/** Create and return a new MergeThread */
protected synchronized MergeThread getMergeThread(IndexWriter writer, OneMerge merge)
throws IOException {
final MergeThread thread = new MergeThread(writer, merge);
thread.setDaemon(true);
thread.setName("Lucene Merge Thread #" + mergeThreadCount++);
return thread;
}
/**
* Wait for any running merge threads to finish. This call is not interruptible as used by {@link
* #close()}.
*/
public void sync() {
boolean interrupted = false;
try {
while (true) {
MergeThread toSync = null;
synchronized (this) {
for (MergeThread t : mergeThreads) {
if (t.isAlive()) {
toSync = t;
break;
}
}
}
if (toSync != null) {
try {
toSync.join();
} catch (InterruptedException ie) {
// ignore this Exception, we will retry until all threads are dead
interrupted = true;
}
} else {
break;
}
}
} finally {
// finally, restore interrupt status:
if (interrupted) Thread.currentThread().interrupt();
}
}
/**
* Returns the number of merge threads that are alive, ignoring the calling thread if it is a
* merge thread. Note that this number is ≤ {@link #mergeThreads} size.
*
* @lucene.internal
*/
public synchronized int mergeThreadCount() {
Thread currentThread = Thread.currentThread();
int count = 0;
for (MergeThread mergeThread : mergeThreads) {
if (currentThread != mergeThread
&& mergeThread.isAlive()
&& mergeThread.merge.rateLimiter.getAbort() == false) {
count++;
}
}
return count;
}
private boolean isBacklog(long now, OneMerge merge) {
double mergeMB = bytesToMB(merge.estimatedMergeBytes);
for (MergeThread mergeThread : mergeThreads) {
long mergeStartNS = mergeThread.merge.mergeStartNS;
if (mergeThread.isAlive()
&& mergeThread.merge != merge
&& mergeStartNS != -1
&& mergeThread.merge.estimatedMergeBytes >= MIN_BIG_MERGE_MB * 1024 * 1024
&& nsToSec(now - mergeStartNS) > 3.0) {
double otherMergeMB = bytesToMB(mergeThread.merge.estimatedMergeBytes);
double ratio = otherMergeMB / mergeMB;
if (ratio > 0.3 && ratio < 3.0) {
return true;
}
}
}
return false;
}
@Override
public synchronized void merge(IndexWriter writer, MergeTrigger trigger, boolean newMergesFound)
throws IOException {
assert !Thread.holdsLock(writer);
initDynamicDefaults(writer);
if (trigger == MergeTrigger.CLOSING) {
// Disable throttling on close:
targetMBPerSec = MAX_MERGE_MB_PER_SEC;
updateMergeThreads();
}
// First, quickly run through the newly proposed merges
// and add any orthogonal merges (ie a merge not
// involving segments already pending to be merged) to
// the queue. If we are way behind on merging, many of
// these newly proposed merges will likely already be
// registered.
if (verbose()) {
message("now merge");
message(" index: " + writer.segString());
}
// Iterate, pulling from the IndexWriter's queue of
// pending merges, until it's empty:
while (true) {
if (maybeStall(writer) == false) {
break;
}
OneMerge merge = writer.getNextMerge();
if (merge == null) {
if (verbose()) {
message(" no more merges pending; now return");
}
return;
}
updateIOThrottle(merge);
boolean success = false;
try {
if (verbose()) {
message(" consider merge " + writer.segString(merge.segments));
}
// OK to spawn a new merge thread to handle this
// merge:
final MergeThread merger = getMergeThread(writer, merge);
mergeThreads.add(merger);
if (verbose()) {
message(" launch new thread [" + merger.getName() + "]");
}
merger.start();
updateMergeThreads();
success = true;
} finally {
if (!success) {
writer.mergeFinish(merge);
}
}
}
}
/**
* Called whenever the running merges have changed, to set merge IO limits. This method sorts the
* merge threads by their merge size in descending order and then pauses/unpauses threads from
* first to last -- that way, smaller merges are guaranteed to run before larger ones.
*/
protected synchronized void updateMergeThreads() {
// Only look at threads that are alive & not in the
// process of stopping (ie have an active merge):
final List<MergeThread> activeMerges = new ArrayList<>();
int threadIdx = 0;
while (threadIdx < mergeThreads.size()) {
final MergeThread mergeThread = mergeThreads.get(threadIdx);
if (!mergeThread.isAlive()) {
// Prune any dead threads
mergeThreads.remove(threadIdx);
continue;
}
activeMerges.add(mergeThread);
threadIdx++;
}
// Sort the merge threads, largest first:
CollectionUtil.timSort(activeMerges);
final int activeMergeCount = activeMerges.size();
int bigMergeCount = 0;
for (threadIdx = activeMergeCount - 1; threadIdx >= 0; threadIdx--) {
MergeThread mergeThread = activeMerges.get(threadIdx);
if (mergeThread.merge.estimatedMergeBytes > MIN_BIG_MERGE_MB * 1024 * 1024) {
bigMergeCount = 1 + threadIdx;
break;
}
}
long now = System.nanoTime();
StringBuilder message;
if (verbose()) {
message = new StringBuilder();
message.append(
String.format(
Locale.ROOT,
"updateMergeThreads ioThrottle=%s targetMBPerSec=%.1f MB/sec",
doAutoIOThrottle,
targetMBPerSec));
} else {
message = null;
}
for (threadIdx = 0; threadIdx < activeMergeCount; threadIdx++) {
MergeThread mergeThread = activeMerges.get(threadIdx);
OneMerge merge = mergeThread.merge;
// pause the thread if maxThreadCount is smaller than the number of merge threads.
final boolean doPause = threadIdx < bigMergeCount - maxThreadCount;
double newMBPerSec;
if (doPause) {
newMBPerSec = 0.0;
} else if (merge.maxNumSegments != -1) {
newMBPerSec = forceMergeMBPerSec;
} else if (doAutoIOThrottle == false) {
newMBPerSec = Double.POSITIVE_INFINITY;
} else if (merge.estimatedMergeBytes < MIN_BIG_MERGE_MB * 1024 * 1024) {
// Don't rate limit small merges:
newMBPerSec = Double.POSITIVE_INFINITY;
} else {
newMBPerSec = targetMBPerSec;
}
double curMBPerSec = merge.rateLimiter.getMBPerSec();
if (verbose()) {
long mergeStartNS = merge.mergeStartNS;
if (mergeStartNS == -1) {
// IndexWriter didn't start the merge yet:
mergeStartNS = now;
}
message.append('\n');
message.append(
String.format(
Locale.ROOT,
"merge thread %s estSize=%.1f MB (written=%.1f MB) runTime=%.1fs (stopped=%.1fs, paused=%.1fs) rate=%s\n",
mergeThread.getName(),
bytesToMB(merge.estimatedMergeBytes),
bytesToMB(merge.rateLimiter.totalBytesWritten),
nsToSec(now - mergeStartNS),
nsToSec(merge.rateLimiter.getTotalStoppedNS()),
nsToSec(merge.rateLimiter.getTotalPausedNS()),
rateToString(merge.rateLimiter.getMBPerSec())));
if (newMBPerSec != curMBPerSec) {
if (newMBPerSec == 0.0) {
message.append(" now stop");
} else if (curMBPerSec == 0.0) {
if (newMBPerSec == Double.POSITIVE_INFINITY) {
message.append(" now resume");
} else {
message.append(
String.format(Locale.ROOT, " now resume to %.1f MB/sec", newMBPerSec));
}
} else {
message.append(
String.format(
Locale.ROOT,
" now change from %.1f MB/sec to %.1f MB/sec",
curMBPerSec,
newMBPerSec));
}
} else if (curMBPerSec == 0.0) {
message.append(" leave stopped");
} else {
message.append(String.format(Locale.ROOT, " leave running at %.1f MB/sec", curMBPerSec));
}
}
merge.rateLimiter.setMBPerSec(newMBPerSec);
}
if (verbose()) {
message(message.toString());
}
}
@Test
public void testInMemoryAndOnDiskMerger() throws Throwable {
JobID jobId = new JobID("a", 0);
TaskAttemptID reduceId1 = new TaskAttemptID(new TaskID(jobId, TaskType.REDUCE, 0), 0);
TaskAttemptID mapId1 = new TaskAttemptID(new TaskID(jobId, TaskType.MAP, 1), 0);
TaskAttemptID mapId2 = new TaskAttemptID(new TaskID(jobId, TaskType.MAP, 2), 0);
LocalDirAllocator lda = new LocalDirAllocator(MRConfig.LOCAL_DIR);
MergeManagerImpl<Text, Text> mergeManager =
new MergeManagerImpl<Text, Text>(
reduceId1,
jobConf,
fs,
lda,
Reporter.NULL,
null,
null,
null,
null,
null,
null,
null,
new Progress(),
new MROutputFiles());
// write map outputs
Map<String, String> map1 = new TreeMap<String, String>();
map1.put("apple", "disgusting");
map1.put("carrot", "delicious");
Map<String, String> map2 = new TreeMap<String, String>();
map1.put("banana", "pretty good");
byte[] mapOutputBytes1 = writeMapOutput(conf, map1);
byte[] mapOutputBytes2 = writeMapOutput(conf, map2);
InMemoryMapOutput<Text, Text> mapOutput1 =
new InMemoryMapOutput<Text, Text>(
conf, mapId1, mergeManager, mapOutputBytes1.length, null, true);
InMemoryMapOutput<Text, Text> mapOutput2 =
new InMemoryMapOutput<Text, Text>(
conf, mapId2, mergeManager, mapOutputBytes2.length, null, true);
System.arraycopy(mapOutputBytes1, 0, mapOutput1.getMemory(), 0, mapOutputBytes1.length);
System.arraycopy(mapOutputBytes2, 0, mapOutput2.getMemory(), 0, mapOutputBytes2.length);
// create merger and run merge
MergeThread<InMemoryMapOutput<Text, Text>, Text, Text> inMemoryMerger =
mergeManager.createInMemoryMerger();
List<InMemoryMapOutput<Text, Text>> mapOutputs1 =
new ArrayList<InMemoryMapOutput<Text, Text>>();
mapOutputs1.add(mapOutput1);
mapOutputs1.add(mapOutput2);
inMemoryMerger.merge(mapOutputs1);
Assert.assertEquals(1, mergeManager.onDiskMapOutputs.size());
TaskAttemptID reduceId2 = new TaskAttemptID(new TaskID(jobId, TaskType.REDUCE, 3), 0);
TaskAttemptID mapId3 = new TaskAttemptID(new TaskID(jobId, TaskType.MAP, 4), 0);
TaskAttemptID mapId4 = new TaskAttemptID(new TaskID(jobId, TaskType.MAP, 5), 0);
// write map outputs
Map<String, String> map3 = new TreeMap<String, String>();
map3.put("apple", "awesome");
map3.put("carrot", "amazing");
Map<String, String> map4 = new TreeMap<String, String>();
map4.put("banana", "bla");
byte[] mapOutputBytes3 = writeMapOutput(conf, map3);
byte[] mapOutputBytes4 = writeMapOutput(conf, map4);
InMemoryMapOutput<Text, Text> mapOutput3 =
new InMemoryMapOutput<Text, Text>(
conf, mapId3, mergeManager, mapOutputBytes3.length, null, true);
InMemoryMapOutput<Text, Text> mapOutput4 =
new InMemoryMapOutput<Text, Text>(
conf, mapId4, mergeManager, mapOutputBytes4.length, null, true);
System.arraycopy(mapOutputBytes3, 0, mapOutput3.getMemory(), 0, mapOutputBytes3.length);
System.arraycopy(mapOutputBytes4, 0, mapOutput4.getMemory(), 0, mapOutputBytes4.length);
// // create merger and run merge
MergeThread<InMemoryMapOutput<Text, Text>, Text, Text> inMemoryMerger2 =
mergeManager.createInMemoryMerger();
List<InMemoryMapOutput<Text, Text>> mapOutputs2 =
new ArrayList<InMemoryMapOutput<Text, Text>>();
mapOutputs2.add(mapOutput3);
mapOutputs2.add(mapOutput4);
inMemoryMerger2.merge(mapOutputs2);
Assert.assertEquals(2, mergeManager.onDiskMapOutputs.size());
List<CompressAwarePath> paths = new ArrayList<CompressAwarePath>();
Iterator<CompressAwarePath> iterator = mergeManager.onDiskMapOutputs.iterator();
List<String> keys = new ArrayList<String>();
List<String> values = new ArrayList<String>();
while (iterator.hasNext()) {
CompressAwarePath next = iterator.next();
readOnDiskMapOutput(conf, fs, next, keys, values);
paths.add(next);
}
Assert.assertEquals(
keys, Arrays.asList("apple", "banana", "carrot", "apple", "banana", "carrot"));
Assert.assertEquals(
values,
Arrays.asList("awesome", "bla", "amazing", "disgusting", "pretty good", "delicious"));
mergeManager.close();
mergeManager =
new MergeManagerImpl<Text, Text>(
reduceId2,
jobConf,
fs,
lda,
Reporter.NULL,
null,
null,
null,
null,
null,
null,
null,
new Progress(),
new MROutputFiles());
MergeThread<CompressAwarePath, Text, Text> onDiskMerger = mergeManager.createOnDiskMerger();
onDiskMerger.merge(paths);
Assert.assertEquals(1, mergeManager.onDiskMapOutputs.size());
keys = new ArrayList<String>();
values = new ArrayList<String>();
readOnDiskMapOutput(conf, fs, mergeManager.onDiskMapOutputs.iterator().next(), keys, values);
Assert.assertEquals(
keys, Arrays.asList("apple", "apple", "banana", "banana", "carrot", "carrot"));
Assert.assertEquals(
values,
Arrays.asList("awesome", "disgusting", "pretty good", "bla", "amazing", "delicious"));
mergeManager.close();
Assert.assertEquals(0, mergeManager.inMemoryMapOutputs.size());
Assert.assertEquals(0, mergeManager.inMemoryMergedMapOutputs.size());
Assert.assertEquals(0, mergeManager.onDiskMapOutputs.size());
}
