HDFS-5276. FileSystem.Statistics should use thread-local counters to avoid multi-threaded performance issues on read/write. (Colin Patrick McCabe)

git-svn-id: https://svn.apache.org/repos/asf/hadoop/common/trunk@1533668 13f79535-47bb-0310-9956-ffa450edef68
This commit is contained in:
Colin McCabe 2013-10-19 00:17:37 +00:00
parent 2baa42dd01
commit e86f4a2e25
3 changed files with 312 additions and 28 deletions

View File

@ -363,6 +363,9 @@ Release 2.3.0 - UNRELEASED
HADOOP-9078. enhance unit-test coverage of class
org.apache.hadoop.fs.FileContext (Ivan A. Veselovsky via jeagles)
HDFS-5276. FileSystem.Statistics should use thread-local counters to avoid
multi-threaded performance issues on read/write. (Colin Patrick McCabe)
OPTIMIZATIONS
HADOOP-9748. Reduce blocking on UGI.ensureInitialized (daryn)

View File

@ -20,6 +20,7 @@
import java.io.Closeable;
import java.io.FileNotFoundException;
import java.io.IOException;
import java.lang.ref.WeakReference;
import java.net.URI;
import java.net.URISyntaxException;
import java.security.PrivilegedExceptionAction;
@ -31,6 +32,7 @@
import java.util.HashSet;
import java.util.IdentityHashMap;
import java.util.Iterator;
import java.util.LinkedList;
import java.util.List;
import java.util.Map;
import java.util.NoSuchElementException;
@ -2501,110 +2503,69 @@ public String toString() {
}
}
public static final class Statistics {
private final String scheme;
private AtomicLong bytesRead = new AtomicLong();
private AtomicLong bytesWritten = new AtomicLong();
private AtomicInteger readOps = new AtomicInteger();
private AtomicInteger largeReadOps = new AtomicInteger();
private AtomicInteger writeOps = new AtomicInteger();
public Statistics(String scheme) {
this.scheme = scheme;
}
/**
* Copy constructor.
* Tracks statistics about how many reads, writes, and so forth have been
* done in a FileSystem.
*
* @param st
* The input Statistics object which is cloned.
* Since there is only one of these objects per FileSystem, there will
* typically be many threads writing to this object. Almost every operation
* on an open file will involve a write to this object. In contrast, reading
* statistics is done infrequently by most programs, and not at all by others.
* Hence, this is optimized for writes.
*
* Each thread writes to its own thread-local area of memory. This removes
* contention and allows us to scale up to many, many threads. To read
* statistics, the reader thread totals up the contents of all of the
* thread-local data areas.
*/
public Statistics(Statistics st) {
this.scheme = st.scheme;
this.bytesRead = new AtomicLong(st.bytesRead.longValue());
this.bytesWritten = new AtomicLong(st.bytesWritten.longValue());
public static final class Statistics {
/**
* Statistics data.
*
* There is only a single writer to thread-local StatisticsData objects.
* Hence, volatile is adequate here-- we do not need AtomicLong or similar
* to prevent lost updates.
* The Java specification guarantees that updates to volatile longs will
* be perceived as atomic with respect to other threads, which is all we
* need.
*/
private static class StatisticsData {
volatile long bytesRead;
volatile long bytesWritten;
volatile int readOps;
volatile int largeReadOps;
volatile int writeOps;
/**
* Stores a weak reference to the thread owning this StatisticsData.
* This allows us to remove StatisticsData objects that pertain to
* threads that no longer exist.
*/
final WeakReference<Thread> owner;
StatisticsData(WeakReference<Thread> owner) {
this.owner = owner;
}
/**
* Increment the bytes read in the statistics
* @param newBytes the additional bytes read
* Add another StatisticsData object to this one.
*/
public void incrementBytesRead(long newBytes) {
bytesRead.getAndAdd(newBytes);
void add(StatisticsData other) {
this.bytesRead += other.bytesRead;
this.bytesWritten += other.bytesWritten;
this.readOps += other.readOps;
this.largeReadOps += other.largeReadOps;
this.writeOps += other.writeOps;
}
/**
* Increment the bytes written in the statistics
* @param newBytes the additional bytes written
* Negate the values of all statistics.
*/
public void incrementBytesWritten(long newBytes) {
bytesWritten.getAndAdd(newBytes);
}
/**
* Increment the number of read operations
* @param count number of read operations
*/
public void incrementReadOps(int count) {
readOps.getAndAdd(count);
}
/**
* Increment the number of large read operations
* @param count number of large read operations
*/
public void incrementLargeReadOps(int count) {
largeReadOps.getAndAdd(count);
}
/**
* Increment the number of write operations
* @param count number of write operations
*/
public void incrementWriteOps(int count) {
writeOps.getAndAdd(count);
}
/**
* Get the total number of bytes read
* @return the number of bytes
*/
public long getBytesRead() {
return bytesRead.get();
}
/**
* Get the total number of bytes written
* @return the number of bytes
*/
public long getBytesWritten() {
return bytesWritten.get();
}
/**
* Get the number of file system read operations such as list files
* @return number of read operations
*/
public int getReadOps() {
return readOps.get() + largeReadOps.get();
}
/**
* Get the number of large file system read operations such as list files
* under a large directory
* @return number of large read operations
*/
public int getLargeReadOps() {
return largeReadOps.get();
}
/**
* Get the number of file system write operations such as create, append
* rename etc.
* @return number of write operations
*/
public int getWriteOps() {
return writeOps.get();
void negate() {
this.bytesRead = -this.bytesRead;
this.bytesWritten = -this.bytesWritten;
this.readOps = -this.readOps;
this.largeReadOps = -this.largeReadOps;
this.writeOps = -this.writeOps;
}
@Override
@ -2613,13 +2574,299 @@ public String toString() {
+ readOps + " read ops, " + largeReadOps + " large read ops, "
+ writeOps + " write ops";
}
}
private interface StatisticsAggregator<T> {
void accept(StatisticsData data);
T aggregate();
}
private final String scheme;
/**
* Reset the counts of bytes to 0.
* rootData is data that doesn't belong to any thread, but will be added
* to the totals. This is useful for making copies of Statistics objects,
* and for storing data that pertains to threads that have been garbage
* collected. Protected by the Statistics lock.
*/
private final StatisticsData rootData;
/**
* Thread-local data.
*/
private final ThreadLocal<StatisticsData> threadData;
/**
* List of all thread-local data areas. Protected by the Statistics lock.
*/
private LinkedList<StatisticsData> allData;
public Statistics(String scheme) {
this.scheme = scheme;
this.rootData = new StatisticsData(null);
this.threadData = new ThreadLocal<StatisticsData>();
this.allData = null;
}
/**
* Copy constructor.
*
* @param other The input Statistics object which is cloned.
*/
public Statistics(Statistics other) {
this.scheme = other.scheme;
this.rootData = new StatisticsData(null);
other.visitAll(new StatisticsAggregator<Void>() {
@Override
public void accept(StatisticsData data) {
rootData.add(data);
}
public Void aggregate() {
return null;
}
});
this.threadData = new ThreadLocal<StatisticsData>();
}
/**
* Get or create the thread-local data associated with the current thread.
*/
private StatisticsData getThreadData() {
StatisticsData data = threadData.get();
if (data == null) {
data = new StatisticsData(
new WeakReference<Thread>(Thread.currentThread()));
threadData.set(data);
synchronized(this) {
if (allData == null) {
allData = new LinkedList<StatisticsData>();
}
allData.add(data);
}
}
return data;
}
/**
* Increment the bytes read in the statistics
* @param newBytes the additional bytes read
*/
public void incrementBytesRead(long newBytes) {
getThreadData().bytesRead += newBytes;
}
/**
* Increment the bytes written in the statistics
* @param newBytes the additional bytes written
*/
public void incrementBytesWritten(long newBytes) {
getThreadData().bytesWritten += newBytes;
}
/**
* Increment the number of read operations
* @param count number of read operations
*/
public void incrementReadOps(int count) {
getThreadData().readOps += count;
}
/**
* Increment the number of large read operations
* @param count number of large read operations
*/
public void incrementLargeReadOps(int count) {
getThreadData().largeReadOps += count;
}
/**
* Increment the number of write operations
* @param count number of write operations
*/
public void incrementWriteOps(int count) {
getThreadData().writeOps += count;
}
/**
* Apply the given aggregator to all StatisticsData objects associated with
* this Statistics object.
*
* For each StatisticsData object, we will call accept on the visitor.
* Finally, at the end, we will call aggregate to get the final total.
*
* @param The visitor to use.
* @return The total.
*/
private synchronized <T> T visitAll(StatisticsAggregator<T> visitor) {
visitor.accept(rootData);
if (allData != null) {
for (Iterator<StatisticsData> iter = allData.iterator();
iter.hasNext(); ) {
StatisticsData data = iter.next();
visitor.accept(data);
if (data.owner.get() == null) {
/*
* If the thread that created this thread-local data no
* longer exists, remove the StatisticsData from our list
* and fold the values into rootData.
*/
rootData.add(data);
iter.remove();
}
}
}
return visitor.aggregate();
}
/**
* Get the total number of bytes read
* @return the number of bytes
*/
public long getBytesRead() {
return visitAll(new StatisticsAggregator<Long>() {
private long bytesRead = 0;
@Override
public void accept(StatisticsData data) {
bytesRead += data.bytesRead;
}
public Long aggregate() {
return bytesRead;
}
});
}
/**
* Get the total number of bytes written
* @return the number of bytes
*/
public long getBytesWritten() {
return visitAll(new StatisticsAggregator<Long>() {
private long bytesWritten = 0;
@Override
public void accept(StatisticsData data) {
bytesWritten += data.bytesWritten;
}
public Long aggregate() {
return bytesWritten;
}
});
}
/**
* Get the number of file system read operations such as list files
* @return number of read operations
*/
public int getReadOps() {
return visitAll(new StatisticsAggregator<Integer>() {
private int readOps = 0;
@Override
public void accept(StatisticsData data) {
readOps += data.readOps;
readOps += data.largeReadOps;
}
public Integer aggregate() {
return readOps;
}
});
}
/**
* Get the number of large file system read operations such as list files
* under a large directory
* @return number of large read operations
*/
public int getLargeReadOps() {
return visitAll(new StatisticsAggregator<Integer>() {
private int largeReadOps = 0;
@Override
public void accept(StatisticsData data) {
largeReadOps += data.largeReadOps;
}
public Integer aggregate() {
return largeReadOps;
}
});
}
/**
* Get the number of file system write operations such as create, append
* rename etc.
* @return number of write operations
*/
public int getWriteOps() {
return visitAll(new StatisticsAggregator<Integer>() {
private int writeOps = 0;
@Override
public void accept(StatisticsData data) {
writeOps += data.writeOps;
}
public Integer aggregate() {
return writeOps;
}
});
}
@Override
public String toString() {
return visitAll(new StatisticsAggregator<String>() {
private StatisticsData total = new StatisticsData(null);
@Override
public void accept(StatisticsData data) {
total.add(data);
}
public String aggregate() {
return total.toString();
}
});
}
/**
* Resets all statistics to 0.
*
* In order to reset, we add up all the thread-local statistics data, and
* set rootData to the negative of that.
*
* This may seem like a counterintuitive way to reset the statsitics. Why
* can't we just zero out all the thread-local data? Well, thread-local
* data can only be modified by the thread that owns it. If we tried to
* modify the thread-local data from this thread, our modification might get
* interleaved with a read-modify-write operation done by the thread that
* owns the data. That would result in our update getting lost.
*
* The approach used here avoids this problem because it only ever reads
* (not writes) the thread-local data. Both reads and writes to rootData
* are done under the lock, so we're free to modify rootData from any thread
* that holds the lock.
*/
public void reset() {
bytesWritten.set(0);
bytesRead.set(0);
visitAll(new StatisticsAggregator<Void>() {
private StatisticsData total = new StatisticsData(null);
@Override
public void accept(StatisticsData data) {
total.add(data);
}
public Void aggregate() {
total.negate();
rootData.add(total);
return null;
}
});
}
/**

View File

@ -27,6 +27,8 @@
import org.junit.Assert;
import org.junit.Test;
import com.google.common.util.concurrent.Uninterruptibles;
import static org.apache.hadoop.fs.FileContextTestHelper.*;
/**
@ -44,6 +46,38 @@ public abstract class FCStatisticsBaseTest {
//fc should be set appropriately by the deriving test.
protected static FileContext fc = null;
@Test(timeout=60000)
public void testStatisticsOperations() throws Exception {
final Statistics stats = new Statistics("file");
Assert.assertEquals(0L, stats.getBytesRead());
Assert.assertEquals(0L, stats.getBytesWritten());
Assert.assertEquals(0, stats.getWriteOps());
stats.incrementBytesWritten(1000);
Assert.assertEquals(1000L, stats.getBytesWritten());
Assert.assertEquals(0, stats.getWriteOps());
stats.incrementWriteOps(123);
Assert.assertEquals(123, stats.getWriteOps());
Thread thread = new Thread() {
@Override
public void run() {
stats.incrementWriteOps(1);
}
};
thread.start();
Uninterruptibles.joinUninterruptibly(thread);
Assert.assertEquals(124, stats.getWriteOps());
// Test copy constructor and reset function
Statistics stats2 = new Statistics(stats);
stats.reset();
Assert.assertEquals(0, stats.getWriteOps());
Assert.assertEquals(0L, stats.getBytesWritten());
Assert.assertEquals(0L, stats.getBytesRead());
Assert.assertEquals(124, stats2.getWriteOps());
Assert.assertEquals(1000L, stats2.getBytesWritten());
Assert.assertEquals(0L, stats2.getBytesRead());
}
@Test
public void testStatistics() throws IOException, URISyntaxException {
URI fsUri = getFsUri();