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 1.3 by dl, Fri Jul 22 13:25:12 2011 UTC vs.
Revision 1.10 by dl, Fri Jul 29 13:50:54 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 27 | Line 26 | import java.io.ObjectOutputStream;
26		* update a common sum that is used for purposes such as collecting
27		* statistics. In this case, performance may be significantly faster
28		* than using a shared {@link AtomicLong}, at the expense of using
29	<	* significantly more space.  On the other hand, if it is known that
30	<	* only one thread can ever update the sum, performance may be
31	<	* significantly slower than just updating a local variable.
29	>	* more space.  On the other hand, if it is known that only one thread
30	>	* can ever update the sum, performance may be significantly slower
31	>	* than just updating a local variable.
32		*
33		* <p>A StripedAdder may optionally be constructed with a given
34		* expected contention level; i.e., the number of threads that are
#	Line 43 | Line 42 | public class StripedAdder implements Ser
42		private static final long serialVersionUID = 7249069246863182397L;
43
44		/*
45	<	* Overview: We maintain a table of Atomic long variables. The
46	<	* table is indexed by per-thread hash codes that are initialized
47	<	* to random values.
48	<	*
49	<	* The table doubles in size upon contention (as indicated by
50	<	* failed CASes when performing add()), but is capped at the
51	<	* nearest power of two >= #CPUS. This reflects the idea that,
52	<	* when there are more threads than CPUs, then if each thread were
45	>	* A StripedAdder maintains a table of Atomic long variables. The
46	>	* table is indexed by per-thread hash codes.
47	>	*
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
63	>	* operation (see method retryAdd). Upon a collision, if the table
64	>	* size is less than the capacity, it is doubled in size unless
65	>	* some other thread holds lock. If a hashed slot is empty, and
66	>	* lock is available, a new Adder is created. Otherwise, if the
67	>	* slot exists, a CAS is tried.  Retries proceed by "double
68	>	* hashing", using a secondary hash (Marsaglia XorShift) to try to
69	>	* find a free slot.
70	>	*
71	>	* By default, the table is lazily initialized.  Upon first use,
72	>	* the table is set to size 1, and contains a single Adder. The
73	>	* maximum table size is bounded by nearest power of two >= the
74	>	* number of CPUS.  The table size is capped because, when there
75	>	* are more threads than CPUs, supposing that each thread were
76		* bound to a CPU, there would exist a perfect hash function
77		* mapping threads to slots that eliminates collisions. When we
78		* reach capacity, we search for this mapping by randomly varying
#	Line 58 | Line 80 | public class StripedAdder implements Ser
80		* and failures only become known via CAS failures, convergence
81		* will be slow, and because threads are typically not bound to
82		* CPUS forever, may not occur at all. However, despite these
83	<	* limitations, observed contention is typically very low in these
83	>	* limitations, observed contention is typically low in these
84		* cases.
85		*
64	–	* Table entries are of class Adder; a form of AtomicLong padded
65	–	* to reduce cache contention on most processors. Padding is
66	–	* overkill for most Atomics because they are most often
67	–	* irregularly scattered in memory and thus don't interfere much
68	–	* with each other. But Atomic objects residing in arrays will
69	–	* tend to be placed adjacent to each other, and so will most
70	–	* often share cache lines without this precaution.  Except for
71	–	* slot adders[0], Adders are constructed upon first use, which
72	–	* further improves per-thread locality and helps reduce (an
73	–	* already large) footprint.
74	–	*
86		* A single spinlock is used for resizing the table as well as
87	<	* populating slots with new Adders. Upon lock contention, threads
88	<	* try other slots rather than blocking. We guarantee that at
89	<	* least one slot (0) exists, so retries will eventually find a
90	<	* candidate Adder. During these retries, there is increased
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
92		* contention and reduced locality, which is still better than
93		* alternatives.
94		*/
95
96	<	/**
85	<	* Number of processors, to place a cap on table growth.
86	<	*/
87	<	static final int NCPU = Runtime.getRuntime().availableProcessors();
96	>	private static final int NCPU = Runtime.getRuntime().availableProcessors();
97
98		/**
99	<	* Padded version of AtomicLong
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 extends AtomicLong {
104	<	long p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, pa, pb, pc, pd;
105	<	Adder(long x) { super(x); }
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 starts off with
112	<	* a given random value, but may be set to a different
100	<	* pseudo-random value (using a cheaper but adequate xorshift
101	<	* generator) upon collisions.
111	>	* Holder for the thread-local hash code. The code is initially
112	>	* random, but may be set to a different value upon collisions.
113		*/
114		static final class HashCode {
115	+	static final Random rng = new Random();
116		int code;
117	<	HashCode(int h) { code = h; }
117	>	HashCode() {
118	>	int h = rng.nextInt(); // Avoid zero, because of xorShift rehash
119	>	code = (h == 0) ? 1 : h;
120	>	}
121		}
122
123		/**
124		* The corresponding ThreadLocal class
125		*/
126		static final class ThreadHashCode extends ThreadLocal<HashCode> {
127	<	static final Random rng = new Random();
113	<	public HashCode initialValue() {
114	<	int h = rng.nextInt();
115	<	return new HashCode((h == 0) ? 1 : h); // ensure nonzero
116	<	}
127	>	public HashCode initialValue() { return new HashCode(); }
128		}
129
130		/**
131		* Static per-thread hash codes. Shared across all StripedAdders
132	<	* because adjustments due to collisions in one table are likely
133	<	* to be appropriate for others.
132	>	* to reduce ThreadLocal pollution and because adjustments due to
133	>	* collisions in one table are likely to be appropriate for
134	>	* others.
135		*/
136		static final ThreadHashCode threadHashCode = new ThreadHashCode();
137
138		/**
139	<	* Table of adders. Minimum size 2. Size grows to be at most NCPU.
139	>	* Table of adders. When non-null, size is a power of 2.
140		*/
141		private transient volatile Adder[] adders;
142
143		/**
144	<	* Serves as a lock when resizing and/or creating Adders.  There
133	<	* is no need for a blocking lock: When busy, other threads try
134	<	* other slots.
135	<	*/
136	<	private final AtomicInteger mutex;
137	<
138	<	/**
139	<	* Marsaglia XorShift random generator for rehashing on collisions
144	>	* Spinlock (locked via CAS) used when resizing and/or creating Adders.
145		*/
146	<	private static int xorShift(int r) {
142	<	r ^= r << 13;
143	<	r ^= r >>> 17;
144	<	return r ^ (r << 5);
145	<	}
146	>	private volatile int busy;
147
148		/**
149		* Creates a new adder with zero sum.
150		*/
151		public StripedAdder() {
151	–	this(2);
152		}
153
154		/**
#	Line 160 | Line 160 | public class StripedAdder implements Ser
160		*/
161		public StripedAdder(int expectedContention) {
162		int cap = (expectedContention < NCPU) ? expectedContention : NCPU;
163	<	int size = 2;
163	>	int size = 1;
164		while (size < cap)
165		size <<= 1;
166		Adder[] as = new Adder[size];
167	<	as[0] = new Adder(0); // ensure at least one available adder
167	>	for (int i = 0; i < size; ++i)
168	>	as[i] = new Adder(0);
169		this.adders = as;
169	–	this.mutex = new AtomicInteger();
170		}
171
172		/**
#	Line 175 | 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;  // locals to hold volatile reads
179		HashCode hc = threadHashCode.get();
180	<	for (int h = hc.code;;) {
181	<	Adder[] as = adders;
182	<	int n = as.length;
183	<	Adder a = as[h & (n - 1)];
184	<	if (a != null) {
185	<	long v = a.get();
186	<	if (a.compareAndSet(v, v + x))
187	<	break;
187	<	if (n >= NCPU) {                 // Collision when table at max
188	<	h = hc.code = xorShift(h);   // change code
189	<	continue;
190	<	}
191	<	}
192	<	final AtomicInteger mutex = this.mutex;
193	<	if (mutex.get() != 0)
194	<	h = xorShift(h);                 // Try elsewhere
195	<	else if (mutex.compareAndSet(0, 1)) {
196	<	boolean created = false;
197	<	try {
198	<	Adder[] rs = adders;
199	<	if (a != null && rs == as)   // Resize table
200	<	rs = adders = Arrays.copyOf(as, as.length << 1);
201	<	int j = h & (rs.length - 1);
202	<	if (rs[j] == null) {         // Create adder
203	<	rs[j] = new Adder(x);
204	<	created = true;
205	<	}
206	<	} finally {
207	<	mutex.set(0);
208	<	}
209	<	if (created) {
210	<	hc.code = h;                 // Use this adder next time
211	<	break;
212	<	}
213	<	}
180	>	int h = hc.code;
181	>	boolean contended;
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	>	contended = true;
188		}
189	+	else
190	+	contended = false;
191	+	retryAdd(x, hc, contended);
192		}
193
194		/**
195	<	* Returns an estimate of the current sum.  The result is
196	<	* calculated by summing multiple variables, so may not be
197	<	* accurate if updates occur concurrently with this method.
195	>	* Handle cases of add involving initialization, resizing,
196	>	* creating new Adders, and/or contention. See above for
197	>	* explanation. This method suffers the usual non-modularity
198	>	* problems of optimistic retry code, relying on rechecked sets of
199	>	* reads.
200		*
201	<	* @return the estimated sum
202	<	*/
203	<	public long sum() {
225	<	long sum = 0;
226	<	Adder[] as = adders;
227	<	int n = as.length;
228	<	for (int i = 0; i < n; ++i) {
229	<	Adder a = as[i];
230	<	if (a != null)
231	<	sum += a.get();
232	<	}
233	<	return sum;
234	<	}
235	<
236	<	/**
237	<	* Resets each of the variables to zero. This is effective in
238	<	* fully resetting the sum only if there are no concurrent
239	<	* updates.
201	>	* @param x the value to add
202	>	* @param hc the hash code holder
203	>	* @param precontended true if CAS failed before call
204		*/
205	<	public void reset() {
206	<	Adder[] as = adders;
207	<	int n = as.length;
208	<	for (int i = 0; i < n; ++i) {
209	<	Adder a = as[i];
210	<	if (a != null)
211	<	a.set(0L);
205	>	private void retryAdd(long x, HashCode hc, boolean precontended) {
206	>	int h = hc.code;
207	>	boolean collide = false; // true if last slot nonempty
208	>	for (;;) {
209	>	Adder[] as; Adder a; int n;
210	>	if ((as = adders) != null && (n = as.length) > 0) {
211	>	if ((a = as[(n - 1) & h]) == null) {
212	>	if (busy == 0 &&            // Try to attach new Adder
213	>	UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
214	>	boolean created = false;
215	>	try {                   // Recheck under lock
216	>	Adder[] rs; int m, j;
217	>	if ((rs = adders) != null && (m = rs.length) > 0 &&
218	>	rs[j = (m - 1) & h] == null) {
219	>	rs[j] = new Adder(x);
220	>	created = true;
221	>	}
222	>	} finally {
223	>	busy = 0;
224	>	}
225	>	if (created)
226	>	break;
227	>	continue;               // Slot is now non-empty
228	>	}
229	>	collide = false;
230	>	}
231	>	else if (precontended)          // CAS already known to fail
232	>	precontended = false;       // Continue after rehash
233	>	else {
234	>	long v = a.value;
235	>	if (UNSAFE.compareAndSwapLong(a, valueOffset, v, v + x))
236	>	break;
237	>	if (!collide)
238	>	collide = true;
239	>	else if (n >= NCPU || adders != as)
240	>	collide = false;        // Don't expand
241	>	else if (busy == 0 &&
242	>	UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
243	>	collide = false;
244	>	try {
245	>	if (adders == as) { // Expand table
246	>	Adder[] rs = new Adder[n << 1];
247	>	for (int i = 0; i < n; ++i)
248	>	rs[i] = as[i];
249	>	adders = rs;
250	>	}
251	>	} finally {
252	>	busy = 0;
253	>	}
254	>	continue;
255	>	}
256	>	}
257	>	h ^= h << 13;                   // Rehash
258	>	h ^= h >>> 17;
259	>	h ^= h << 5;
260	>	}
261	>	else if (adders == as) {            // Try to default-initialize
262	>	Adder[] rs = new Adder[1];
263	>	rs[0] = new Adder(x);
264	>	boolean init = false;
265	>	while (adders == as) {
266	>	if (UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
267	>	try {
268	>	if (adders == as) {
269	>	adders = rs;
270	>	init = true;
271	>	}
272	>	} finally {
273	>	busy = 0;
274	>	}
275	>	break;
276	>	}
277	>	if (adders != as)
278	>	break;
279	>	Thread.yield();              // Back off
280	>	}
281	>	if (init)
282	>	break;
283	>	}
284		}
285	+	hc.code = h;                            // Record index for next time
286		}
287
288		/**
#	Line 263 | Line 300 | public class StripedAdder implements Ser
300		}
301
302		/**
303	<	* Equivalent to {@link #sum} followed by {@link #reset}.
303	>	* Returns an estimate of the current sum.  The result is
304	>	* calculated by summing multiple variables, so may not be
305	>	* accurate if updates occur concurrently with this method.
306		*
307		* @return the estimated sum
308		*/
309	<	public long sumAndReset() {
310	<	long sum = 0;
309	>	public long sum() {
310	>	long sum = 0L;
311	>	Adder[] as = adders;
312	>	if (as != null) {
313	>	int n = as.length;
314	>	for (int i = 0; i < n; ++i) {
315	>	Adder a = as[i];
316	>	if (a != null)
317	>	sum += a.value;
318	>	}
319	>	}
320	>	return sum;
321	>	}
322	>
323	>	/**
324	>	* Resets each of the variables to zero, returning the estimated
325	>	* previous sum. This is effective in fully resetting the sum only
326	>	* if there are no concurrent updates.
327	>	*
328	>	* @return the estimated previous sum
329	>	*/
330	>	public long reset() {
331	>	long sum = 0L;
332		Adder[] as = adders;
333	<	int n = as.length;
334	<	for (int i = 0; i < n; ++i) {
335	<	Adder a = as[i];
336	<	if (a != null) {
337	<	sum += a.get();
338	<	a.set(0L);
333	>	if (as != null) {
334	>	int n = as.length;
335	>	for (int i = 0; i < n; ++i) {
336	>	Adder a = as[i];
337	>	if (a != null) {
338	>	sum += a.value;
339	>	a.value = 0L;
340	>	}
341		}
342		}
343		return sum;
#	Line 290 | Line 352 | public class StripedAdder implements Ser
352		private void readObject(ObjectInputStream s)
353		throws IOException, ClassNotFoundException {
354		s.defaultReadObject();
355	<	long c = s.readLong();
356	<	Adder[] as = new Adder[2];
295	<	as[0] = new Adder(c);
296	<	this.adders = as;
297	<	mutex.set(0);
355	>	busy = 0;
356	>	add(s.readLong());
357		}
358
359	<	}
359	>	// Unsafe mechanics
360	>	private static final sun.misc.Unsafe UNSAFE;
361	>	private static final long busyOffset;
362	>	private static final long valueOffset;
363	>	static {
364	>	try {
365	>	UNSAFE = getUnsafe();
366	>	Class<?> sk = StripedAdder.class;
367	>	busyOffset = UNSAFE.objectFieldOffset
368	>	(sk.getDeclaredField("busy"));
369	>	Class<?> ak = Adder.class;
370	>	valueOffset = UNSAFE.objectFieldOffset
371	>	(ak.getDeclaredField("value"));
372	>	} catch (Exception e) {
373	>	throw new Error(e);
374	>	}
375	>	}
376
377	+	/**
378	+	* Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
379	+	* Replace with a simple call to Unsafe.getUnsafe when integrating
380	+	* into a jdk.
381	+	*
382	+	* @return a sun.misc.Unsafe
383	+	*/
384	+	private static sun.misc.Unsafe getUnsafe() {
385	+	try {
386	+	return sun.misc.Unsafe.getUnsafe();
387	+	} catch (SecurityException se) {
388	+	try {
389	+	return java.security.AccessController.doPrivileged
390	+	(new java.security
391	+	.PrivilegedExceptionAction<sun.misc.Unsafe>() {
392	+	public sun.misc.Unsafe run() throws Exception {
393	+	java.lang.reflect.Field f = sun.misc
394	+	.Unsafe.class.getDeclaredField("theUnsafe");
395	+	f.setAccessible(true);
396	+	return (sun.misc.Unsafe) f.get(null);
397	+	}});
398	+	} catch (java.security.PrivilegedActionException e) {
399	+	throw new RuntimeException("Could not initialize intrinsics",
400	+	e.getCause());
401	+	}
402	+	}
403	+	}
404
405	+	}

Diff Legend

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