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 1.12 by dl, Sat Jul 30 16:26:34 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 15 | Line 14 | import java.io.ObjectInputStream;
14		import java.io.ObjectOutputStream;
15
16		/**
17	<	* A set of variables that together maintain a sum.  When updates
18	<	* (method {@link #add}) are contended across threads, this set of
19	<	* adder variables may grow dynamically to reduce contention. Method
21	<	* {@link #sum} returns the current combined total across these
22	<	* adders. This value is <em>NOT</em> an atomic snapshot (concurrent
23	<	* updates may occur while the sum is being calculated), and so cannot
24	<	* be used alone for fine-grained synchronization control.
17	>	* One or more variables that together maintain an initially zero sum.
18	>	* When updates (method {@link #add}) are contended across threads,
19	>	* the set of variables may grow dynamically to reduce contention.
20		*
21	<	* <p> This class may be applicable when many threads frequently
22	<	* update a common sum that is used for purposes such as collecting
23	<	* statistics. In this case, performance may be significantly faster
24	<	* than using a shared {@link AtomicLong}, at the expense of using
25	<	* significantly more space.  On the other hand, if it is known that
26	<	* only one thread can ever update the sum, performance may be
27	<	* significantly slower than just updating a local variable.
21	>	* <p> This class is usually preferable to {@link AtomicLong} when
22	>	* multiple threads update a common sum that is used for purposes such
23	>	* as collecting statistics, not for fine-grained synchronization
24	>	* control.  Under high update contention, throughput of this class is
25	>	* expected to be significantly higher, at the expense of higher space
26	>	* consumption.  Under low contention, this class imposes very little
27	>	* time and space overhead compared to AtomicLong.  On the other hand,
28	>	* in contexts where it is statically known that only one thread can
29	>	* ever update a sum, time and space overhead is noticeably greater
30	>	* than just updating a local variable.
31		*
32	<	* <p>A StripedAdder may optionally be constructed with a given
33	<	* expected contention level; i.e., the number of threads that are
34	<	* expected to concurrently update the sum. Supplying an accurate
35	<	* value may improve performance by reducing the need for dynamic
36	<	* adjustment.
32	>	* <p> Method {@link #sum} returns the current combined total across
33	>	* the variables maintaining the sum.  This value is <em>NOT</em> an
34	>	* atomic snapshot: Concurrent updates may occur while the sum is
35	>	* being calculated.  However, updates cannot be "lost", so invocation
36	>	* of <code>sum</code> in the absence of concurrent updates always
37	>	* returns an accurate result.  The sum may also be <code>reset</code>
38	>	* to zero, as an alternative to creating a new adder.  However,
39	>	* method {@link #reset} is intrinsically racy, so should only be used
40	>	* when it is known that no threads are concurrently updating the sum.
41	>	*
42	>	* <p><em>jsr166e note: This class is targeted to be placed in
43	>	* java.util.concurrent.atomic<em>
44		*
45		* @author Doug Lea
46		*/
#	Line 43 | Line 48 | public class StripedAdder implements Ser
48		private static final long serialVersionUID = 7249069246863182397L;
49
50		/*
51	<	* Overview: We maintain a table of Atomic long variables. The
52	<	* table is indexed by per-thread hash codes that are initialized
53	<	* to random values.
54	<	*
50	<	* The table doubles in size upon contention (as indicated by
51	<	* failed CASes when performing add()), but is capped at the
52	<	* nearest power of two >= #CPUS. This reflects the idea that,
53	<	* when there are more threads than CPUs, then if each thread were
54	<	* bound to a CPU, there would exist a perfect hash function
55	<	* mapping threads to slots that eliminates collisions. When we
56	<	* reach capacity, we search for this mapping by randomly varying
57	<	* the hash codes of colliding threads.  Because search is random,
58	<	* and failures only become known via CAS failures, convergence
59	<	* will be slow, and because threads are typically not bound to
60	<	* CPUS forever, may not occur at all. However, despite these
61	<	* limitations, observed contention is typically very low in these
62	<	* cases.
51	>	* A StripedAdder maintains a lazily-initialized table of
52	>	* atomically updated variables, plus an extra "base" field. The
53	>	* table size is a power of two. Indexing uses masked per-thread
54	>	* hash codes
55		*
56	<	* Table entries are of class Adder; a form of AtomicLong padded
56	>	* Table entries are of class Cell; a variant of AtomicLong padded
57		* to reduce cache contention on most processors. Padding is
58	<	* overkill for most Atomics because they are most often
59	<	* irregularly scattered in memory and thus don't interfere much
60	<	* with each other. But Atomic objects residing in arrays will
61	<	* tend to be placed adjacent to each other, and so will most
62	<	* often share cache lines without this precaution.  Except for
63	<	* slot adders[0], Adders are constructed upon first use, which
64	<	* further improves per-thread locality and helps reduce (an
65	<	* already large) footprint.
58	>	* overkill for most Atomics because they are usually irregularly
59	>	* scattered in memory and thus don't interfere much with each
60	>	* other. But Atomic objects residing in arrays will tend to be
61	>	* placed adjacent to each other, and so will most often share
62	>	* cache lines (with a huge negative performance impact) without
63	>	* this precaution.
64	>	*
65	>	* In part because Cells are relatively large, we avoid creating
66	>	* them until they are needed.  When there is no contention, all
67	>	* updates are made to the base field.  Upon first contention (a
68	>	* failed CAS on base update), the table is initialized to size 2.
69	>	* The table size is doubled upon further contention until
70	>	* reaching the nearest power of two greater than or equal to the
71	>	* number of CPUS.
72	>	*
73	>	* Per-thread hash codes are initialized to random values.
74	>	* Contention and/or table collisions are indicated by failed
75	>	* CASes when performing an add operation (see method
76	>	* retryAdd). Upon a collision, if the table size is less than the
77	>	* capacity, it is doubled in size unless some other thread holds
78	>	* the lock. If a hashed slot is empty, and lock is available, a
79	>	* new Cell is created. Otherwise, if the slot exists, a CAS is
80	>	* tried.  Retries proceed by "double hashing", using a secondary
81	>	* hash (Marsaglia XorShift) to try to find a free slot.
82	>	*
83	>	* The table size is capped because, when there are more threads
84	>	* than CPUs, supposing that each thread were bound to a CPU,
85	>	* there would exist a perfect hash function mapping threads to
86	>	* slots that eliminates collisions. When we reach capacity, we
87	>	* search for this mapping by randomly varying the hash codes of
88	>	* colliding threads.  Because search is random, and collisions
89	>	* only become known via CAS failures, convergence can be slow,
90	>	* and because threads are typically not bound to CPUS forever,
91	>	* may not occur at all. However, despite these limitations,
92	>	* observed contention rates are typically low in these cases.
93	>	*
94	>	* A single spinlock is used for initializing and resizing the
95	>	* table, as well as populating slots with new Cells.  There is no
96	>	* need for a blocking lock: Upon lock contention, threads try
97	>	* other slots (or the base) rather than blocking.  During these
98	>	* retries, there is increased contention and reduced locality,
99	>	* which is still better than alternatives.
100	>	*
101	>	* It is possible for a Cell to become unused when threads that
102	>	* once hashed to it terminate, as well as in the case where
103	>	* doubling the table causes no thread to hash to it under
104	>	* expanded mask.  We do not try to detect or remove such cells,
105	>	* under the assumption that for long-running adders, observed
106	>	* contention levels will recur, so the cells will eventually be
107	>	* needed again; and for short-lived ones, it does not matter.
108		*
75	–	* A single spinlock is used for resizing the table as well as
76	–	* populating slots with new Adders. Upon lock contention, threads
77	–	* try other slots rather than blocking. We guarantee that at
78	–	* least one slot (0) exists, so retries will eventually find a
79	–	* candidate Adder. During these retries, there is increased
80	–	* contention and reduced locality, which is still better than
81	–	* alternatives.
109		*/
110
111	<	/**
85	<	* Number of processors, to place a cap on table growth.
86	<	*/
87	<	static final int NCPU = Runtime.getRuntime().availableProcessors();
111	>	private static final int NCPU = Runtime.getRuntime().availableProcessors();
112
113		/**
114	<	* Padded version of AtomicLong
114	>	* Padded variant of AtomicLong.  The value field is placed
115	>	* between pads, hoping that the JVM doesn't reorder them.
116	>	* Updates are via inlined CAS in methods add and retryAdd.
117		*/
118	<	static final class Adder extends AtomicLong {
119	<	long p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, pa, pb, pc, pd;
120	<	Adder(long x) { super(x); }
118	>	static final class Cell {
119	>	volatile long p0, p1, p2, p3, p4, p5, p6;
120	>	volatile long value;
121	>	volatile long q0, q1, q2, q3, q4, q5, q6;
122	>	Cell(long x) { value = x; }
123		}
124
125		/**
126	<	* Holder for the thread-local hash code. The code starts off with
127	<	* 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.
126	>	* Holder for the thread-local hash code. The code is initially
127	>	* random, but may be set to a different value upon collisions.
128		*/
129		static final class HashCode {
130	+	static final Random rng = new Random();
131		int code;
132	<	HashCode(int h) { code = h; }
132	>	HashCode() {
133	>	int h = rng.nextInt(); // Avoid zero to allow xorShift rehash
134	>	code = (h == 0) ? 1 : h;
135	>	}
136		}
137
138		/**
139		* The corresponding ThreadLocal class
140		*/
141		static final class ThreadHashCode extends ThreadLocal<HashCode> {
142	<	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	<	}
142	>	public HashCode initialValue() { return new HashCode(); }
143		}
144
145		/**
146		* Static per-thread hash codes. Shared across all StripedAdders
147	<	* because adjustments due to collisions in one table are likely
148	<	* to be appropriate for others.
147	>	* to reduce ThreadLocal pollution and because adjustments due to
148	>	* collisions in one table are likely to be appropriate for
149	>	* others.
150		*/
151		static final ThreadHashCode threadHashCode = new ThreadHashCode();
152
153		/**
154	<	* Table of adders. Minimum size 2. Size grows to be at most NCPU.
154	>	* Table of cells. When non-null, size is a power of 2.
155		*/
156	<	private transient volatile Adder[] adders;
156	>	private transient volatile Cell[] cells;
157
158		/**
159	<	* Serves as a lock when resizing and/or creating Adders.  There
160	<	* is no need for a blocking lock: When busy, other threads try
134	<	* other slots.
159	>	* Base sum, used mainly when there is no contention, but also as
160	>	* a fallback during table initializion races. Updated via CAS.
161		*/
162	<	private final AtomicInteger mutex;
162	>	private transient volatile long base;
163
164		/**
165	<	* Marsaglia XorShift random generator for rehashing on collisions
165	>	* Spinlock (locked via CAS) used when resizing and/or creating Cells.
166		*/
167	<	private static int xorShift(int r) {
142	<	r ^= r << 13;
143	<	r ^= r >>> 17;
144	<	return r ^ (r << 5);
145	<	}
167	>	private transient volatile int busy;
168
169		/**
170	<	* Creates a new adder with zero sum.
170	>	* Creates a new adder with initial sum of zero.
171		*/
172		public StripedAdder() {
151	–	this(2);
173		}
174
175		/**
176	<	* Creates a new adder with zero sum, and with stripes presized
156	<	* for the given expected contention level.
176	>	* Adds the given value.
177		*
178	<	* @param expectedContention the expected number of threads that
159	<	* will concurrently update the sum.
178	>	* @param x the value to add
179		*/
180	<	public StripedAdder(int expectedContention) {
181	<	int cap = (expectedContention < NCPU) ? expectedContention : NCPU;
182	<	int size = 2;
183	<	while (size < cap)
184	<	size <<= 1;
185	<	Adder[] as = new Adder[size];
186	<	as[0] = new Adder(0); // ensure at least one available adder
187	<	this.adders = as;
188	<	this.mutex = new AtomicInteger();
180	>	public void add(long x) {
181	>	Cell[] as; long v; HashCode hc; Cell a; int n; boolean contended;
182	>	if ((as = cells) != null ||
183	>	!UNSAFE.compareAndSwapLong(this, baseOffset, v = base, v + x)) {
184	>	int h = (hc = threadHashCode.get()).code;
185	>	if (as != null && (n = as.length) > 0 &&
186	>	(a = as[(n - 1) & h]) != null) {
187	>	if (UNSAFE.compareAndSwapLong(a, valueOffset,
188	>	v = a.value, v + x))
189	>	return;
190	>	contended = true;
191	>	}
192	>	else
193	>	contended = false;
194	>	retryAdd(x, hc, contended);
195	>	}
196		}
197
198		/**
199	<	* Adds the given value.
199	>	* Handle cases of add involving initialization, resizing,
200	>	* creating new Cells, and/or contention. See above for
201	>	* explanation. This method suffers the usual non-modularity
202	>	* problems of optimistic retry code, relying on rechecked sets of
203	>	* reads.
204		*
205		* @param x the value to add
206	+	* @param hc the hash code holder
207	+	* @param precontended true if CAS failed before call
208		*/
209	<	public void add(long x) {
210	<	HashCode hc = threadHashCode.get();
211	<	for (int h = hc.code;;) {
212	<	Adder[] as = adders;
213	<	int n = as.length;
214	<	Adder a = as[h & (n - 1)];
215	<	if (a != null) {
216	<	long v = a.get();
217	<	if (a.compareAndSet(v, v + x))
218	<	break;
219	<	if (n >= NCPU) {                 // Collision when table at max
220	<	h = hc.code = xorShift(h);   // change code
221	<	continue;
209	>	private void retryAdd(long x, HashCode hc, boolean precontended) {
210	>	int h = hc.code;
211	>	boolean collide = false; // true if last slot nonempty
212	>	for (;;) {
213	>	Cell[] as; Cell a; int n;
214	>	if ((as = cells) != null && (n = as.length) > 0) {
215	>	if ((a = as[(n - 1) & h]) == null) {
216	>	if (busy == 0) {            // Try to attach new Cell
217	>	Cell r = new Cell(x);   // Optimistically create
218	>	if (busy == 0 &&
219	>	UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
220	>	boolean created = false;
221	>	try {               // Recheck under lock
222	>	Cell[] rs; int m, j;
223	>	if ((rs = cells) != null &&
224	>	(m = rs.length) > 0 &&
225	>	rs[j = (m - 1) & h] == null) {
226	>	rs[j] = r;
227	>	created = true;
228	>	}
229	>	} finally {
230	>	busy = 0;
231	>	}
232	>	if (created)
233	>	break;
234	>	continue;           // Slot is now non-empty
235	>	}
236	>	}
237	>	collide = false;
238	>	}
239	>	else if (precontended)          // CAS already known to fail
240	>	precontended = false;       // Continue after rehash
241	>	else {
242	>	long v = a.value;
243	>	if (UNSAFE.compareAndSwapLong(a, valueOffset, v, v + x))
244	>	break;
245	>	if (!collide)
246	>	collide = true;
247	>	else if (n >= NCPU || cells != as)
248	>	collide = false;        // Can't expand
249	>	else if (busy == 0 &&
250	>	UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
251	>	collide = false;
252	>	try {
253	>	if (cells == as) { // Expand table
254	>	Cell[] rs = new Cell[n << 1];
255	>	for (int i = 0; i < n; ++i)
256	>	rs[i] = as[i];
257	>	cells = rs;
258	>	}
259	>	} finally {
260	>	busy = 0;
261	>	}
262	>	continue;
263	>	}
264		}
265	+	h ^= h << 13;                   // Rehash
266	+	h ^= h >>> 17;
267	+	h ^= h << 5;
268		}
269	<	final AtomicInteger mutex = this.mutex;
270	<	if (mutex.get() != 0)
271	<	h = xorShift(h);                 // Try elsewhere
272	<	else if (mutex.compareAndSet(0, 1)) {
273	<	boolean created = false;
274	<	try {
275	<	Adder[] rs = adders;
276	<	if (a != null && rs == as)   // Resize table
277	<	rs = adders = Arrays.copyOf(as, as.length << 1);
278	<	int j = h & (rs.length - 1);
202	<	if (rs[j] == null) {         // Create adder
203	<	rs[j] = new Adder(x);
204	<	created = true;
269	>	else if (busy == 0 && cells == as &&
270	>	UNSAFE.compareAndSwapInt(this, busyOffset, 0, 1)) {
271	>	boolean init = false;
272	>	try {                           // Initialize
273	>	if (cells == as) {
274	>	Cell r = new Cell(x);
275	>	Cell[] rs = new Cell[2];
276	>	rs[h & 1] = r;
277	>	cells = rs;
278	>	init = true;
279		}
280		} finally {
281	<	mutex.set(0);
281	>	busy = 0;
282		}
283	<	if (created) {
284	<	hc.code = h;                 // Use this adder next time
283	>	if (init)
284	>	break;
285	>	}
286	>	else {                              // Lost initialization race
287	>	long b = base;                  // Fall back on using base
288	>	if (UNSAFE.compareAndSwapLong(this, baseOffset, b, b + x))
289		break;
212	–	}
290		}
291		}
292	+	hc.code = h;                            // Record index for next time
293		}
294
295		/**
296	<	* Returns an estimate of the current sum.  The result is
219	<	* calculated by summing multiple variables, so may not be
220	<	* accurate if updates occur concurrently with this method.
221	<	*
222	<	* @return the estimated sum
296	>	* Equivalent to {@code add(1)}.
297		*/
298	<	public long sum() {
299	<	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;
298	>	public void increment() {
299	>	add(1L);
300		}
301
302		/**
303	<	* Resets each of the variables to zero. This is effective in
238	<	* fully resetting the sum only if there are no concurrent
239	<	* updates.
303	>	* Equivalent to {@code add(-1)}.
304		*/
305	<	public void reset() {
306	<	Adder[] as = adders;
243	<	int n = as.length;
244	<	for (int i = 0; i < n; ++i) {
245	<	Adder a = as[i];
246	<	if (a != null)
247	<	a.set(0L);
248	<	}
305	>	public void decrement() {
306	>	add(-1L);
307		}
308
309		/**
310	<	* Equivalent to {@code add(1)}.
310	>	* Returns the current sum.  The result is only guaranteed to be
311	>	* accurate in the absence of concurrent updates. Otherwise, it
312	>	* may fail to reflect one or more updates occuring while
313	>	* calculating the result.
314	>	*
315	>	* @return the sum
316		*/
317	<	public void increment() {
318	<	add(1L);
317	>	public long sum() {
318	>	Cell[] as = cells;
319	>	long sum = base;
320	>	if (as != null) {
321	>	int n = as.length;
322	>	for (int i = 0; i < n; ++i) {
323	>	Cell a = as[i];
324	>	if (a != null)
325	>	sum += a.value;
326	>	}
327	>	}
328	>	return sum;
329		}
330
331		/**
332	<	* Equivalent to {@code add(-1)}.
332	>	* Resets variables maintaining the sum to zero.  This is
333	>	* effective in setting the sum to zero only if there are no
334	>	* concurrent updates.
335		*/
336	<	public void decrement() {
337	<	add(-1L);
336	>	public void reset() {
337	>	Cell[] as = cells;
338	>	base = 0L;
339	>	if (as != null) {
340	>	int n = as.length;
341	>	for (int i = 0; i < n; ++i) {
342	>	Cell a = as[i];
343	>	if (a != null)
344	>	a.value = 0L;
345	>	}
346	>	}
347		}
348
349		/**
350	<	* Equivalent to {@link #sum} followed by {@link #reset}.
350	>	* Equivalent in effect to {@link #sum} followed by {@link
351	>	* #reset}. This method may apply for example during quiescent
352	>	* points between multithreaded computations.  If there are
353	>	* updates concurrent with this method, the returned value is
354	>	* <em>not</em> guaranteed to be the final sum occurring before
355	>	* the reset.
356		*
357	<	* @return the estimated sum
357	>	* @return the sum
358		*/
359	<	public long sumAndReset() {
360	<	long sum = 0;
361	<	Adder[] as = adders;
362	<	int n = as.length;
363	<	for (int i = 0; i < n; ++i) {
364	<	Adder a = as[i];
365	<	if (a != null) {
366	<	sum += a.get();
367	<	a.set(0L);
359	>	public long sumThenReset() {
360	>	Cell[] as = cells;
361	>	long sum = base;
362	>	base = 0L;
363	>	if (as != null) {
364	>	int n = as.length;
365	>	for (int i = 0; i < n; ++i) {
366	>	Cell a = as[i];
367	>	if (a != null) {
368	>	sum += a.value;
369	>	a.value = 0L;
370	>	}
371		}
372		}
373		return sum;
#	Line 290 | Line 382 | public class StripedAdder implements Ser
382		private void readObject(ObjectInputStream s)
383		throws IOException, ClassNotFoundException {
384		s.defaultReadObject();
385	<	long c = s.readLong();
386	<	Adder[] as = new Adder[2];
387	<	as[0] = new Adder(c);
388	<	this.adders = as;
389	<	mutex.set(0);
385	>	busy = 0;
386	>	cells = null;
387	>	base = s.readLong();
388	>	}
389	>
390	>	// Unsafe mechanics
391	>	private static final sun.misc.Unsafe UNSAFE;
392	>	private static final long baseOffset;
393	>	private static final long busyOffset;
394	>	private static final long valueOffset;
395	>	static {
396	>	try {
397	>	UNSAFE = getUnsafe();
398	>	Class<?> sk = StripedAdder.class;
399	>	baseOffset = UNSAFE.objectFieldOffset
400	>	(sk.getDeclaredField("base"));
401	>	busyOffset = UNSAFE.objectFieldOffset
402	>	(sk.getDeclaredField("busy"));
403	>	Class<?> ak = Cell.class;
404	>	valueOffset = UNSAFE.objectFieldOffset
405	>	(ak.getDeclaredField("value"));
406	>	} catch (Exception e) {
407	>	throw new Error(e);
408	>	}
409		}
410
411	<	}
412	<
411	>	/**
412	>	* Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
413	>	* Replace with a simple call to Unsafe.getUnsafe when integrating
414	>	* into a jdk.
415	>	*
416	>	* @return a sun.misc.Unsafe
417	>	*/
418	>	private static sun.misc.Unsafe getUnsafe() {
419	>	try {
420	>	return sun.misc.Unsafe.getUnsafe();
421	>	} catch (SecurityException se) {
422	>	try {
423	>	return java.security.AccessController.doPrivileged
424	>	(new java.security
425	>	.PrivilegedExceptionAction<sun.misc.Unsafe>() {
426	>	public sun.misc.Unsafe run() throws Exception {
427	>	java.lang.reflect.Field f = sun.misc
428	>	.Unsafe.class.getDeclaredField("theUnsafe");
429	>	f.setAccessible(true);
430	>	return (sun.misc.Unsafe) f.get(null);
431	>	}});
432	>	} catch (java.security.PrivilegedActionException e) {
433	>	throw new RuntimeException("Could not initialize intrinsics",
434	>	e.getCause());
435	>	}
436	>	}
437	>	}
438
439	+	}

Diff Legend

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