HDDS-3892. Datanode initialization is too slow when there are thousan…

hadoop-hdds/container-service/src/main/java/org/apache/hadoop/ozone/container/common/utils/ContainerCache.java

@@ -23,6 +23,7 @@
 import java.util.concurrent.locks.Lock;
 import java.util.concurrent.locks.ReentrantLock;
 
+import com.google.common.util.concurrent.Striped;
 import org.apache.hadoop.hdds.conf.ConfigurationSource;
 import org.apache.hadoop.hdds.utils.MetadataStore;
 import org.apache.hadoop.hdds.utils.MetadataStoreBuilder;
@@ -43,6 +44,7 @@ public final class ContainerCache extends LRUMap {
   private final Lock lock = new ReentrantLock();
   private static ContainerCache cache;
   private static final float LOAD_FACTOR = 0.75f;
+  private final Striped<Lock> rocksDBLock = Striped.lazyWeakLock(1024);
   /**
    * Constructs a cache that holds DBHandle references.
    */
@@ -117,30 +119,57 @@ public ReferenceCountedDB getDB(long containerID, String containerDBType,
       throws IOException {
     Preconditions.checkState(containerID >= 0,
         "Container ID cannot be negative.");
-    lock.lock();
+    ReferenceCountedDB db;
+    Lock containerLock = rocksDBLock.get(containerDBPath);
+    containerLock.lock();
     try {
-      ReferenceCountedDB db = (ReferenceCountedDB) this.get(containerDBPath);
+      lock.lock();
+      try {
+        db = (ReferenceCountedDB) this.get(containerDBPath);
+        if (db != null) {
+          db.incrementReference();
+          return db;
+        }
+      } finally {
+        lock.unlock();
+      }
 
-      if (db == null) {
+      try {
         MetadataStore metadataStore =
             MetadataStoreBuilder.newBuilder()
-            .setDbFile(new File(containerDBPath))
-            .setCreateIfMissing(false)
-            .setConf(conf)
-            .setDBType(containerDBType)
-            .build();
+                .setDbFile(new File(containerDBPath))
+                .setCreateIfMissing(false)
+                .setConf(conf)
+                .setDBType(containerDBType)
+                .build();
         db = new ReferenceCountedDB(metadataStore, containerDBPath);
-        this.put(containerDBPath, db);
+      } catch (Exception e) {
+        LOG.error("Error opening DB. Container:{} ContainerPath:{}",
+            containerID, containerDBPath, e);
+        throw e;
+      }
+
+      lock.lock();
+      try {
+        ReferenceCountedDB currentDB =
+            (ReferenceCountedDB) this.get(containerDBPath);
+        if (currentDB != null) {
+          // increment the reference before returning the object
+          currentDB.incrementReference();
+          // clean the db created in previous step
+          db.cleanup();
+          return currentDB;
+        } else {
+          this.put(containerDBPath, db);
+          // increment the reference before returning the object
+          db.incrementReference();
+          return db;
+        }
+      } finally {
+        lock.unlock();
       }
-      // increment the reference before returning the object
-      db.incrementReference();
-      return db;
-    } catch (Exception e) {
-      LOG.error("Error opening DB. Container:{} ContainerPath:{}",
-          containerID, containerDBPath, e);
-      throw e;
     } finally {
-      lock.unlock();
+      containerLock.unlock();
     }
   }
 

hadoop-hdds/container-service/src/main/java/org/apache/hadoop/ozone/container/ozoneimpl/ContainerReader.java

@@ -120,6 +120,7 @@ public boolean accept(File pathname) {
       return;
     }
 
+    LOG.info("Start to verify containers on volume {}", hddsVolumeRootDir);
     for (File scmLoc : scmDir) {
       File currentDir = new File(scmLoc, Storage.STORAGE_DIR_CURRENT);
       File[] containerTopDirs = currentDir.listFiles();
@@ -144,6 +145,7 @@ public boolean accept(File pathname) {
         }
       }
     }
+    LOG.info("Finish verifying containers on volume {}", hddsVolumeRootDir);
   }
 
   private void verifyContainerFile(long containerID, File containerFile) {

hadoop-hdds/container-service/src/main/java/org/apache/hadoop/ozone/container/ozoneimpl/OzoneContainer.java

@@ -162,6 +162,7 @@ private void buildContainerSet() {
     Iterator<HddsVolume> volumeSetIterator = volumeSet.getVolumesList()
         .iterator();
     ArrayList<Thread> volumeThreads = new ArrayList<Thread>();
+    long startTime = System.currentTimeMillis();
 
     //TODO: diskchecker should be run before this, to see how disks are.
     // And also handle disk failure tolerance need to be added
@@ -182,6 +183,8 @@ private void buildContainerSet() {
       Thread.currentThread().interrupt();
     }
 
+    LOG.info("Build ContainerSet costs {}s",
+        (System.currentTimeMillis() - startTime) / 1000);
   }
 
   /**

hadoop-hdds/container-service/src/test/java/org/apache/hadoop/ozone/container/common/TestContainerCache.java

@@ -31,6 +31,13 @@
 import org.junit.rules.ExpectedException;
 
 import java.io.File;
+import java.io.IOException;
+import java.util.ArrayList;
+import java.util.List;
+import java.util.concurrent.ExecutionException;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+import java.util.concurrent.Future;
 
 
 /**
@@ -63,6 +70,8 @@ public void testContainerCacheEviction() throws Exception {
     conf.setInt(OzoneConfigKeys.OZONE_CONTAINER_CACHE_SIZE, 2);
 
     ContainerCache cache = ContainerCache.getInstance(conf);
+    cache.clear();
+    Assert.assertTrue(cache.size() == 0);
     File containerDir1 = new File(root, "cont1");
     File containerDir2 = new File(root, "cont2");
     File containerDir3 = new File(root, "cont3");
@@ -123,4 +132,47 @@ public void testContainerCacheEviction() throws Exception {
     thrown.expect(IllegalArgumentException.class);
     db5.close();
   }
+
+  @Test
+  public void testConcurrentDBGet() throws Exception {
+    File root = new File(testRoot);
+    root.mkdirs();
+    root.deleteOnExit();
+
+    OzoneConfiguration conf = new OzoneConfiguration();
+    conf.setInt(OzoneConfigKeys.OZONE_CONTAINER_CACHE_SIZE, 2);
+    ContainerCache cache = ContainerCache.getInstance(conf);
+    cache.clear();
+    Assert.assertTrue(cache.size() == 0);
+    File containerDir = new File(root, "cont1");
+    createContainerDB(conf, containerDir);
+    ExecutorService executorService = Executors.newFixedThreadPool(2);
+    Runnable task = () -> {
+      try {
+        ReferenceCountedDB db1 = cache.getDB(1, "RocksDB",
+            containerDir.getPath(), conf);
+        Assert.assertTrue(db1 != null);
+      } catch (IOException e) {
+        Assert.fail("Should get the DB instance");
+      }
+    };
+    List<Future> futureList = new ArrayList<>();
+    futureList.add(executorService.submit(task));
+    futureList.add(executorService.submit(task));
+    for (Future future: futureList) {
+      try {
+        future.get();
+      } catch (InterruptedException| ExecutionException e) {
+        Assert.fail("Should get the DB instance");
+      }
+    }
+
+    ReferenceCountedDB db = cache.getDB(1, "RocksDB",
+        containerDir.getPath(), conf);
+    db.close();
+    db.close();
+    db.close();
+    Assert.assertTrue(cache.size() == 1);
+    db.cleanup();
+  }
 }

hadoop-hdds/container-service/src/test/java/org/apache/hadoop/ozone/container/ozoneimpl/TestContainerReader.java

@@ -22,15 +22,16 @@
 import org.apache.hadoop.conf.StorageUnit;
 import org.apache.hadoop.hdds.StringUtils;
 import org.apache.hadoop.hdds.client.BlockID;
-import org.apache.hadoop.hdds.conf.ConfigurationSource;
 import org.apache.hadoop.hdds.conf.OzoneConfiguration;
 import org.apache.hadoop.hdds.protocol.datanode.proto.ContainerProtos;
+import org.apache.hadoop.hdds.scm.ScmConfigKeys;
 import org.apache.hadoop.ozone.OzoneConsts;
 import org.apache.hadoop.ozone.container.common.helpers.BlockData;
 import org.apache.hadoop.ozone.container.common.helpers.ChunkInfo;
 import org.apache.hadoop.ozone.container.common.impl.ChunkLayOutVersion;
 import org.apache.hadoop.ozone.container.common.impl.ContainerSet;
 import org.apache.hadoop.ozone.container.common.interfaces.Container;
+import org.apache.hadoop.ozone.container.common.utils.ContainerCache;
 import org.apache.hadoop.ozone.container.common.utils.ReferenceCountedDB;
 import org.apache.hadoop.ozone.container.common.volume.HddsVolume;
 import org.apache.hadoop.ozone.container.common.volume.MutableVolumeSet;
@@ -68,7 +69,7 @@ public class TestContainerReader {
   private MutableVolumeSet volumeSet;
   private HddsVolume hddsVolume;
   private ContainerSet containerSet;
-  private ConfigurationSource conf;
+  private OzoneConfiguration conf;
 
 
   private RoundRobinVolumeChoosingPolicy volumeChoosingPolicy;
@@ -219,4 +220,68 @@ public void testContainerReader() throws Exception {
           keyValueContainerData.getNumPendingDeletionBlocks());
     }
   }
+
+  @Test
+  public void testMultipleContainerReader() throws Exception {
+    final int volumeNum = 10;
+    StringBuffer datanodeDirs = new StringBuffer();
+    File[] volumeDirs = new File[volumeNum];
+    for (int i = 0; i < volumeNum; i++) {
+      volumeDirs[i] = tempDir.newFolder();
+      datanodeDirs = datanodeDirs.append(volumeDirs[i]).append(",");
+    }
+    conf.set(ScmConfigKeys.HDDS_DATANODE_DIR_KEY,
+        datanodeDirs.toString());
+    MutableVolumeSet volumeSets =
+        new MutableVolumeSet(datanodeId.toString(), conf);
+    ContainerCache cache = ContainerCache.getInstance(conf);
+    cache.clear();
+
+    RoundRobinVolumeChoosingPolicy policy =
+        new RoundRobinVolumeChoosingPolicy();
+
+    final int containerCount = 100;
+    blockCount = containerCount;
+    for (int i = 0; i < containerCount; i++) {
+      KeyValueContainerData keyValueContainerData =
+          new KeyValueContainerData(i, ChunkLayOutVersion.FILE_PER_BLOCK,
+              (long) StorageUnit.GB.toBytes(5), UUID.randomUUID().toString(),
+              datanodeId.toString());
+
+      KeyValueContainer keyValueContainer =
+          new KeyValueContainer(keyValueContainerData,
+              conf);
+      keyValueContainer.create(volumeSets, policy, scmId);
+
+      List<Long> blkNames;
+      if (i % 2 == 0) {
+        blkNames = addBlocks(keyValueContainer, true);
+        markBlocksForDelete(keyValueContainer, true, blkNames, i);
+      } else {
+        blkNames = addBlocks(keyValueContainer, false);
+        markBlocksForDelete(keyValueContainer, false, blkNames, i);
+      }
+    }
+
+    List<HddsVolume> hddsVolumes = volumeSets.getVolumesList();
+    ContainerReader[] containerReaders = new ContainerReader[volumeNum];
+    Thread[] threads = new Thread[volumeNum];
+    for (int i = 0; i < volumeNum; i++) {
+      containerReaders[i] = new ContainerReader(volumeSets,
+          hddsVolumes.get(i), containerSet, conf);
+      threads[i] = new Thread(containerReaders[i]);
+    }
+    long startTime = System.currentTimeMillis();
+    for (int i = 0; i < volumeNum; i++) {
+      threads[i].start();
+    }
+    for (int i = 0; i < volumeNum; i++) {
+      threads[i].join();
+    }
+    System.out.println("Open " + volumeNum + " Volume with " + containerCount +
+        " costs " + (System.currentTimeMillis() - startTime) / 1000 + "s");
+    Assert.assertTrue(
+        containerSet.getContainerMap().entrySet().size() == containerCount);
+    Assert.assertTrue(cache.size() == containerCount);
+  }
 }