move cleaning filter cache on closed readers to separate thread

improve cleaning the global weighted cache when a reader closes, move it to a separate thread, so iterating over the cache entries will nto happen on each segment closed, but instead be "bulked"
This commit is contained in:
Shay Banon 2012-09-29 14:06:28 +02:00
parent 9dd8e1ea05
commit 05653cbdaa
3 changed files with 75 additions and 18 deletions

View File

@ -83,12 +83,7 @@ public class WeightedFilterCache extends AbstractIndexComponent implements Filte
return;
}
seenReadersCount.dec();
for (FilterCacheKey key : indicesFilterCache.cache().asMap().keySet()) {
if (key.readerKey() == readerKey) {
// invalidate will cause a removal and will be notified
indicesFilterCache.cache().invalidate(key);
}
}
indicesFilterCache.addReaderKeyToClean(readerKey);
}
}
@ -106,13 +101,7 @@ public class WeightedFilterCache extends AbstractIndexComponent implements Filte
return;
}
seenReadersCount.dec();
Cache<FilterCacheKey, DocSet> cache = indicesFilterCache.cache();
for (FilterCacheKey key : cache.asMap().keySet()) {
if (key.readerKey() == reader.getCoreCacheKey()) {
// invalidate will cause a removal and will be notified
cache.invalidate(key);
}
}
indicesFilterCache.addReaderKeyToClean(reader.getCoreCacheKey());
}
@Override
@ -247,7 +236,7 @@ public class WeightedFilterCache extends AbstractIndexComponent implements Filte
if (this == o) return true;
// if (o == null || getClass() != o.getClass()) return false;
FilterCacheKey that = (FilterCacheKey) o;
return (readerKey == that.readerKey && filterKey.equals(that.filterKey));
return (readerKey.equals(that.readerKey) && filterKey.equals(that.filterKey));
}
@Override

View File

@ -719,7 +719,7 @@ public class InternalIndexShard extends AbstractIndexShardComponent implements I
}
}
private class EngineRefresher implements Runnable {
class EngineRefresher implements Runnable {
@Override
public void run() {
// we check before if a refresh is needed, if not, we reschedule, otherwise, we fork, refresh, and then reschedule
@ -767,7 +767,7 @@ public class InternalIndexShard extends AbstractIndexShardComponent implements I
}
}
private class EngineMerger implements Runnable {
class EngineMerger implements Runnable {
@Override
public void run() {
if (!engine().possibleMergeNeeded()) {

View File

@ -25,7 +25,9 @@ import com.google.common.cache.CacheBuilder;
import com.google.common.cache.RemovalListener;
import com.google.common.cache.RemovalNotification;
import com.google.common.collect.ImmutableMap;
import gnu.trove.set.hash.THashSet;
import org.elasticsearch.cluster.metadata.MetaData;
import org.elasticsearch.common.CacheRecycler;
import org.elasticsearch.common.collect.MapBuilder;
import org.elasticsearch.common.component.AbstractComponent;
import org.elasticsearch.common.inject.Inject;
@ -33,21 +35,33 @@ import org.elasticsearch.common.lucene.docset.DocSet;
import org.elasticsearch.common.settings.Settings;
import org.elasticsearch.common.unit.ByteSizeValue;
import org.elasticsearch.common.unit.TimeValue;
import org.elasticsearch.common.util.concurrent.ConcurrentCollections;
import org.elasticsearch.index.cache.filter.weighted.WeightedFilterCache;
import org.elasticsearch.monitor.jvm.JvmInfo;
import org.elasticsearch.node.settings.NodeSettingsService;
import org.elasticsearch.threadpool.ThreadPool;
import java.util.Iterator;
import java.util.Map;
import java.util.Set;
import java.util.concurrent.TimeUnit;
public class IndicesFilterCache extends AbstractComponent implements RemovalListener<WeightedFilterCache.FilterCacheKey, DocSet> {
private final ThreadPool threadPool;
private Cache<WeightedFilterCache.FilterCacheKey, DocSet> cache;
private volatile String size;
private volatile long sizeInBytes;
private volatile TimeValue expire;
private final TimeValue cleanInterval;
private final Set<Object> readersKeysToClean = ConcurrentCollections.newConcurrentSet();
private volatile boolean closed;
private volatile Map<String, RemovalListener<WeightedFilterCache.FilterCacheKey, DocSet>> removalListeners =
ImmutableMap.of();
@ -85,15 +99,19 @@ public class IndicesFilterCache extends AbstractComponent implements RemovalList
}
@Inject
public IndicesFilterCache(Settings settings, NodeSettingsService nodeSettingsService) {
public IndicesFilterCache(Settings settings, ThreadPool threadPool, NodeSettingsService nodeSettingsService) {
super(settings);
this.threadPool = threadPool;
this.size = componentSettings.get("size", "20%");
this.expire = componentSettings.getAsTime("expire", null);
this.cleanInterval = componentSettings.getAsTime("clean_interval", TimeValue.timeValueSeconds(1));
computeSizeInBytes();
buildCache();
logger.debug("using [node] filter cache with size [{}], actual_size [{}]", size, new ByteSizeValue(sizeInBytes));
nodeSettingsService.addListener(new ApplySettings());
threadPool.schedule(cleanInterval, ThreadPool.Names.SAME, new ReaderCleaner());
}
private void buildCache() {
@ -102,7 +120,7 @@ public class IndicesFilterCache extends AbstractComponent implements RemovalList
.maximumWeight(sizeInBytes).weigher(new WeightedFilterCache.FilterCacheValueWeigher());
// defaults to 4, but this is a busy map for all indices, increase it a bit
cacheBuilder.concurrencyLevel(8);
cacheBuilder.concurrencyLevel(16);
if (expire != null) {
cacheBuilder.expireAfterAccess(expire.millis(), TimeUnit.MILLISECONDS);
@ -128,7 +146,12 @@ public class IndicesFilterCache extends AbstractComponent implements RemovalList
removalListeners = MapBuilder.newMapBuilder(removalListeners).remove(index).immutableMap();
}
public void addReaderKeyToClean(Object readerKey) {
readersKeysToClean.add(readerKey);
}
public void close() {
closed = true;
cache.invalidateAll();
}
@ -147,4 +170,49 @@ public class IndicesFilterCache extends AbstractComponent implements RemovalList
listener.onRemoval(removalNotification);
}
}
/**
* The reason we need this class ie because we need to clean all the filters that are associated
* with a reader. We don't want to do it every time a reader closes, since iterating over all the map
* is expensive. There doesn't seem to be a nicer way to do it (and maintaining a list per reader
* of the filters will cost more).
*/
class ReaderCleaner implements Runnable {
@Override
public void run() {
if (closed) {
return;
}
if (readersKeysToClean.isEmpty()) {
threadPool.schedule(cleanInterval, ThreadPool.Names.SAME, this);
return;
}
threadPool.executor(ThreadPool.Names.GENERIC).execute(new Runnable() {
@Override
public void run() {
THashSet<Object> keys = CacheRecycler.popHashSet();
try {
for (Iterator<Object> it = readersKeysToClean.iterator(); it.hasNext(); ) {
keys.add(it.next());
it.remove();
}
cache.cleanUp();
if (!keys.isEmpty()) {
for (Iterator<WeightedFilterCache.FilterCacheKey> it = cache.asMap().keySet().iterator(); it.hasNext(); ) {
WeightedFilterCache.FilterCacheKey filterCacheKey = it.next();
if (keys.contains(filterCacheKey.readerKey())) {
// same as invalidate
it.remove();
}
}
}
threadPool.schedule(cleanInterval, ThreadPool.Names.SAME, ReaderCleaner.this);
} finally {
CacheRecycler.pushHashSet(keys);
}
}
});
}
}
}