ViewVC Help

View File | Revision Log | Show Annotations | Download File | Root Listing

root/jsr166/jsr166/src/jsr166e/StripedAdder.java

Comparing jsr166/src/jsr166e/StripedAdder.java (file contents):
Revision 1.7 by dl, Tue Jul 26 18:30:35 2011 UTC vs.
Revision 1.8 by dl, Thu Jul 28 15:05:55 2011 UTC

#	Line 5 | Line 5
5		*/
6
7		package jsr166e;
8	–	import java.util.Arrays;
8		import java.util.Random;
9		import java.util.concurrent.atomic.AtomicInteger;
10		import java.util.concurrent.atomic.AtomicLong;
#	Line 46 | Line 45 | public class StripedAdder implements Ser
45		* A StripedAdder maintains a table of Atomic long variables. The
46		* table is indexed by per-thread hash codes.
47		*
48	<	* By default, the table is lazily initialized, to minimize
49	<	* footprint until adders are used. On first use, the table is set
50	<	* to size DEFAULT_INITIAL_SIZE (currently 8). Table size is
51	<	* bounded by the number of CPUS (if larger than the default
52	<	* size).
48	>	* Table entries are of class Adder; a variant of AtomicLong
49	>	* padded to reduce cache contention on most processors. Padding
50	>	* is overkill for most Atomics because they are usually
51	>	* irregularly scattered in memory and thus don't interfere much
52	>	* with each other. But Atomic objects residing in arrays will
53	>	* tend to be placed adjacent to each other, and so will most
54	>	* often share cache lines (with a huge negative performance
55	>	* impact) without this precaution.
56	>	*
57	>	* Because Adders are relatively large, we avoid creating them
58	>	* until they are needed. On the other hand, we try to create them
59	>	* on any sign of contention.
60		*
61		* Per-thread hash codes are initialized to random values.
62		* Collisions are indicated by failed CASes when performing an add
#	Line 62 | Line 68 | public class StripedAdder implements Ser
68		* hashing", using a secondary hash (Marsaglia XorShift) to try to
69		* find a free slot.
70		*
71	<	* The table size is capped because, when there are more threads
72	<	* than CPUs, supposing that each thread were bound to a CPU,
73	<	* there would exist a perfect hash function mapping threads to
74	<	* slots that eliminates collisions. When we reach capacity, we
75	<	* search for this mapping by randomly varying the hash codes of
71	>	* By default, the table is lazily initialized.  Upon first use,
72	>	* the table is set to size 2 (the minimum non-empty size), but
73	>	* containing only a single Adder. The maximum table size is
74	>	* bounded by nearest power of two >= the number of CPUS.  The
75	>	* table size is capped because, when there are more threads than
76	>	* CPUs, supposing that each thread were bound to a CPU, there
77	>	* would exist a perfect hash function mapping threads to slots
78	>	* that eliminates collisions. When we reach capacity, we search
79	>	* for this mapping by randomly varying the hash codes of
80		* colliding threads.  Because search is random, and failures only
81		* become known via CAS failures, convergence will be slow, and
82		* because threads are typically not bound to CPUS forever, may
83		* not occur at all. However, despite these limitations, observed
84		* contention is typically low in these cases.
85		*
76	–	* Table entries are of class Adder; a form of AtomicLong padded
77	–	* to reduce cache contention on most processors. Padding is
78	–	* overkill for most Atomics because they are usually irregularly
79	–	* scattered in memory and thus don't interfere much with each
80	–	* other. But Atomic objects residing in arrays will tend to be
81	–	* placed adjacent to each other, and so will most often share
82	–	* cache lines without this precaution.  Adders are by default
83	–	* constructed upon first use, which further improves per-thread
84	–	* locality and helps reduce footprint.
85	–	*
86		* A single spinlock is used for resizing the table as well as
87	<	* populating slots with new Adders. Upon lock contention, threads
87	>	* populating slots with new Adders. After initialization, there
88	>	* is no need for a blocking lock: Upon lock contention, threads
89		* try other slots rather than blocking. After initialization, at
90		* least one slot exists, so retries will eventually find a
91		* candidate Adder.  During these retries, there is increased
#	Line 92 | Line 93 | public class StripedAdder implements Ser
93		* alternatives.
94		*/
95
95	–	/**
96	–	* Padded version of AtomicLong
97	–	*/
98	–	static final class Adder extends AtomicLong {
99	–	long p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, pa, pb, pc, pd, pe;
100	–	Adder(long x) { super(x); }
101	–	}
102	–
96		private static final int NCPU = Runtime.getRuntime().availableProcessors();
97
98		/**
99	<	* Table bounds. DEFAULT_INITIAL_SIZE is the table size set upon
100	<	* first use under default constructor, and must be a power of
101	<	* two. There is not much point in making size a lot smaller than
102	<	* that of Adders though.  CAP is the maximum allowed table size.
103	<	*/
104	<	private static final int DEFAULT_INITIAL_SIZE = 8;
105	<	private static final int CAP = Math.max(NCPU, DEFAULT_INITIAL_SIZE);
99	>	* Padded variant of AtomicLong.  The value field is placed
100	>	* between pads, hoping that the JVM doesn't reorder them.
101	>	* Updates are via inlined CAS in methods add and retryAdd.
102	>	*/
103	>	static final class Adder {
104	>	volatile long p0, p1, p2, p3, p4, p5, p6;
105	>	volatile long value;
106	>	volatile long q0, q1, q2, q3, q4, q5, q6;
107	>	Adder(long x) { value = x; }
108	>	}
109
110		/**
111		* Holder for the thread-local hash code. The code is initially
#	Line 119 | Line 115 | public class StripedAdder implements Ser
115		static final Random rng = new Random();
116		int code;
117		HashCode() {
118	<	int h = rng.nextInt();
119	<	code = (h == 0) ? 1 : h; // ensure nonzero
118	>	int h = rng.nextInt(); // Avoid zero, because of xorShift rehash
119	>	code = (h == 0) ? 1 : h;
120		}
121		}
122
#	Line 140 | Line 136 | public class StripedAdder implements Ser
136		static final ThreadHashCode threadHashCode = new ThreadHashCode();
137
138		/**
139	<	* Common placeholder for empty arrays.
144	<	*/
145	<	static final Adder[] EMPTY_ARRAY = new Adder[0];
146	<
147	<	/**
148	<	* Table of adders. Size is either zero or a power of two, grows
149	<	* to be at most CAP.
139	>	* Table of adders. When non-null, size is a power of 2, at least 2.
140		*/
141		private transient volatile Adder[] adders;
142
143		/**
144	<	* Serves as a lock when resizing and/or creating Adders.  There
155	<	* is no need for a blocking lock: Except during initialization
156	<	* races, when busy, other threads try other slots. However,
157	<	* during (double-checked) initializations, we use the
158	<	* "synchronized" lock on this object.
144	>	* Spinlock (locked via CAS) used when resizing and/or creating Adders.
145		*/
146	<	private final AtomicInteger mutex;
146	>	private volatile int busy;
147
148		/**
149		* Creates a new adder with zero sum.
150		*/
151		public StripedAdder() {
166	–	this.adders = EMPTY_ARRAY;
167	–	this.mutex = new AtomicInteger();
168	–	// remaining initialization on first call to add.
152		}
153
154		/**
#	Line 176 | Line 159 | public class StripedAdder implements Ser
159		* will concurrently update the sum.
160		*/
161		public StripedAdder(int expectedContention) {
162	<	if (expectedContention > 0) {
163	<	int cap = (expectedContention < CAP) ? expectedContention : CAP;
164	<	int size = 1;
165	<	while (size < cap)
166	<	size <<= 1;
167	<	Adder[] as = new Adder[size];
168	<	for (int i = 0; i < size; ++i)
169	<	as[i] = new Adder(0);
187	<	this.adders = as;
188	<	}
189	<	else
190	<	this.adders = EMPTY_ARRAY;
191	<	this.mutex = new AtomicInteger();
162	>	int cap = (expectedContention < NCPU) ? expectedContention : NCPU;
163	>	int size = 2;
164	>	while (size < cap)
165	>	size <<= 1;
166	>	Adder[] as = new Adder[size];
167	>	for (int i = 0; i < size; ++i)
168	>	as[i] = new Adder(0);
169	>	this.adders = as;
170		}
171
172		/**
#	Line 197 | Line 175 | public class StripedAdder implements Ser
175		* @param x the value to add
176		*/
177		public void add(long x) {
178	<	Adder[] as; Adder a; int n; long v; // locals to hold volatile reads
178	>	Adder[] as; Adder a; int n;  // locals to hold volatile reads
179		HashCode hc = threadHashCode.get();
180		int h = hc.code;
181	<	if ((as = adders) == null || (n = as.length) < 1 ||
182	<	(a = as[(n - 1) & h]) == null ||
183	<	!a.compareAndSet(v = a.get(), v + x))
184	<	retryAdd(x, hc);
181	>	boolean collide;
182	>	if ((as = adders) != null && (n = as.length) > 0 &&
183	>	(a = as[(n - 1) & h]) != null) {
184	>	long v = a.value;
185	>	if (UNSAFE.compareAndSwapLong(a, valueOffset, v, v + x))
186	>	return;
187	>	collide = true;
188	>	}
189	>	else
190	>	collide = false;
191	>	retryAdd(x, hc, collide);
192		}
193
194		/**
195		* Handle cases of add involving initialization, resizing,
196		* creating new Adders, and/or contention. See above for
197	<	* explanation.
197	>	* explanation. This method suffers the usual non-modularity
198	>	* problems of optimistic retry code, relying on rechecked sets of
199	>	* reads.
200		*/
201	<	private void retryAdd(long x, HashCode hc) {
201	>	private void retryAdd(long x, HashCode hc, boolean collide) {
202		int h = hc.code;
216	–	final AtomicInteger mutex = this.mutex;
217	–	int collisions = 1 - mutex.get(); // first guess: collides if not locked
203		for (;;) {
204	<	Adder[] as; Adder a; long v; int k, n;
205	<	while ((as = adders) == null || (n = as.length) < 1) {
206	<	synchronized(mutex) {                // Try to initialize
207	<	if (adders == as) {
208	<	Adder[] rs = new Adder[DEFAULT_INITIAL_SIZE];
209	<	rs[h & (DEFAULT_INITIAL_SIZE - 1)] = new Adder(0);
210	<	adders = rs;
204	>	Adder[] as; Adder a; int n;
205	>	if ((as = adders) != null && (n = as.length) > 0) {
206	>	if ((a = as[(n - 1) & h]) != null) {
207	>	boolean shared = true;      // Slot exists
208	>	if (collide && n < NCPU && busy == 0 &&
209	>	UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
210	>	try {
211	>	if (adders == as) { // Expand table
212	>	Adder[] rs = new Adder[n << 1];
213	>	for (int i = 0; i < n; ++i)
214	>	rs[i] = as[i];
215	>	adders = rs;
216	>	shared = false;
217	>	}
218	>	} finally {
219	>	busy = 0;
220	>	}
221	>	if (shared || (h & n) != 0) {
222	>	collide = false;
223	>	continue;           // Array or index changed
224	>	}
225	>	}
226	>	long v = a.value;
227	>	if (UNSAFE.compareAndSwapLong(a, valueOffset, v, v + x))
228	>	break;
229	>	collide = shared;
230	>	}
231	>	else {                          // Try to attach new Adder
232	>	if (busy == 0 &&
233	>	UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
234	>	boolean created = false;
235	>	try {                   // Recheck under lock
236	>	Adder[] rs; int m, j;
237	>	if ((rs = adders) != null && (m = rs.length) > 0 &&
238	>	rs[j = (m - 1) & h] == null) {
239	>	rs[j] = new Adder(x);
240	>	created = true;
241	>	}
242	>	} finally {
243	>	busy = 0;
244	>	}
245	>	if (created)
246	>	break;
247	>	continue;               // Slot is now non-empty
248		}
249	+	collide = false;
250		}
251	<	collisions = 0;
251	>	h ^= h << 13;                   // Rehash
252	>	h ^= h >>> 17;
253	>	h ^= h << 5;
254		}
255	<
256	<	if ((a = as[k = (n - 1) & h]) == null) { // Try to add slot
257	<	if (mutex.get() == 0 && mutex.compareAndSet(0, 1)) {
255	>	else if (busy == 0) {               // Default-initialize
256	>	Adder r = new Adder(x);
257	>	Adder[] rs = new Adder[2];
258	>	rs[h & 1] = r;
259	>	if (adders == as && busy == 0 &&
260	>	UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
261	>	boolean init = false;
262		try {
263	<	if (adders == as && as[k] == null)
264	<	a = as[k] = new Adder(x);
263	>	if (adders == as) {
264	>	adders = rs;
265	>	init = true;
266	>	}
267		} finally {
268	<	mutex.set(0);
268	>	busy = 0;
269		}
270	<	if (a != null)
270	>	if (init)
271		break;
272		}
242	–	collisions = 0;
243	–	}
244	–	else if (collisions != 0 && n < CAP &&   // Try to expand table
245	–	mutex.get() == 0 && mutex.compareAndSet(0, 1)) {
246	–	try {
247	–	if (adders == as) {
248	–	Adder[] rs = new Adder[n << 1];
249	–	for (int i = 0; i < n; ++i)
250	–	rs[i] = as[i];
251	–	adders = rs;
252	–	}
253	–	} finally {
254	–	mutex.set(0);
255	–	}
256	–	collisions = 0;
273		}
274	<	else if (a.compareAndSet(v = a.get(), v + x))
275	<	break;
260	<	else
261	<	collisions = 1;
262	<	h ^= h << 13;                            // Rehash
263	<	h ^= h >>> 17;
264	<	h ^= h << 5;
274	>	else if (adders == as)              // Lost initialization race
275	>	Thread.yield();
276		}
277	<	hc.code = h;
277	>	hc.code = h;                            // Record index for next time
278		}
279
280		/**
#	Line 281 | Line 292 | public class StripedAdder implements Ser
292		for (int i = 0; i < n; ++i) {
293		Adder a = as[i];
294		if (a != null)
295	<	sum += a.get();
295	>	sum += a.value;
296		}
297		}
298		return sum;
#	Line 299 | Line 310 | public class StripedAdder implements Ser
310		for (int i = 0; i < n; ++i) {
311		Adder a = as[i];
312		if (a != null)
313	<	a.set(0L);
313	>	a.value = 0L;
314		}
315		}
316		}
#	Line 331 | Line 342 | public class StripedAdder implements Ser
342		for (int i = 0; i < n; ++i) {
343		Adder a = as[i];
344		if (a != null) {
345	<	sum += a.get();
346	<	a.set(0L);
345	>	sum += a.value;
346	>	a.value = 0L;
347		}
348		}
349		}
#	Line 348 | Line 359 | public class StripedAdder implements Ser
359		private void readObject(ObjectInputStream s)
360		throws IOException, ClassNotFoundException {
361		s.defaultReadObject();
362	<	mutex.set(0);
362	>	busy = 0;
363		add(s.readLong());
364		}
365
366	+	// Unsafe mechanics
367	+	private static final sun.misc.Unsafe UNSAFE;
368	+	private static final long busyOffset;
369	+	private static final long valueOffset;
370	+	static {
371	+	try {
372	+	UNSAFE = getUnsafe();
373	+	Class<?> sk = StripedAdder.class;
374	+	busyOffset = UNSAFE.objectFieldOffset
375	+	(sk.getDeclaredField("busy"));
376	+	Class<?> ak = Adder.class;
377	+	valueOffset = UNSAFE.objectFieldOffset
378	+	(ak.getDeclaredField("value"));
379	+	} catch (Exception e) {
380	+	throw new Error(e);
381	+	}
382	+	}
383	+
384	+	/**
385	+	* Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
386	+	* Replace with a simple call to Unsafe.getUnsafe when integrating
387	+	* into a jdk.
388	+	*
389	+	* @return a sun.misc.Unsafe
390	+	*/
391	+	private static sun.misc.Unsafe getUnsafe() {
392	+	try {
393	+	return sun.misc.Unsafe.getUnsafe();
394	+	} catch (SecurityException se) {
395	+	try {
396	+	return java.security.AccessController.doPrivileged
397	+	(new java.security
398	+	.PrivilegedExceptionAction<sun.misc.Unsafe>() {
399	+	public sun.misc.Unsafe run() throws Exception {
400	+	java.lang.reflect.Field f = sun.misc
401	+	.Unsafe.class.getDeclaredField("theUnsafe");
402	+	f.setAccessible(true);
403	+	return (sun.misc.Unsafe) f.get(null);
404	+	}});
405	+	} catch (java.security.PrivilegedActionException e) {
406	+	throw new RuntimeException("Could not initialize intrinsics",
407	+	e.getCause());
408	+	}
409	+	}
410	+	}
411	+
412		}

Diff Legend

–	Removed lines
+	Added lines
<	Changed lines
>	Changed lines