivalent in effect to {@link #sum} followed by {@link
+ * #reset}. This method may apply for example during quiescent
+ * points between multithreaded computations. If there are
+ * updates concurrent with this method, the returned value is
+ * not guaranteed to be the final sum occurring before
+ * the reset.
*
- * @return the estimated sum
+ * @return the sum
*/
- public long sumAndReset() {
- long sum = 0;
- Adder[] as = adders;
- int n = as.length;
- for (int i = 0; i < n; ++i) {
- Adder a = as[i];
- if (a != null) {
- sum += a.get();
- a.set(0L);
+ public long sumThenReset() {
+ Cell[] as = cells;
+ long sum = base;
+ base = 0L;
+ if (as != null) {
+ int n = as.length;
+ for (int i = 0; i < n; ++i) {
+ Cell a = as[i];
+ if (a != null) {
+ sum += a.value;
+ a.value = 0L;
+ }
}
}
return sum;
@@ -244,13 +382,58 @@ public class StripedAdder implements Ser
private void readObject(ObjectInputStream s)
throws IOException, ClassNotFoundException {
s.defaultReadObject();
- long c = s.readLong();
- Adder[] as = new Adder[2];
- as[0] = new Adder(c);
- this.adders = as;
- mutex.set(0);
+ busy = 0;
+ cells = null;
+ base = s.readLong();
+ }
+
+ // Unsafe mechanics
+ private static final sun.misc.Unsafe UNSAFE;
+ private static final long baseOffset;
+ private static final long busyOffset;
+ private static final long valueOffset;
+ static {
+ try {
+ UNSAFE = getUnsafe();
+ Class> sk = StripedAdder.class;
+ baseOffset = UNSAFE.objectFieldOffset
+ (sk.getDeclaredField("base"));
+ busyOffset = UNSAFE.objectFieldOffset
+ (sk.getDeclaredField("busy"));
+ Class> ak = Cell.class;
+ valueOffset = UNSAFE.objectFieldOffset
+ (ak.getDeclaredField("value"));
+ } catch (Exception e) {
+ throw new Error(e);
+ }
}
-}
-
+ /**
+ * Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package.
+ * Replace with a simple call to Unsafe.getUnsafe when integrating
+ * into a jdk.
+ *
+ * @return a sun.misc.Unsafe
+ */
+ private static sun.misc.Unsafe getUnsafe() {
+ try {
+ return sun.misc.Unsafe.getUnsafe();
+ } catch (SecurityException se) {
+ try {
+ return java.security.AccessController.doPrivileged
+ (new java.security
+ .PrivilegedExceptionAction() {
+ public sun.misc.Unsafe run() throws Exception {
+ java.lang.reflect.Field f = sun.misc
+ .Unsafe.class.getDeclaredField("theUnsafe");
+ f.setAccessible(true);
+ return (sun.misc.Unsafe) f.get(null);
+ }});
+ } catch (java.security.PrivilegedActionException e) {
+ throw new RuntimeException("Could not initialize intrinsics",
+ e.getCause());
+ }
+ }
+ }
+}