| 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 |
| 43 |
|
private static final long serialVersionUID = 7249069246863182397L; |
| 44 |
|
|
| 45 |
|
/* |
| 46 |
< |
* Overview: We maintain a table of Atomic long variables. The |
| 46 |
> |
* A StripedAdder maintains a table of Atomic long variables. The |
| 47 |
|
* table is indexed by per-thread hash codes that are initialized |
| 48 |
|
* to random values. |
| 49 |
|
* |
| 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 |
| 61 |
> |
* limitations, observed contention is typically low in these |
| 62 |
|
* cases. |
| 63 |
|
* |
| 64 |
|
* Table entries are of class Adder; a form of AtomicLong padded |
| 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. |
| 70 |
> |
* often share cache lines without this precaution. Adders are |
| 71 |
> |
* constructed upon first use, which further improves per-thread |
| 72 |
> |
* locality and helps reduce (an already large) footprint. |
| 73 |
|
* |
| 74 |
|
* A single spinlock is used for resizing the table as well as |
| 75 |
|
* populating slots with new Adders. Upon lock contention, threads |
| 76 |
< |
* try other slots rather than blocking. We guarantee that at |
| 77 |
< |
* least one slot (0) exists, so retries will eventually find a |
| 76 |
> |
* try other slots rather than blocking. After initialization, at |
| 77 |
> |
* least one slot exists, so retries will eventually find a |
| 78 |
|
* candidate Adder. During these retries, there is increased |
| 79 |
|
* contention and reduced locality, which is still better than |
| 80 |
|
* alternatives. |
| 95 |
|
|
| 96 |
|
/** |
| 97 |
|
* Holder for the thread-local hash code. The code starts off with |
| 98 |
< |
* a given random value, but may be set to a different |
| 99 |
< |
* pseudo-random value (using a cheaper but adequate xorshift |
| 101 |
< |
* generator) upon collisions. |
| 98 |
> |
* a given random value, but may be set to a different value upon |
| 99 |
> |
* collisions in retryAdd. |
| 100 |
|
*/ |
| 101 |
|
static final class HashCode { |
| 102 |
|
int code; |
| 134 |
|
private final AtomicInteger mutex; |
| 135 |
|
|
| 136 |
|
/** |
| 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 |
– |
|
| 147 |
– |
/** |
| 137 |
|
* Creates a new adder with zero sum. |
| 138 |
|
*/ |
| 139 |
|
public StripedAdder() { |
| 140 |
< |
this(2); |
| 140 |
> |
this.mutex = new AtomicInteger(); |
| 141 |
> |
// remaining initialization on first call to add. |
| 142 |
|
} |
| 143 |
|
|
| 144 |
|
/** |
| 154 |
|
while (size < cap) |
| 155 |
|
size <<= 1; |
| 156 |
|
Adder[] as = new Adder[size]; |
| 157 |
< |
as[0] = new Adder(0); // ensure at least one available adder |
| 157 |
> |
for (int i = 0; i < size; ++i) |
| 158 |
> |
as[i] = new Adder(0); |
| 159 |
|
this.adders = as; |
| 160 |
|
this.mutex = new AtomicInteger(); |
| 161 |
|
} |
| 166 |
|
* @param x the value to add |
| 167 |
|
*/ |
| 168 |
|
public void add(long x) { |
| 169 |
+ |
Adder[] as; Adder a; int n; long v; // locals to hold volatile reads |
| 170 |
|
HashCode hc = threadHashCode.get(); |
| 171 |
< |
for (int h = hc.code;;) { |
| 172 |
< |
Adder[] as = adders; |
| 173 |
< |
int n = as.length; |
| 174 |
< |
Adder a = as[h & (n - 1)]; |
| 175 |
< |
if (a != null) { |
| 176 |
< |
long v = a.get(); |
| 177 |
< |
if (a.compareAndSet(v, v + x)) |
| 178 |
< |
break; |
| 179 |
< |
if (n >= NCPU) { // Collision when table at max |
| 180 |
< |
h = hc.code = xorShift(h); // change code |
| 181 |
< |
continue; |
| 171 |
> |
if ((as = adders) == null || (n = as.length) < 1 || |
| 172 |
> |
(a = as[hc.code & (n - 1)]) == null || |
| 173 |
> |
!a.compareAndSet(v = a.get(), v + x)) |
| 174 |
> |
retryAdd(x, hc); |
| 175 |
> |
} |
| 176 |
> |
|
| 177 |
> |
/** |
| 178 |
> |
* Handle cases of add involving initialization, resizing, |
| 179 |
> |
* creating new Adders, and/or contention. |
| 180 |
> |
*/ |
| 181 |
> |
private void retryAdd(long x, HashCode hc) { |
| 182 |
> |
int h = hc.code; |
| 183 |
> |
final AtomicInteger mutex = this.mutex; |
| 184 |
> |
AtomicInteger lock = null; // nonnull when held |
| 185 |
> |
try { |
| 186 |
> |
for (;;) { |
| 187 |
> |
Adder[] as; Adder a; long v; int n, k; // locals for volatiles |
| 188 |
> |
boolean needLock = true; |
| 189 |
> |
if ((as = adders) == null || (n = as.length) < 1) { |
| 190 |
> |
if (lock != null) // default-initialize |
| 191 |
> |
adders = new Adder[2]; |
| 192 |
|
} |
| 193 |
< |
} |
| 194 |
< |
final AtomicInteger mutex = this.mutex; |
| 195 |
< |
if (mutex.get() != 0) |
| 196 |
< |
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; |
| 193 |
> |
else if ((a = as[k = h & (n - 1)]) == null) { |
| 194 |
> |
if (lock != null) { // attach new adder |
| 195 |
> |
as[k] = new Adder(x); |
| 196 |
> |
break; |
| 197 |
|
} |
| 206 |
– |
} finally { |
| 207 |
– |
mutex.set(0); |
| 198 |
|
} |
| 199 |
< |
if (created) { |
| 210 |
< |
hc.code = h; // Use this adder next time |
| 199 |
> |
else if (a.compareAndSet(v = a.get(), v + x)) |
| 200 |
|
break; |
| 201 |
+ |
else if (n >= NCPU) // cannot expand |
| 202 |
+ |
needLock = false; |
| 203 |
+ |
else if (lock != null) // expand table |
| 204 |
+ |
adders = Arrays.copyOf(as, n << 1); |
| 205 |
+ |
|
| 206 |
+ |
if (lock == null) { |
| 207 |
+ |
if (needLock && mutex.get() == 0 && |
| 208 |
+ |
mutex.compareAndSet(0, 1)) |
| 209 |
+ |
lock = mutex; |
| 210 |
+ |
else { // try elsewhere |
| 211 |
+ |
h ^= h << 13; // Marsaglia XorShift |
| 212 |
+ |
h ^= h >>> 17; |
| 213 |
+ |
h ^= h << 5; |
| 214 |
+ |
} |
| 215 |
|
} |
| 216 |
|
} |
| 217 |
+ |
} finally { |
| 218 |
+ |
if (lock != null) |
| 219 |
+ |
lock.set(0); |
| 220 |
|
} |
| 221 |
+ |
if (hc.code != h) // avoid unneeded writes |
| 222 |
+ |
hc.code = h; |
| 223 |
|
} |
| 224 |
|
|
| 225 |
|
/** |
| 230 |
|
* @return the estimated sum |
| 231 |
|
*/ |
| 232 |
|
public long sum() { |
| 233 |
< |
long sum = 0; |
| 233 |
> |
long sum = 0L; |
| 234 |
|
Adder[] as = adders; |
| 235 |
< |
int n = as.length; |
| 236 |
< |
for (int i = 0; i < n; ++i) { |
| 237 |
< |
Adder a = as[i]; |
| 238 |
< |
if (a != null) |
| 239 |
< |
sum += a.get(); |
| 235 |
> |
if (as != null) { |
| 236 |
> |
int n = as.length; |
| 237 |
> |
for (int i = 0; i < n; ++i) { |
| 238 |
> |
Adder a = as[i]; |
| 239 |
> |
if (a != null) |
| 240 |
> |
sum += a.get(); |
| 241 |
> |
} |
| 242 |
|
} |
| 243 |
|
return sum; |
| 244 |
|
} |
| 250 |
|
*/ |
| 251 |
|
public void reset() { |
| 252 |
|
Adder[] as = adders; |
| 253 |
< |
int n = as.length; |
| 254 |
< |
for (int i = 0; i < n; ++i) { |
| 255 |
< |
Adder a = as[i]; |
| 256 |
< |
if (a != null) |
| 257 |
< |
a.set(0L); |
| 253 |
> |
if (as != null) { |
| 254 |
> |
int n = as.length; |
| 255 |
> |
for (int i = 0; i < n; ++i) { |
| 256 |
> |
Adder a = as[i]; |
| 257 |
> |
if (a != null) |
| 258 |
> |
a.set(0L); |
| 259 |
> |
} |
| 260 |
|
} |
| 261 |
|
} |
| 262 |
|
|
| 280 |
|
* @return the estimated sum |
| 281 |
|
*/ |
| 282 |
|
public long sumAndReset() { |
| 283 |
< |
long sum = 0; |
| 283 |
> |
long sum = 0L; |
| 284 |
|
Adder[] as = adders; |
| 285 |
< |
int n = as.length; |
| 286 |
< |
for (int i = 0; i < n; ++i) { |
| 287 |
< |
Adder a = as[i]; |
| 288 |
< |
if (a != null) { |
| 289 |
< |
sum += a.get(); |
| 290 |
< |
a.set(0L); |
| 285 |
> |
if (as != null) { |
| 286 |
> |
int n = as.length; |
| 287 |
> |
for (int i = 0; i < n; ++i) { |
| 288 |
> |
Adder a = as[i]; |
| 289 |
> |
if (a != null) { |
| 290 |
> |
sum += a.get(); |
| 291 |
> |
a.set(0L); |
| 292 |
> |
} |
| 293 |
|
} |
| 294 |
|
} |
| 295 |
|
return sum; |
| 304 |
|
private void readObject(ObjectInputStream s) |
| 305 |
|
throws IOException, ClassNotFoundException { |
| 306 |
|
s.defaultReadObject(); |
| 293 |
– |
long c = s.readLong(); |
| 294 |
– |
Adder[] as = new Adder[2]; |
| 295 |
– |
as[0] = new Adder(c); |
| 296 |
– |
this.adders = as; |
| 307 |
|
mutex.set(0); |
| 308 |
+ |
add(s.readLong()); |
| 309 |
|
} |
| 310 |
|
|
| 311 |
|
} |
| 301 |
– |
|
| 302 |
– |
|
