| 44 |
|
|
| 45 |
|
/* |
| 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. |
| 47 |
> |
* table is indexed by per-thread hash codes. |
| 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 |
| 53 |
< |
* bound to a CPU, there would exist a perfect hash function |
| 54 |
< |
* mapping threads to slots that eliminates collisions. When we |
| 55 |
< |
* reach capacity, we search for this mapping by randomly varying |
| 56 |
< |
* the hash codes of colliding threads. Because search is random, |
| 57 |
< |
* and failures only become known via CAS failures, convergence |
| 58 |
< |
* will be slow, and because threads are typically not bound to |
| 59 |
< |
* CPUS forever, may not occur at all. However, despite these |
| 60 |
< |
* limitations, observed contention is typically low in these |
| 61 |
< |
* cases. |
| 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 most often |
| 79 |
< |
* irregularly scattered in memory and thus don't interfere much |
| 80 |
< |
* with each other. But Atomic objects residing in arrays will |
| 81 |
< |
* tend to be placed adjacent to each other, and so will most |
| 82 |
< |
* often share cache lines without this precaution. Adders are |
| 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 (an already large) footprint. |
| 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 |
|
* try other slots rather than blocking. After initialization, at |
| 89 |
|
* least one slot exists, so retries will eventually find a |
| 90 |
< |
* candidate Adder. During these retries, there is increased |
| 90 |
> |
* candidate Adder. During these retries, there is increased |
| 91 |
|
* contention and reduced locality, which is still better than |
| 92 |
|
* alternatives. |
| 93 |
|
*/ |
| 94 |
|
|
| 95 |
|
/** |
| 84 |
– |
* Number of processors, to place a cap on table growth. |
| 85 |
– |
*/ |
| 86 |
– |
static final int NCPU = Runtime.getRuntime().availableProcessors(); |
| 87 |
– |
|
| 88 |
– |
/** |
| 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. The code starts off with |
| 107 |
< |
* a given random value, but may be set to a different value upon |
| 108 |
< |
* collisions in retryAdd. |
| 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(); |
| 111 |
< |
public HashCode initialValue() { |
| 112 |
< |
int h = rng.nextInt(); |
| 113 |
< |
return new HashCode((h == 0) ? 1 : h); // ensure nonzero |
| 114 |
< |
} |
| 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. Minimum size 2. Size grows to be at most NCPU. |
| 143 |
> |
* Common placeholder for empty arrays. |
| 144 |
> |
*/ |
| 145 |
> |
static final Adder[] EMPTY_ARRAY = new Adder[0]; |
| 146 |
> |
|
| 147 |
> |
/** |
| 148 |
> |
* Table of adders. Size is either zero or a power of two, grows |
| 149 |
> |
* to be at most CAP. |
| 150 |
|
*/ |
| 151 |
|
private transient volatile Adder[] adders; |
| 152 |
|
|
| 153 |
|
/** |
| 154 |
|
* Serves as a lock when resizing and/or creating Adders. There |
| 155 |
< |
* is no need for a blocking lock: When busy, other threads try |
| 156 |
< |
* other slots. |
| 155 |
> |
* is no need for a blocking lock: Except during initialization |
| 156 |
> |
* races, when busy, other threads try other slots. However, |
| 157 |
> |
* during (double-checked) initializations, we use the |
| 158 |
> |
* "synchronized" lock on this object. |
| 159 |
|
*/ |
| 160 |
|
private final AtomicInteger mutex; |
| 161 |
|
|
| 163 |
|
* Creates a new adder with zero sum. |
| 164 |
|
*/ |
| 165 |
|
public StripedAdder() { |
| 166 |
+ |
this.adders = EMPTY_ARRAY; |
| 167 |
|
this.mutex = new AtomicInteger(); |
| 168 |
|
// remaining initialization on first call to add. |
| 169 |
|
} |
| 176 |
|
* will concurrently update the sum. |
| 177 |
|
*/ |
| 178 |
|
public StripedAdder(int expectedContention) { |
| 179 |
< |
int cap = (expectedContention < NCPU) ? expectedContention : NCPU; |
| 180 |
< |
int size = 2; |
| 181 |
< |
while (size < cap) |
| 182 |
< |
size <<= 1; |
| 183 |
< |
Adder[] as = new Adder[size]; |
| 184 |
< |
for (int i = 0; i < size; ++i) |
| 185 |
< |
as[i] = new Adder(0); |
| 186 |
< |
this.adders = as; |
| 179 |
> |
if (expectedContention > 0) { |
| 180 |
> |
int cap = (expectedContention < CAP) ? expectedContention : CAP; |
| 181 |
> |
int size = 1; |
| 182 |
> |
while (size < cap) |
| 183 |
> |
size <<= 1; |
| 184 |
> |
Adder[] as = new Adder[size]; |
| 185 |
> |
for (int i = 0; i < size; ++i) |
| 186 |
> |
as[i] = new Adder(0); |
| 187 |
> |
this.adders = as; |
| 188 |
> |
} |
| 189 |
> |
else |
| 190 |
> |
this.adders = EMPTY_ARRAY; |
| 191 |
|
this.mutex = new AtomicInteger(); |
| 192 |
|
} |
| 193 |
|
|
| 199 |
|
public void add(long x) { |
| 200 |
|
Adder[] as; Adder a; int n; long v; // locals to hold volatile reads |
| 201 |
|
HashCode hc = threadHashCode.get(); |
| 202 |
+ |
int h = hc.code; |
| 203 |
|
if ((as = adders) == null || (n = as.length) < 1 || |
| 204 |
< |
(a = as[hc.code & (n - 1)]) == null || |
| 204 |
> |
(a = as[(n - 1) & h]) == null || |
| 205 |
|
!a.compareAndSet(v = a.get(), v + x)) |
| 206 |
|
retryAdd(x, hc); |
| 207 |
|
} |
| 208 |
|
|
| 209 |
|
/** |
| 210 |
|
* Handle cases of add involving initialization, resizing, |
| 211 |
< |
* creating new Adders, and/or contention. |
| 211 |
> |
* creating new Adders, and/or contention. See above for |
| 212 |
> |
* explanation. |
| 213 |
|
*/ |
| 214 |
|
private void retryAdd(long x, HashCode hc) { |
| 215 |
|
int h = hc.code; |
| 216 |
|
final AtomicInteger mutex = this.mutex; |
| 217 |
< |
AtomicInteger lock = null; // nonnull when held |
| 218 |
< |
try { |
| 219 |
< |
for (;;) { |
| 220 |
< |
Adder[] as; Adder a; long v; int n, k; // locals for volatiles |
| 221 |
< |
boolean needLock = true; |
| 222 |
< |
if ((as = adders) == null || (n = as.length) < 1) { |
| 223 |
< |
if (lock != null) // default-initialize |
| 224 |
< |
adders = new Adder[2]; |
| 217 |
> |
int collisions = 1 - mutex.get(); // first guess: collides if not locked |
| 218 |
> |
for (;;) { |
| 219 |
> |
Adder[] as; Adder a; long v; int k, n; |
| 220 |
> |
while ((as = adders) == null || (n = as.length) < 1) { |
| 221 |
> |
synchronized(mutex) { // Try to initialize |
| 222 |
> |
if (adders == as) { |
| 223 |
> |
Adder[] rs = new Adder[DEFAULT_INITIAL_SIZE]; |
| 224 |
> |
rs[h & (DEFAULT_INITIAL_SIZE - 1)] = new Adder(0); |
| 225 |
> |
adders = rs; |
| 226 |
> |
} |
| 227 |
|
} |
| 228 |
< |
else if ((a = as[k = h & (n - 1)]) == null) { |
| 229 |
< |
if (lock != null) { // attach new adder |
| 230 |
< |
as[k] = new Adder(x); |
| 231 |
< |
break; |
| 228 |
> |
collisions = 0; |
| 229 |
> |
} |
| 230 |
> |
|
| 231 |
> |
if ((a = as[k = (n - 1) & h]) == null) { // Try to add slot |
| 232 |
> |
if (mutex.get() == 0 && mutex.compareAndSet(0, 1)) { |
| 233 |
> |
try { |
| 234 |
> |
if (adders == as && as[k] == null) |
| 235 |
> |
a = as[k] = new Adder(x); |
| 236 |
> |
} finally { |
| 237 |
> |
mutex.set(0); |
| 238 |
|
} |
| 239 |
+ |
if (a != null) |
| 240 |
+ |
break; |
| 241 |
|
} |
| 242 |
< |
else if (a.compareAndSet(v = a.get(), v + x)) |
| 243 |
< |
break; |
| 244 |
< |
else if (n >= NCPU) // cannot expand |
| 245 |
< |
needLock = false; |
| 246 |
< |
else if (lock != null) // expand table |
| 247 |
< |
adders = Arrays.copyOf(as, n << 1); |
| 248 |
< |
|
| 249 |
< |
if (lock == null) { |
| 250 |
< |
if (needLock && mutex.get() == 0 && |
| 251 |
< |
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; |
| 242 |
> |
collisions = 0; |
| 243 |
> |
} |
| 244 |
> |
else if (collisions != 0 && n < CAP && // Try to expand table |
| 245 |
> |
mutex.get() == 0 && mutex.compareAndSet(0, 1)) { |
| 246 |
> |
try { |
| 247 |
> |
if (adders == as) { |
| 248 |
> |
Adder[] rs = new Adder[n << 1]; |
| 249 |
> |
for (int i = 0; i < n; ++i) |
| 250 |
> |
rs[i] = as[i]; |
| 251 |
> |
adders = rs; |
| 252 |
|
} |
| 253 |
+ |
} finally { |
| 254 |
+ |
mutex.set(0); |
| 255 |
|
} |
| 256 |
+ |
collisions = 0; |
| 257 |
|
} |
| 258 |
< |
} finally { |
| 259 |
< |
if (lock != null) |
| 260 |
< |
lock.set(0); |
| 258 |
> |
else if (a.compareAndSet(v = a.get(), v + x)) |
| 259 |
> |
break; |
| 260 |
> |
else |
| 261 |
> |
collisions = 1; |
| 262 |
> |
h ^= h << 13; // Rehash |
| 263 |
> |
h ^= h >>> 17; |
| 264 |
> |
h ^= h << 5; |
| 265 |
|
} |
| 266 |
< |
if (hc.code != h) // avoid unneeded writes |
| 222 |
< |
hc.code = h; |
| 266 |
> |
hc.code = h; |
| 267 |
|
} |
| 268 |
|
|
| 269 |
|
/** |
