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; |
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 |
> |
* much more space. On the other hand, if it is known that only one |
30 |
> |
* thread can ever update the sum, performance may be significantly |
31 |
> |
* slower 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 |
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 |
48 |
< |
* to random values. |
45 |
> |
* A StripedAdder maintains a table of Atomic long variables. The |
46 |
> |
* table is indexed by per-thread hash codes. |
47 |
|
* |
48 |
< |
* The table doubles in size upon contention (as indicated by |
49 |
< |
* failed CASes when performing add()), but is capped at the |
50 |
< |
* 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. |
63 |
< |
* |
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 |
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 without this precaution. Except for |
55 |
< |
* slot adders[0], Adders are constructed upon first use, which |
56 |
< |
* further improves per-thread locality and helps reduce (an |
57 |
< |
* already large) footprint. |
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 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 |
|
* |
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, at least 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. |
144 |
> |
* Spinlock (locked via CAS) used when resizing and/or creating Adders. |
145 |
|
*/ |
146 |
< |
private final AtomicInteger mutex; |
137 |
< |
|
138 |
< |
/** |
139 |
< |
* Marsaglia XorShift random generator for rehashing on collisions |
140 |
< |
*/ |
141 |
< |
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 |
|
/** |
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 |
|
/** |
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; |
188 |
< |
if (n >= NCPU) { // Collision when table at max |
189 |
< |
h = hc.code = xorShift(h); // change code |
190 |
< |
continue; |
180 |
> |
int h = hc.code; |
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. 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, boolean collide) { |
202 |
> |
int h = hc.code; |
203 |
> |
for (;;) { |
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 |
< |
} |
232 |
< |
final AtomicInteger mutex = this.mutex; |
233 |
< |
if (mutex.get() != 0) |
234 |
< |
h = xorShift(h); // Try elsewhere |
235 |
< |
else if (mutex.compareAndSet(0, 1)) { |
236 |
< |
boolean created = false; |
237 |
< |
try { |
238 |
< |
Adder[] rs = adders; |
239 |
< |
if (a != null && rs == as) // Resize table |
240 |
< |
rs = adders = Arrays.copyOf(as, as.length << 1); |
241 |
< |
int j = h & (rs.length - 1); |
242 |
< |
if (rs[j] == null) { // Create adder |
243 |
< |
rs[j] = new Adder(x); |
244 |
< |
created = true; |
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 |
< |
} finally { |
207 |
< |
mutex.set(0); |
249 |
> |
collide = false; |
250 |
|
} |
251 |
< |
if (created) { |
252 |
< |
hc.code = h; // Use this adder next time |
253 |
< |
break; |
251 |
> |
h ^= h << 13; // Rehash |
252 |
> |
h ^= h >>> 17; |
253 |
> |
h ^= h << 5; |
254 |
> |
} |
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) { |
264 |
> |
adders = rs; |
265 |
> |
init = true; |
266 |
> |
} |
267 |
> |
} finally { |
268 |
> |
busy = 0; |
269 |
> |
} |
270 |
> |
if (init) |
271 |
> |
break; |
272 |
|
} |
273 |
|
} |
274 |
+ |
else if (adders == as) // Lost initialization race |
275 |
+ |
Thread.yield(); |
276 |
|
} |
277 |
< |
} |
216 |
< |
|
217 |
< |
/** |
218 |
< |
* 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 |
223 |
< |
*/ |
224 |
< |
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. |
240 |
< |
*/ |
241 |
< |
public void reset() { |
242 |
< |
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 |
< |
} |
277 |
> |
hc.code = h; // Record index for next time |
278 |
|
} |
279 |
|
|
280 |
|
/** |
292 |
|
} |
293 |
|
|
294 |
|
/** |
295 |
< |
* Equivalent to {@link #sum} followed by {@link #reset}. |
295 |
> |
* Returns an estimate of the current sum. The result is |
296 |
> |
* calculated by summing multiple variables, so may not be |
297 |
> |
* accurate if updates occur concurrently with this method. |
298 |
|
* |
299 |
|
* @return the estimated sum |
300 |
|
*/ |
301 |
< |
public long sumAndReset() { |
302 |
< |
long sum = 0; |
301 |
> |
public long sum() { |
302 |
> |
long sum = 0L; |
303 |
> |
Adder[] as = adders; |
304 |
> |
if (as != null) { |
305 |
> |
int n = as.length; |
306 |
> |
for (int i = 0; i < n; ++i) { |
307 |
> |
Adder a = as[i]; |
308 |
> |
if (a != null) |
309 |
> |
sum += a.value; |
310 |
> |
} |
311 |
> |
} |
312 |
> |
return sum; |
313 |
> |
} |
314 |
> |
|
315 |
> |
/** |
316 |
> |
* Resets each of the variables to zero, returning the estimated |
317 |
> |
* previous sum. This is effective in fully resetting the sum only |
318 |
> |
* if there are no concurrent updates. |
319 |
> |
* |
320 |
> |
* @return the estimated previous sum |
321 |
> |
*/ |
322 |
> |
public long reset() { |
323 |
> |
long sum = 0L; |
324 |
|
Adder[] as = adders; |
325 |
< |
int n = as.length; |
326 |
< |
for (int i = 0; i < n; ++i) { |
327 |
< |
Adder a = as[i]; |
328 |
< |
if (a != null) { |
329 |
< |
sum += a.get(); |
330 |
< |
a.set(0L); |
325 |
> |
if (as != null) { |
326 |
> |
int n = as.length; |
327 |
> |
for (int i = 0; i < n; ++i) { |
328 |
> |
Adder a = as[i]; |
329 |
> |
if (a != null) { |
330 |
> |
sum += a.value; |
331 |
> |
a.value = 0L; |
332 |
> |
} |
333 |
|
} |
334 |
|
} |
335 |
|
return sum; |
344 |
|
private void readObject(ObjectInputStream s) |
345 |
|
throws IOException, ClassNotFoundException { |
346 |
|
s.defaultReadObject(); |
347 |
< |
long c = s.readLong(); |
348 |
< |
Adder[] as = new Adder[2]; |
295 |
< |
as[0] = new Adder(c); |
296 |
< |
this.adders = as; |
297 |
< |
mutex.set(0); |
347 |
> |
busy = 0; |
348 |
> |
add(s.readLong()); |
349 |
|
} |
350 |
|
|
351 |
< |
} |
351 |
> |
// Unsafe mechanics |
352 |
> |
private static final sun.misc.Unsafe UNSAFE; |
353 |
> |
private static final long busyOffset; |
354 |
> |
private static final long valueOffset; |
355 |
> |
static { |
356 |
> |
try { |
357 |
> |
UNSAFE = getUnsafe(); |
358 |
> |
Class<?> sk = StripedAdder.class; |
359 |
> |
busyOffset = UNSAFE.objectFieldOffset |
360 |
> |
(sk.getDeclaredField("busy")); |
361 |
> |
Class<?> ak = Adder.class; |
362 |
> |
valueOffset = UNSAFE.objectFieldOffset |
363 |
> |
(ak.getDeclaredField("value")); |
364 |
> |
} catch (Exception e) { |
365 |
> |
throw new Error(e); |
366 |
> |
} |
367 |
> |
} |
368 |
|
|
369 |
+ |
/** |
370 |
+ |
* Returns a sun.misc.Unsafe. Suitable for use in a 3rd party package. |
371 |
+ |
* Replace with a simple call to Unsafe.getUnsafe when integrating |
372 |
+ |
* into a jdk. |
373 |
+ |
* |
374 |
+ |
* @return a sun.misc.Unsafe |
375 |
+ |
*/ |
376 |
+ |
private static sun.misc.Unsafe getUnsafe() { |
377 |
+ |
try { |
378 |
+ |
return sun.misc.Unsafe.getUnsafe(); |
379 |
+ |
} catch (SecurityException se) { |
380 |
+ |
try { |
381 |
+ |
return java.security.AccessController.doPrivileged |
382 |
+ |
(new java.security |
383 |
+ |
.PrivilegedExceptionAction<sun.misc.Unsafe>() { |
384 |
+ |
public sun.misc.Unsafe run() throws Exception { |
385 |
+ |
java.lang.reflect.Field f = sun.misc |
386 |
+ |
.Unsafe.class.getDeclaredField("theUnsafe"); |
387 |
+ |
f.setAccessible(true); |
388 |
+ |
return (sun.misc.Unsafe) f.get(null); |
389 |
+ |
}}); |
390 |
+ |
} catch (java.security.PrivilegedActionException e) { |
391 |
+ |
throw new RuntimeException("Could not initialize intrinsics", |
392 |
+ |
e.getCause()); |
393 |
+ |
} |
394 |
+ |
} |
395 |
+ |
} |
396 |
|
|
397 |
+ |
} |