org.cliffc.high_scale_lib.ConcurrentAutoTable Maven / Gradle / Ivy
/*
* Written by Cliff Click and released to the public domain, as explained at
* http://creativecommons.org/licenses/publicdomain
*/
package org.cliffc.high_scale_lib;
import java.io.Serializable;
import java.util.concurrent.atomic.*;
import sun.misc.Unsafe;
/**
* An auto-resizing table of {@code longs}, supporting low-contention CAS
* operations. Updates are done with CAS's to no particular table element.
* The intent is to support highly scalable counters, r/w locks, and other
* structures where the updates are associative, loss-free (no-brainer), and
* otherwise happen at such a high volume that the cache contention for
* CAS'ing a single word is unacceptable.
*
* This API is overkill for simple counters (e.g. no need for the 'mask')
* and is untested as an API for making a scalable r/w lock and so is likely
* to change!
*
* @since 1.5
* @author Cliff Click
*/
public class ConcurrentAutoTable implements Serializable {
// --- public interface ---
/**
* Add the given value to current counter value. Concurrent updates will
* not be lost, but addAndGet or getAndAdd are not implemented because the
* total counter value (i.e., {@link #get}) is not atomically updated.
* Updates are striped across an array of counters to avoid cache contention
* and has been tested with performance scaling linearly up to 768 CPUs.
*/
public void add( long x ) { add_if_mask( x,0); }
/** {@link #add} with -1 */
public void decrement() { add_if_mask(-1L,0); }
/** {@link #add} with +1 */
public void increment() { add_if_mask( 1L,0); }
/** Atomically set the sum of the striped counters to specified value.
* Rather more expensive than a simple store, in order to remain atomic.
*/
public void set( long x ) {
CAT newcat = new CAT(null,4,x);
// Spin until CAS works
while( !CAS_cat(_cat,newcat) );
}
/**
* Current value of the counter. Since other threads are updating furiously
* the value is only approximate, but it includes all counts made by the
* current thread. Requires a pass over the internally striped counters.
*/
public long get() { return _cat.sum(0); }
/** Same as {@link #get}, included for completeness. */
public int intValue() { return (int)_cat.sum(0); }
/** Same as {@link #get}, included for completeness. */
public long longValue() { return _cat.sum(0); }
/**
* A cheaper {@link #get}. Updated only once/millisecond, but as fast as a
* simple load instruction when not updating.
*/
public long estimate_get( ) { return _cat.estimate_sum(0); }
/**
* Return the counter's {@code long} value converted to a string.
*/
public String toString() { return _cat.toString(0); }
/**
* A more verbose print than {@link #toString}, showing internal structure.
* Useful for debugging.
*/
public void print() { _cat.print(); }
/**
* Return the internal counter striping factor. Useful for diagnosing
* performance problems.
*/
public int internal_size() { return _cat._t.length; }
// Only add 'x' to some slot in table, hinted at by 'hash', if bits under
// the mask are all zero. The sum can overflow or 'x' can contain bits in
// the mask. Value is CAS'd so no counts are lost. The CAS is retried until
// it succeeds or bits are found under the mask. Returned value is the old
// value - which WILL have zero under the mask on success and WILL NOT have
// zero under the mask for failure.
private long add_if_mask( long x, long mask ) { return _cat.add_if_mask(x,mask,hash(),this); }
// The underlying array of concurrently updated long counters
private volatile CAT _cat = new CAT(null,4/*Start Small, Think Big!*/,0L);
private static final AtomicReferenceFieldUpdater _catUpdater =
AtomicReferenceFieldUpdater.newUpdater(ConcurrentAutoTable.class,CAT.class, "_cat");
private boolean CAS_cat( CAT oldcat, CAT newcat ) { return _catUpdater.compareAndSet(this,oldcat,newcat); }
// Hash spreader
private static final int hash() {
int h = System.identityHashCode(Thread.currentThread());
// You would think that System.identityHashCode on the current thread
// would be a good hash fcn, but actually on SunOS 5.8 it is pretty lousy
// in the low bits.
h ^= (h>>>20) ^ (h>>>12); // Bit spreader, borrowed from Doug Lea
h ^= (h>>> 7) ^ (h>>> 4);
return h<<2; // Pad out cache lines. The goal is to avoid cache-line contention
}
// --- CAT -----------------------------------------------------------------
private static class CAT implements Serializable {
// Unsafe crud: get a function which will CAS arrays
private static final Unsafe _unsafe = UtilUnsafe.getUnsafe();
private static final int _Lbase = _unsafe.arrayBaseOffset(long[].class);
private static final int _Lscale = _unsafe.arrayIndexScale(long[].class);
private static long rawIndex(long[] ary, int i) {
assert i >= 0 && i < ary.length;
return _Lbase + i * _Lscale;
}
private final static boolean CAS( long[] A, int idx, long old, long nnn ) {
return _unsafe.compareAndSwapLong( A, rawIndex(A,idx), old, nnn );
}
volatile long _resizers; // count of threads attempting a resize
static private final AtomicLongFieldUpdater _resizerUpdater =
AtomicLongFieldUpdater.newUpdater(CAT.class, "_resizers");
private final CAT _next;
private volatile long _sum_cache;
private volatile long _fuzzy_sum_cache;
private volatile long _fuzzy_time;
private static final int MAX_SPIN=2;
private long[] _t; // Power-of-2 array of longs
CAT( CAT next, int sz, long init ) {
_next = next;
_sum_cache = Long.MIN_VALUE;
_t = new long[sz];
_t[0] = init;
}
// Only add 'x' to some slot in table, hinted at by 'hash', if bits under
// the mask are all zero. The sum can overflow or 'x' can contain bits in
// the mask. Value is CAS'd so no counts are lost. The CAS is attempted
// ONCE.
public long add_if_mask( long x, long mask, int hash, ConcurrentAutoTable master ) {
long[] t = _t;
int idx = hash & (t.length-1);
// Peel loop; try once fast
long old = t[idx];
boolean ok = CAS( t, idx, old&~mask, old+x );
if( _sum_cache != Long.MIN_VALUE )
_sum_cache = Long.MIN_VALUE; // Blow out cache
if( ok ) return old; // Got it
if( (old&mask) != 0 ) return old; // Failed for bit-set under mask
// Try harder
int cnt=0;
while( true ) {
old = t[idx];
if( (old&mask) != 0 ) return old; // Failed for bit-set under mask
if( CAS( t, idx, old, old+x ) ) break; // Got it!
cnt++;
}
if( cnt < MAX_SPIN ) return old; // Allowable spin loop count
if( t.length >= 1024*1024 ) return old; // too big already
// Too much contention; double array size in an effort to reduce contention
long r = _resizers;
int newbytes = (t.length<<1)<<3/*word to bytes*/;
while( !_resizerUpdater.compareAndSet(this,r,r+newbytes) )
r = _resizers;
r += newbytes;
if( master._cat != this ) return old; // Already doubled, don't bother
if( (r>>17) != 0 ) { // Already too much allocation attempts?
// TODO - use a wait with timeout, so we'll wakeup as soon as the new
// table is ready, or after the timeout in any case. Annoyingly, this
// breaks the non-blocking property - so for now we just briefly sleep.
//synchronized( this ) { wait(8*megs); } // Timeout - we always wakeup
try { Thread.sleep(r>>17); } catch( InterruptedException e ) { }
if( master._cat != this ) return old;
}
CAT newcat = new CAT(this,t.length*2,0);
// Take 1 stab at updating the CAT with the new larger size. If this
// fails, we assume some other thread already expanded the CAT - so we
// do not need to retry until it succeeds.
master.CAS_cat(this,newcat);
return old;
}
// Return the current sum of all things in the table, stripping off mask
// before the add. Writers can be updating the table furiously, so the
// sum is only locally accurate.
public long sum( long mask ) {
long sum = _sum_cache;
if( sum != Long.MIN_VALUE ) return sum;
sum = _next == null ? 0 : _next.sum(mask); // Recursively get cached sum
long[] t = _t;
for( int i=0; i