C'est comment je pourrais (ou a) écrire un efficace tampon circulaire en Java. Il est appuyé par un simple tableau. Pour mon cas d'utilisation particulier, j'avais besoin de haut débit concurrent, j'ai donc utilisé la CAS pour l'allocation de l'indice. J'ai ensuite créé des mécanismes pour la fiabilité des copies, y compris un CAS exemplaire de l'ensemble de la mémoire tampon. J'ai utilisé ce dans un cas qui est décrit plus en détail dans un court article.
import java.util.concurrent.atomic.AtomicLong;
import java.lang.reflect.Array;
/**
* A circular array buffer with a copy-and-swap cursor.
*
* <p>This class provides an list of T objects who's size is <em>unstable</em>.
* It's intended for capturing data where the frequency of sampling greatly
* outweighs the frequency of inspection (for instance, monitoring).</p>
*
* <p>This object keeps in memory a fixed size buffer which is used for
* capturing objects. It copies the objects to a snapshot array which may be
* worked with. The size of the snapshot array will vary based on the
* stability of the array during the copy operation.</p>
*
* <p>Adding buffer to the buffer is <em>O(1)</em>, and lockless. Taking a
* stable copy of the sample is <em>O(n)</em>.</p>
*/
public class ConcurrentCircularBuffer <T> {
private final AtomicLong cursor = new AtomicLong();
private final T[] buffer;
private final Class<T> type;
/**
* Create a new concurrent circular buffer.
*
* @param type The type of the array. This is captured for the same reason
* it's required by {@link java.util.List.toArray()}.
*
* @param bufferSize The size of the buffer.
*
* @throws IllegalArgumentException if the bufferSize is a non-positive
* value.
*/
public ConcurrentCircularBuffer (final Class <T> type,
final int bufferSize)
{
if (bufferSize < 1) {
throw new IllegalArgumentException(
"Buffer size must be a positive value"
);
}
this.type = type;
this.buffer = (T[]) new Object [ bufferSize ];
}
/**
* Add a new object to this buffer.
*
* <p>Add a new object to the cursor-point of the buffer.</p>
*
* @param sample The object to add.
*/
public void add (T sample) {
buffer[(int) (cursor.getAndIncrement() % buffer.length)] = sample;
}
/**
* Return a stable snapshot of the buffer.
*
* <p>Capture a stable snapshot of the buffer as an array. The snapshot
* may not be the same length as the buffer, any objects which were
* unstable during the copy will be factored out.</p>
*
* @return An array snapshot of the buffer.
*/
public T[] snapshot () {
T[] snapshots = (T[]) new Object [ buffer.length ];
/* Determine the size of the snapshot by the number of affected
* records. Trim the size of the snapshot by the number of records
* which are considered to be unstable during the copy (the amount the
* cursor may have moved while the copy took place).
*
* If the cursor eliminated the sample (if the sample size is so small
* compared to the rate of mutation that it did a full-wrap during the
* copy) then just treat the buffer as though the cursor is
* buffer.length - 1 and it was not changed during copy (this is
* unlikley, but it should typically provide fairly stable results).
*/
long before = cursor.get();
/* If the cursor hasn't yet moved, skip the copying and simply return a
* zero-length array.
*/
if (before == 0) {
return (T[]) Array.newInstance(type, 0);
}
System.arraycopy(buffer, 0, snapshots, 0, buffer.length);
long after = cursor.get();
int size = buffer.length - (int) (after - before);
long snapshotCursor = before - 1;
/* Highly unlikely, but the entire buffer was replaced while we
* waited...so just return a zero length array, since we can't get a
* stable snapshot...
*/
if (size <= 0) {
return (T[]) Array.newInstance(type, 0);
}
long start = snapshotCursor - (size - 1);
long end = snapshotCursor;
if (snapshotCursor < snapshots.length) {
size = (int) snapshotCursor + 1;
start = 0;
}
/* Copy the sample snapshot to a new array the size of our stable
* snapshot area.
*/
T[] result = (T[]) Array.newInstance(type, size);
int startOfCopy = (int) (start % snapshots.length);
int endOfCopy = (int) (end % snapshots.length);
/* If the buffer space wraps the physical end of the array, use two
* copies to construct the new array.
*/
if (startOfCopy > endOfCopy) {
System.arraycopy(snapshots, startOfCopy,
result, 0,
snapshots.length - startOfCopy);
System.arraycopy(snapshots, 0,
result, (snapshots.length - startOfCopy),
endOfCopy + 1);
}
else {
/* Otherwise it's a single continuous segment, copy the whole thing
* into the result.
*/
System.arraycopy(snapshots, startOfCopy,
result, 0, endOfCopy - startOfCopy + 1);
}
return (T[]) result;
}
/**
* Get a stable snapshot of the complete buffer.
*
* <p>This operation fetches a snapshot of the buffer using the algorithm
* defined in {@link snapshot()}. If there was concurrent modification of
* the buffer during the copy, however, it will retry until a full stable
* snapshot of the buffer was acquired.</p>
*
* <p><em>Note, for very busy buffers on large symmetric multiprocessing
* machines and supercomputers running data processing intensive
* applications, this operation has the potential of being fairly
* expensive. In practice on commodity hardware, dualcore processors and
* non-processing intensive systems (such as web services) it very rarely
* retries.</em></p>
*
* @return A full copy of the internal buffer.
*/
public T[] completeSnapshot () {
T[] snapshot = snapshot();
/* Try again until we get a snapshot that's the same size as the
* buffer... This is very often a single iteration, but it depends on
* how busy the system is.
*/
while (snapshot.length != buffer.length) {
snapshot = snapshot();
}
return snapshot;
}
/**
* The size of this buffer.
*/
public int size () {
return buffer.length;
}
}
