org.apache.cassandra.utils.memory.NativeAllocator Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of cassandra-all Show documentation
Show all versions of cassandra-all Show documentation
The Apache Cassandra Project develops a highly scalable second-generation distributed database, bringing together Dynamo's fully distributed design and Bigtable's ColumnFamily-based data model.
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.cassandra.utils.memory;
import java.util.HashMap;
import java.util.Map;
import java.util.concurrent.ConcurrentLinkedQueue;
import java.util.concurrent.Semaphore;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.atomic.AtomicReference;
import org.apache.cassandra.db.*;
import org.apache.cassandra.db.rows.*;
import org.apache.cassandra.utils.concurrent.OpOrder;
/**
* This NativeAllocator uses global slab allocation strategy
* with slab size that scales exponentially from 8kb to 1Mb to
* serve allocation of up to 128kb.
*
*
* The slab allocation reduces heap fragmentation from small
* long-lived objects.
*
*/
public class NativeAllocator extends MemtableAllocator
{
private final static int MAX_REGION_SIZE = 1 * 1024 * 1024;
private final static int MAX_CLONED_SIZE = 128 * 1024; // bigger than this don't go in the region
private final static int MIN_REGION_SIZE = 8 * 1024;
// globally stash any Regions we allocate but are beaten to using, and use these up before allocating any more
private static final Map RACE_ALLOCATED = new HashMap<>();
static
{
for(int i = MIN_REGION_SIZE ; i <= MAX_REGION_SIZE; i *= 2)
RACE_ALLOCATED.put(i, new RaceAllocated());
}
private final AtomicReference currentRegion = new AtomicReference<>();
private final ConcurrentLinkedQueue regions = new ConcurrentLinkedQueue<>();
private final EnsureOnHeap.CloneToHeap cloneToHeap = new EnsureOnHeap.CloneToHeap();
protected NativeAllocator(NativePool pool)
{
super(pool.onHeap.newAllocator(), pool.offHeap.newAllocator());
}
private static class CloningBTreeRowBuilder extends BTreeRow.Builder
{
final OpOrder.Group writeOp;
final NativeAllocator allocator;
private CloningBTreeRowBuilder(OpOrder.Group writeOp, NativeAllocator allocator)
{
super(true);
this.writeOp = writeOp;
this.allocator = allocator;
}
@Override
public void newRow(Clustering clustering)
{
if (clustering != Clustering.STATIC_CLUSTERING)
clustering = new NativeClustering(allocator, writeOp, clustering);
super.newRow(clustering);
}
@Override
public void addCell(Cell cell)
{
super.addCell(new NativeCell(allocator, writeOp, cell));
}
}
public Row.Builder rowBuilder(OpOrder.Group opGroup)
{
return new CloningBTreeRowBuilder(opGroup, this);
}
public DecoratedKey clone(DecoratedKey key, OpOrder.Group writeOp)
{
return new NativeDecoratedKey(key.getToken(), this, writeOp, key.getKey());
}
public EnsureOnHeap ensureOnHeap()
{
return cloneToHeap;
}
public long allocate(int size, OpOrder.Group opGroup)
{
assert size >= 0;
offHeap().allocate(size, opGroup);
// satisfy large allocations directly from JVM since they don't cause fragmentation
// as badly, and fill up our regions quickly
if (size > MAX_CLONED_SIZE)
return allocateOversize(size);
while (true)
{
Region region = currentRegion.get();
long peer;
if (region != null && (peer = region.allocate(size)) > 0)
return peer;
trySwapRegion(region, size);
}
}
private void trySwapRegion(Region current, int minSize)
{
// decide how big we want the new region to be:
// * if there is no prior region, we set it to min size
// * otherwise we double its size; if it's too small to fit the allocation, we round it up to 4-8x its size
int size;
if (current == null) size = MIN_REGION_SIZE;
else size = current.capacity * 2;
if (minSize > size)
size = Integer.highestOneBit(minSize) << 3;
size = Math.min(MAX_REGION_SIZE, size);
// first we try and repurpose a previously allocated region
RaceAllocated raceAllocated = RACE_ALLOCATED.get(size);
Region next = raceAllocated.poll();
// if there are none, we allocate one
if (next == null)
next = new Region(MemoryUtil.allocate(size), size);
// we try to swap in the region we've obtained;
// if we fail to swap the region, we try to stash it for repurposing later; if we're out of stash room, we free it
if (currentRegion.compareAndSet(current, next))
regions.add(next);
else if (!raceAllocated.stash(next))
MemoryUtil.free(next.peer);
}
private long allocateOversize(int size)
{
// satisfy large allocations directly from JVM since they don't cause fragmentation
// as badly, and fill up our regions quickly
Region region = new Region(MemoryUtil.allocate(size), size);
regions.add(region);
long peer;
if ((peer = region.allocate(size)) == -1)
throw new AssertionError();
return peer;
}
public void setDiscarded()
{
for (Region region : regions)
MemoryUtil.free(region.peer);
super.setDiscarded();
}
// used to ensure we don't keep loads of race allocated regions around indefinitely. keeps the total bound on wasted memory low.
private static class RaceAllocated
{
final ConcurrentLinkedQueue stash = new ConcurrentLinkedQueue<>();
final Semaphore permits = new Semaphore(8);
boolean stash(Region region)
{
if (!permits.tryAcquire())
return false;
stash.add(region);
return true;
}
Region poll()
{
Region next = stash.poll();
if (next != null)
permits.release();
return next;
}
}
/**
* A region of memory out of which allocations are sliced.
*
* This serves two purposes:
* - to provide a step between initialization and allocation, so that racing to CAS a
* new region in is harmless
* - encapsulates the allocation offset
*/
private static class Region
{
/**
* Actual underlying data
*/
private final long peer;
private final int capacity;
/**
* Offset for the next allocation, or the sentinel value -1
* which implies that the region is still uninitialized.
*/
private final AtomicInteger nextFreeOffset = new AtomicInteger(0);
/**
* Create an uninitialized region. Note that memory is not allocated yet, so
* this is cheap.
*
* @param peer peer
*/
private Region(long peer, int capacity)
{
this.peer = peer;
this.capacity = capacity;
}
/**
* Try to allocate size bytes from the region.
*
* @return the successful allocation, or -1 to indicate not-enough-space
*/
long allocate(int size)
{
int newOffset = nextFreeOffset.getAndAdd(size);
if (newOffset + size > capacity)
// this region is full
return -1;
return peer + newOffset;
}
@Override
public String toString()
{
return "Region@" + System.identityHashCode(this) +
"waste=" + Math.max(0, capacity - nextFreeOffset.get());
}
}
}