16 |
|
|
17 |
|
/** |
18 |
|
* A set of variables that together maintain a sum. When updates |
19 |
< |
* (method {@link #add}) are contended across threads, the set of |
20 |
< |
* adders may grow to reduce contention. Method {@link #sum} returns |
21 |
< |
* the current combined total across these adders. This value is |
22 |
< |
* <em>NOT</em> an atomic snapshot (concurrent updates may occur while |
23 |
< |
* the sum is being calculated), and so cannot be used alone for |
24 |
< |
* fine-grained synchronization control. |
19 |
> |
* (method {@link #add}) are contended across threads, this set of |
20 |
> |
* 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. |
25 |
|
* |
26 |
|
* <p> This class may be applicable when many threads frequently |
27 |
|
* update a common sum that is used for purposes such as collecting |
28 |
|
* statistics. In this case, performance may be significantly faster |
29 |
|
* than using a shared {@link AtomicLong}, at the expense of using |
30 |
< |
* significantly more space. On the other hand, if it is known that |
31 |
< |
* only one thread can ever update the sum, performance may be |
32 |
< |
* significantly slower than just updating a local variable. |
30 |
> |
* much more space. On the other hand, if it is known that only one |
31 |
> |
* thread can ever update the sum, performance may be significantly |
32 |
> |
* slower than just updating a local variable. |
33 |
> |
* |
34 |
> |
* <p>A StripedAdder may optionally be constructed with a given |
35 |
> |
* expected contention level; i.e., the number of threads that are |
36 |
> |
* expected to concurrently update the sum. Supplying an accurate |
37 |
> |
* value may improve performance by reducing the need for dynamic |
38 |
> |
* adjustment. |
39 |
|
* |
40 |
|
* @author Doug Lea |
41 |
|
*/ |
43 |
|
private static final long serialVersionUID = 7249069246863182397L; |
44 |
|
|
45 |
|
/* |
46 |
< |
* Overview: We maintain a table of AtomicLongs (padded to reduce |
47 |
< |
* false sharing). The table is indexed by per-thread hash codes |
48 |
< |
* that are initialized as random values. The table doubles in |
49 |
< |
* size upon contention (as indicated by failed CASes when |
50 |
< |
* performing add()), but is capped at the nearest power of two >= |
51 |
< |
* #cpus: At that point, contention should be infrequent if each |
52 |
< |
* thread has a unique index; so we instead adjust hash codes to |
53 |
< |
* new random values upon contention rather than expanding. A |
54 |
< |
* single spinlock is used for resizing the table as well as |
46 |
> |
* A StripedAdder maintains a table of Atomic long variables. The |
47 |
> |
* table is indexed by per-thread hash codes. |
48 |
> |
* |
49 |
> |
* By default, the table is lazily initialized, to minimize |
50 |
> |
* footprint until adders are used. On first use, the table is set |
51 |
> |
* to size DEFAULT_INITIAL_SIZE (currently 8). Table size is |
52 |
> |
* bounded by the number of CPUS (if larger than the default |
53 |
> |
* size). |
54 |
> |
* |
55 |
> |
* Per-thread hash codes are initialized to random values. |
56 |
> |
* Collisions are indicated by failed CASes when performing an add |
57 |
> |
* operation (see method retryAdd). Upon a collision, if the table |
58 |
> |
* size is less than the capacity, it is doubled in size unless |
59 |
> |
* some other thread holds lock. If a hashed slot is empty, and |
60 |
> |
* lock is available, a new Adder is created. Otherwise, if the |
61 |
> |
* slot exists, a CAS is tried. Retries proceed by "double |
62 |
> |
* hashing", using a secondary hash (Marsaglia XorShift) to try to |
63 |
> |
* find a free slot. |
64 |
> |
* |
65 |
> |
* The table size is capped because, when there are more threads |
66 |
> |
* than CPUs, supposing that each thread were bound to a CPU, |
67 |
> |
* there would exist a perfect hash function mapping threads to |
68 |
> |
* slots that eliminates collisions. When we reach capacity, we |
69 |
> |
* search for this mapping by randomly varying the hash codes of |
70 |
> |
* colliding threads. Because search is random, and failures only |
71 |
> |
* become known via CAS failures, convergence will be slow, and |
72 |
> |
* because threads are typically not bound to CPUS forever, may |
73 |
> |
* not occur at all. However, despite these limitations, observed |
74 |
> |
* contention is typically low in these cases. |
75 |
> |
* |
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 |
88 |
< |
* just try other slots rather than blocking. We guarantee that at |
88 |
> |
* try other slots rather than blocking. After initialization, at |
89 |
|
* least one slot exists, so retries will eventually find a |
90 |
< |
* candidate Adder. |
90 |
> |
* candidate Adder. During these retries, there is increased |
91 |
> |
* contention and reduced locality, which is still better than |
92 |
> |
* alternatives. |
93 |
|
*/ |
94 |
|
|
95 |
|
/** |
96 |
< |
* Number of processors, to place a cap on table growth. |
57 |
< |
*/ |
58 |
< |
static final int NCPU = Runtime.getRuntime().availableProcessors(); |
59 |
< |
|
60 |
< |
/** |
61 |
< |
* Version of AtomicLong padded to avoid sharing cache |
62 |
< |
* lines on most processors |
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; |
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 |
|
|
103 |
+ |
private static final int NCPU = Runtime.getRuntime().availableProcessors(); |
104 |
+ |
|
105 |
|
/** |
106 |
< |
* Holder for the thread-local hash code. |
106 |
> |
* Table bounds. DEFAULT_INITIAL_SIZE is the table size set upon |
107 |
> |
* first use under default constructor, and must be a power of |
108 |
> |
* two. There is not much point in making size a lot smaller than |
109 |
> |
* that of Adders though. CAP is the maximum allowed table size. |
110 |
> |
*/ |
111 |
> |
private static final int DEFAULT_INITIAL_SIZE = 8; |
112 |
> |
private static final int CAP = Math.max(NCPU, DEFAULT_INITIAL_SIZE); |
113 |
> |
|
114 |
> |
/** |
115 |
> |
* Holder for the thread-local hash code. The code is initially |
116 |
> |
* random, but may be set to a different value upon collisions. |
117 |
|
*/ |
118 |
|
static final class HashCode { |
119 |
+ |
static final Random rng = new Random(); |
120 |
|
int code; |
121 |
< |
HashCode(int h) { code = h; } |
121 |
> |
HashCode() { |
122 |
> |
int h = rng.nextInt(); |
123 |
> |
code = (h == 0) ? 1 : h; // ensure nonzero |
124 |
> |
} |
125 |
|
} |
126 |
|
|
127 |
|
/** |
128 |
|
* The corresponding ThreadLocal class |
129 |
|
*/ |
130 |
|
static final class ThreadHashCode extends ThreadLocal<HashCode> { |
131 |
< |
static final Random rng = new Random(); |
82 |
< |
public HashCode initialValue() { |
83 |
< |
int h = rng.nextInt(); |
84 |
< |
return new HashCode((h == 0) ? 1 : h); // ensure nonzero |
85 |
< |
} |
131 |
> |
public HashCode initialValue() { return new HashCode(); } |
132 |
|
} |
133 |
|
|
134 |
|
/** |
135 |
|
* Static per-thread hash codes. Shared across all StripedAdders |
136 |
< |
* because adjustments due to collisions in one table are likely |
137 |
< |
* to be appropriate for others. |
136 |
> |
* to reduce ThreadLocal pollution and because adjustments due to |
137 |
> |
* collisions in one table are likely to be appropriate for |
138 |
> |
* others. |
139 |
|
*/ |
140 |
|
static final ThreadHashCode threadHashCode = new ThreadHashCode(); |
141 |
|
|
142 |
|
/** |
143 |
< |
* Table of adders. Initially of size 2; grows to be at most NCPU. |
143 |
> |
* Table of adders. Size is power of two, grows to be at most CAP. |
144 |
|
*/ |
145 |
|
private transient volatile Adder[] adders; |
146 |
|
|
147 |
|
/** |
148 |
|
* Serves as a lock when resizing and/or creating Adders. There |
149 |
< |
* is no need for a blocking lock: When busy, other threads try |
150 |
< |
* other slots. |
149 |
> |
* is no need for a blocking lock: Except during initialization |
150 |
> |
* races, when busy, other threads try other slots. However, |
151 |
> |
* during (double-checked) initializations, we use the |
152 |
> |
* "synchronized" lock on this object. |
153 |
|
*/ |
154 |
|
private final AtomicInteger mutex; |
155 |
|
|
156 |
|
/** |
157 |
< |
* Marsaglia XorShift for rehashing on collisions |
157 |
> |
* Creates a new adder with zero sum. |
158 |
|
*/ |
159 |
< |
private static int xorShift(int r) { |
160 |
< |
r ^= r << 13; |
161 |
< |
r ^= r >>> 17; |
113 |
< |
return r ^ (r << 5); |
159 |
> |
public StripedAdder() { |
160 |
> |
this.mutex = new AtomicInteger(); |
161 |
> |
// remaining initialization on first call to add. |
162 |
|
} |
163 |
|
|
164 |
|
/** |
165 |
< |
* Creates a new adder with initially zero sum. |
165 |
> |
* Creates a new adder with zero sum, and with stripes presized |
166 |
> |
* for the given expected contention level. |
167 |
> |
* |
168 |
> |
* @param expectedContention the expected number of threads that |
169 |
> |
* will concurrently update the sum. |
170 |
|
*/ |
171 |
< |
public StripedAdder() { |
172 |
< |
Adder[] as = new Adder[2]; |
173 |
< |
as[0] = new Adder(0); // ensure at least one available adder |
171 |
> |
public StripedAdder(int expectedContention) { |
172 |
> |
int cap = (expectedContention < CAP) ? expectedContention : CAP; |
173 |
> |
int size = 1; |
174 |
> |
while (size < cap) |
175 |
> |
size <<= 1; |
176 |
> |
Adder[] as = new Adder[size]; |
177 |
> |
for (int i = 0; i < size; ++i) |
178 |
> |
as[i] = new Adder(0); |
179 |
|
this.adders = as; |
180 |
|
this.mutex = new AtomicInteger(); |
181 |
|
} |
186 |
|
* @param x the value to add |
187 |
|
*/ |
188 |
|
public void add(long x) { |
189 |
+ |
Adder[] as; Adder a; int n; long v; // locals to hold volatile reads |
190 |
|
HashCode hc = threadHashCode.get(); |
191 |
< |
for (int h = hc.code;;) { |
192 |
< |
Adder[] as = adders; |
193 |
< |
int n = as.length; |
194 |
< |
Adder a = as[h & (n - 1)]; |
195 |
< |
if (a != null) { |
196 |
< |
long v = a.get(); |
197 |
< |
if (a.compareAndSet(v, v + x)) |
198 |
< |
break; |
199 |
< |
if (n >= NCPU) { // Collision when table at max |
200 |
< |
h = hc.code = xorShift(h); // change code |
201 |
< |
continue; |
191 |
> |
int h = hc.code; |
192 |
> |
if ((as = adders) == null || (n = as.length) < 1 || |
193 |
> |
(a = as[(n - 1) & h]) == null || |
194 |
> |
!a.compareAndSet(v = a.get(), v + x)) |
195 |
> |
retryAdd(x, hc); |
196 |
> |
} |
197 |
> |
|
198 |
> |
/** |
199 |
> |
* Handle cases of add involving initialization, resizing, |
200 |
> |
* creating new Adders, and/or contention. See above for |
201 |
> |
* explanation. |
202 |
> |
*/ |
203 |
> |
private void retryAdd(long x, HashCode hc) { |
204 |
> |
int h = hc.code; |
205 |
> |
final AtomicInteger mutex = this.mutex; |
206 |
> |
int collisions = 1 - mutex.get(); // first guess: collides if not locked |
207 |
> |
for (;;) { |
208 |
> |
Adder[] as; Adder a; long v; int k, n; |
209 |
> |
while ((as = adders) == null || (n = as.length) < 1) { |
210 |
> |
synchronized(mutex) { // Try to initialize |
211 |
> |
if (adders == null) { |
212 |
> |
Adder[] rs = new Adder[DEFAULT_INITIAL_SIZE]; |
213 |
> |
rs[h & (DEFAULT_INITIAL_SIZE - 1)] = new Adder(0); |
214 |
> |
adders = rs; |
215 |
> |
} |
216 |
> |
} |
217 |
> |
collisions = 0; |
218 |
> |
} |
219 |
> |
|
220 |
> |
if ((a = as[k = (n - 1) & h]) == null) { // Try to add slot |
221 |
> |
if (mutex.get() == 0 && mutex.compareAndSet(0, 1)) { |
222 |
> |
try { |
223 |
> |
if (adders == as && as[k] == null) |
224 |
> |
a = as[k] = new Adder(x); |
225 |
> |
} finally { |
226 |
> |
mutex.set(0); |
227 |
> |
} |
228 |
> |
if (a != null) |
229 |
> |
break; |
230 |
|
} |
231 |
+ |
collisions = 0; |
232 |
|
} |
233 |
< |
final AtomicInteger mutex = this.mutex; |
234 |
< |
if (mutex.get() != 0) |
148 |
< |
h = xorShift(h); // Try elsewhere |
149 |
< |
else if (mutex.compareAndSet(0, 1)) { |
150 |
< |
boolean created = false; |
233 |
> |
else if (collisions != 0 && n < CAP && // Try to expand table |
234 |
> |
mutex.get() == 0 && mutex.compareAndSet(0, 1)) { |
235 |
|
try { |
236 |
< |
Adder[] rs = adders; |
237 |
< |
if (a != null && rs == as) // Resize table |
238 |
< |
rs = adders = Arrays.copyOf(as, as.length << 1); |
239 |
< |
int j = h & (rs.length - 1); |
240 |
< |
if (rs[j] == null) { // Create adder |
157 |
< |
rs[j] = new Adder(x); |
158 |
< |
created = true; |
236 |
> |
if (adders == as) { |
237 |
> |
Adder[] rs = new Adder[n << 1]; |
238 |
> |
for (int i = 0; i < n; ++i) |
239 |
> |
rs[i] = as[i]; |
240 |
> |
adders = rs; |
241 |
|
} |
242 |
|
} finally { |
243 |
|
mutex.set(0); |
244 |
|
} |
245 |
< |
if (created) { |
164 |
< |
hc.code = h; // Use this adder next time |
165 |
< |
break; |
166 |
< |
} |
245 |
> |
collisions = 0; |
246 |
|
} |
247 |
+ |
else if (a.compareAndSet(v = a.get(), v + x)) |
248 |
+ |
break; |
249 |
+ |
else |
250 |
+ |
collisions = 1; |
251 |
+ |
h ^= h << 13; // Rehash |
252 |
+ |
h ^= h >>> 17; |
253 |
+ |
h ^= h << 5; |
254 |
|
} |
255 |
+ |
hc.code = h; |
256 |
|
} |
257 |
|
|
258 |
|
/** |
263 |
|
* @return the estimated sum |
264 |
|
*/ |
265 |
|
public long sum() { |
266 |
< |
long sum = 0; |
266 |
> |
long sum = 0L; |
267 |
|
Adder[] as = adders; |
268 |
< |
int n = as.length; |
269 |
< |
for (int i = 0; i < n; ++i) { |
270 |
< |
Adder a = as[i]; |
271 |
< |
if (a != null) |
272 |
< |
sum += a.get(); |
268 |
> |
if (as != null) { |
269 |
> |
int n = as.length; |
270 |
> |
for (int i = 0; i < n; ++i) { |
271 |
> |
Adder a = as[i]; |
272 |
> |
if (a != null) |
273 |
> |
sum += a.get(); |
274 |
> |
} |
275 |
|
} |
276 |
|
return sum; |
277 |
|
} |
283 |
|
*/ |
284 |
|
public void reset() { |
285 |
|
Adder[] as = adders; |
286 |
< |
int n = as.length; |
287 |
< |
for (int i = 0; i < n; ++i) { |
288 |
< |
Adder a = as[i]; |
289 |
< |
if (a != null) |
290 |
< |
a.set(0L); |
286 |
> |
if (as != null) { |
287 |
> |
int n = as.length; |
288 |
> |
for (int i = 0; i < n; ++i) { |
289 |
> |
Adder a = as[i]; |
290 |
> |
if (a != null) |
291 |
> |
a.set(0L); |
292 |
> |
} |
293 |
|
} |
294 |
|
} |
295 |
|
|
313 |
|
* @return the estimated sum |
314 |
|
*/ |
315 |
|
public long sumAndReset() { |
316 |
< |
long sum = 0; |
316 |
> |
long sum = 0L; |
317 |
|
Adder[] as = adders; |
318 |
< |
int n = as.length; |
319 |
< |
for (int i = 0; i < n; ++i) { |
320 |
< |
Adder a = as[i]; |
321 |
< |
if (a != null) { |
322 |
< |
sum += a.get(); |
323 |
< |
a.set(0L); |
318 |
> |
if (as != null) { |
319 |
> |
int n = as.length; |
320 |
> |
for (int i = 0; i < n; ++i) { |
321 |
> |
Adder a = as[i]; |
322 |
> |
if (a != null) { |
323 |
> |
sum += a.get(); |
324 |
> |
a.set(0L); |
325 |
> |
} |
326 |
|
} |
327 |
|
} |
328 |
|
return sum; |
337 |
|
private void readObject(ObjectInputStream s) |
338 |
|
throws IOException, ClassNotFoundException { |
339 |
|
s.defaultReadObject(); |
247 |
– |
long c = s.readLong(); |
248 |
– |
Adder[] as = new Adder[2]; |
249 |
– |
as[0] = new Adder(c); |
250 |
– |
this.adders = as; |
340 |
|
mutex.set(0); |
341 |
+ |
add(s.readLong()); |
342 |
|
} |
343 |
|
|
344 |
|
} |
255 |
– |
|
256 |
– |
|