HBASE-16630 Handle Fragmentation in bucket cache

Currently whenever a compaction/bulkload happen and the
blocks are evicted from theirs buckets the buckets become
fragmented and are not available to be used by other
BucketSizes

Bug Fix : Added Memory block type also to the list of
evictions that need to happen when there is a needForExtra

Improvement : Inorder to fix the non availabilty of Buckets and force
the movement of buckets to transformed sizes, whenever we encounter a
situation where an allocation cant be made for a BucketSize, we will
forcefully free the entire buckets that have least occupancy ratio. This
is the same strategy used by MemCached when they encounter a similar
issue going by the name 'Slab Calcification'. Only improvement is that
we use a heuristic to evict from the buckets that are least occupied
and also avoid the BucketSizes where there is a single Bucket

Change-Id: I9e3b4deb8d893953003ddf5f1e66312ed97ea9cb

Signed-off-by: Ramkrishna <ramkrishna.s.vasudevan@intel.com>

hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/bucket/BucketAllocator.java

@@ -20,13 +20,16 @@
 
 package org.apache.hadoop.hbase.io.hfile.bucket;
 
-import java.util.ArrayList;
 import java.util.Arrays;
+import java.util.Comparator;
+import java.util.HashSet;
 import java.util.Iterator;
-import java.util.List;
 import java.util.Map;
+import java.util.Queue;
+import java.util.Set;
 import java.util.concurrent.atomic.AtomicLong;
 
+import com.google.common.collect.MinMaxPriorityQueue;
 import org.apache.commons.collections.map.LinkedMap;
 import org.apache.commons.logging.Log;
 import org.apache.commons.logging.LogFactory;
@@ -583,4 +586,45 @@ public final class BucketAllocator {
     return sz;
   }
 
+  public int getBucketIndex(long offset) {
+    return (int) (offset / bucketCapacity);
+  }
+
+  /**
+   * Returns a set of indices of the buckets that are least filled
+   * excluding the offsets, we also the fully free buckets for the
+   * BucketSizes where everything is empty and they only have one
+   * completely free bucket as a reserved
+   *
+   * @param excludedBuckets the buckets that need to be excluded due to
+   *                        currently being in used
+   * @param bucketCount     max Number of buckets to return
+   * @return set of bucket indices which could be used for eviction
+   */
+  public Set<Integer> getLeastFilledBuckets(Set<Integer> excludedBuckets,
+                                            int bucketCount) {
+    Queue<Integer> queue = MinMaxPriorityQueue.<Integer>orderedBy(
+        new Comparator<Integer>() {
+          @Override
+          public int compare(Integer left, Integer right) {
+            // We will always get instantiated buckets
+            return Float.compare(
+                ((float) buckets[left].usedCount) / buckets[left].itemCount,
+                ((float) buckets[right].usedCount) / buckets[right].itemCount);
+          }
+        }).maximumSize(bucketCount).create();
+
+    for (int i = 0; i < buckets.length; i ++ ) {
+      if (!excludedBuckets.contains(i) && !buckets[i].isUninstantiated() &&
+          // Avoid the buckets that are the only buckets for a sizeIndex
+          bucketSizeInfos[buckets[i].sizeIndex()].bucketList.size() != 1) {
+        queue.add(i);
+      }
+    }
+
+    Set<Integer> result = new HashSet<>(bucketCount);
+    result.addAll(queue);
+
+    return result;
+  }
 }

hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/bucket/BucketCache.java

@@ -31,6 +31,7 @@ import java.io.Serializable;
 import java.nio.ByteBuffer;
 import java.util.ArrayList;
 import java.util.Comparator;
+import java.util.HashSet;
 import java.util.Iterator;
 import java.util.List;
 import java.util.Map;
@@ -107,6 +108,9 @@ public class BucketCache implements BlockCache, HeapSize {
   private static final float DEFAULT_ACCEPT_FACTOR = 0.95f;
   private static final float DEFAULT_MIN_FACTOR = 0.85f;
 
+  // Number of blocks to clear for each of the bucket size that is full
+  private static final int DEFAULT_FREE_ENTIRE_BLOCK_FACTOR = 2;
+
   /** Statistics thread */
   private static final int statThreadPeriod = 5 * 60;
 
@@ -629,6 +633,53 @@ public class BucketCache implements BlockCache, HeapSize {
         * DEFAULT_MIN_FACTOR);
   }
 
+  /**
+   * Return the count of bucketSizeinfos still needf ree space
+   */
+  private int bucketSizesAboveThresholdCount(float minFactor) {
+    BucketAllocator.IndexStatistics[] stats = bucketAllocator.getIndexStatistics();
+    int fullCount = 0;
+    for (int i = 0; i < stats.length; i++) {
+      long freeGoal = (long) Math.floor(stats[i].totalCount() * (1 - minFactor));
+      freeGoal = Math.max(freeGoal, 1);
+      if (stats[i].freeCount() < freeGoal) {
+        fullCount++;
+      }
+    }
+    return fullCount;
+  }
+
+  /**
+   * This method will find the buckets that are minimally occupied
+   * and are not reference counted and will free them completely
+   * without any constraint on the access times of the elements,
+   * and as a process will completely free at most the number of buckets
+   * passed, sometimes it might not due to changing refCounts
+   *
+   * @param completelyFreeBucketsNeeded number of buckets to free
+   **/
+  private void freeEntireBuckets(int completelyFreeBucketsNeeded) {
+    if (completelyFreeBucketsNeeded != 0) {
+      // First we will build a set where the offsets are reference counted, usually
+      // this set is small around O(Handler Count) unless something else is wrong
+      Set<Integer> inUseBuckets = new HashSet<Integer>();
+      for (BucketEntry entry : backingMap.values()) {
+        if (entry.refCount.get() != 0) {
+          inUseBuckets.add(bucketAllocator.getBucketIndex(entry.offset()));
+        }
+      }
+
+      Set<Integer> candidateBuckets = bucketAllocator.getLeastFilledBuckets(
+          inUseBuckets, completelyFreeBucketsNeeded);
+      for (Map.Entry<BlockCacheKey, BucketEntry> entry : backingMap.entrySet()) {
+        if (candidateBuckets.contains(bucketAllocator
+            .getBucketIndex(entry.getValue().offset()))) {
+          evictBlock(entry.getKey(), false);
+        }
+      }
+    }
+  }
+
   /**
    * Free the space if the used size reaches acceptableSize() or one size block
    * couldn't be allocated. When freeing the space, we use the LRU algorithm and
@@ -725,27 +776,14 @@ public class BucketCache implements BlockCache, HeapSize {
         remainingBuckets--;
       }
 
-      /**
+      // Check and free if there are buckets that still need freeing of space
-       * Check whether need extra free because some bucketSizeinfo still needs
+      if (bucketSizesAboveThresholdCount(DEFAULT_MIN_FACTOR) > 0) {
-       * free space
-       */
-      stats = bucketAllocator.getIndexStatistics();
-      boolean needFreeForExtra = false;
-      for (int i = 0; i < stats.length; i++) {
-        long freeGoal = (long) Math.floor(stats[i].totalCount() * (1 - DEFAULT_MIN_FACTOR));
-        freeGoal = Math.max(freeGoal, 1);
-        if (stats[i].freeCount() < freeGoal) {
-          needFreeForExtra = true;
-          break;
-        }
-      }
-
-      if (needFreeForExtra) {
         bucketQueue.clear();
-        remainingBuckets = 2;
+        remainingBuckets = 3;
 
         bucketQueue.add(bucketSingle);
         bucketQueue.add(bucketMulti);
+        bucketQueue.add(bucketMemory);
 
         while ((bucketGroup = bucketQueue.poll()) != null) {
           long bucketBytesToFree = (bytesToFreeWithExtra - bytesFreed) / remainingBuckets;
@@ -754,6 +792,14 @@ public class BucketCache implements BlockCache, HeapSize {
         }
       }
 
+      // Even after the above free we might still need freeing because of the
+      // De-fragmentation of the buckets (also called Slab Calcification problem), i.e
+      // there might be some buckets where the occupancy is very sparse and thus are not
+      // yielding the free for the other bucket sizes, the fix for this to evict some
+      // of the buckets, we do this by evicting the buckets that are least fulled
+      freeEntireBuckets(DEFAULT_FREE_ENTIRE_BLOCK_FACTOR *
+          bucketSizesAboveThresholdCount(1.0f));
+
       if (LOG.isDebugEnabled()) {
         long single = bucketSingle.totalSize();
         long multi = bucketMulti.totalSize();