diff --git a/buffer/src/main/java/io/netty/buffer/PoolChunk.java b/buffer/src/main/java/io/netty/buffer/PoolChunk.java
index 93eda84104..d4b952ffda 100644
--- a/buffer/src/main/java/io/netty/buffer/PoolChunk.java
+++ b/buffer/src/main/java/io/netty/buffer/PoolChunk.java
@@ -16,6 +16,10 @@
 
 package io.netty.buffer;
 
+import java.nio.ByteBuffer;
+import java.util.ArrayDeque;
+import java.util.Deque;
+
 /**
  * Description of algorithm for PageRun/PoolSubpage allocation from PoolChunk
  *
@@ -107,7 +111,6 @@ final class PoolChunk<T> implements PoolChunkMetric {
     final T memory;
     final boolean unpooled;
     final int offset;
-
     private final byte[] memoryMap;
     private final byte[] depthMap;
     private final PoolSubpage<T>[] subpages;
@@ -122,6 +125,13 @@ final class PoolChunk<T> implements PoolChunkMetric {
     /** Used to mark memory as unusable */
     private final byte unusable;
 
+    // Use as cache for ByteBuffer created from the memory. These are just duplicates and so are only a container
+    // around the memory itself. These are often needed for operations within the Pooled*DirectByteBuf and so
+    // may produce extra GC, which can be greatly reduced by caching the duplicates.
+    //
+    // This may be null if the PoolChunk is unpooled as pooling the ByteBuffer instances does not make any sense here.
+    private final Deque<ByteBuffer> cachedNioBuffers;
+
     private int freeBytes;
 
     PoolChunkList<T> parent;
@@ -163,6 +173,7 @@ final class PoolChunk<T> implements PoolChunkMetric {
         }
 
         subpages = newSubpageArray(maxSubpageAllocs);
+        cachedNioBuffers = new ArrayDeque<ByteBuffer>();
     }
 
     /** Creates a special chunk that is not pooled. */
@@ -182,6 +193,21 @@ final class PoolChunk<T> implements PoolChunkMetric {
         chunkSize = size;
         log2ChunkSize = log2(chunkSize);
         maxSubpageAllocs = 0;
+        cachedNioBuffers = null;
+    }
+
+    ByteBuffer pollCachedNioBuffer() {
+        // We use LIFO to increase the chance that its still "hot" and so in the CPU / L* cache.
+        return cachedNioBuffers != null ? cachedNioBuffers.pollLast() : null;
+    }
+
+    void offerCachedNioBuffer(ByteBuffer nioBuffer) {
+        // Only cache if we did not reach the limit yet and if its not unpooled.
+        // If we do just drop it on the floor and let the GC collect it.
+        if (cachedNioBuffers != null &&
+                cachedNioBuffers.size() <= PooledByteBufAllocator.DEFAULT_MAX_CACHED_BYTEBUFFERS_PER_CHUNK) {
+            cachedNioBuffers.offer(nioBuffer);
+        }
     }
 
     @SuppressWarnings("unchecked")
diff --git a/buffer/src/main/java/io/netty/buffer/PooledByteBuf.java b/buffer/src/main/java/io/netty/buffer/PooledByteBuf.java
index 56a4be3872..699ca1ca62 100644
--- a/buffer/src/main/java/io/netty/buffer/PooledByteBuf.java
+++ b/buffer/src/main/java/io/netty/buffer/PooledByteBuf.java
@@ -56,13 +56,13 @@ abstract class PooledByteBuf<T> extends AbstractReferenceCountedByteBuf {
 
         this.chunk = chunk;
         memory = chunk.memory;
+        tmpNioBuf = chunk.pollCachedNioBuffer();
         allocator = chunk.arena.parent;
         this.cache = cache;
         this.handle = handle;
         this.offset = offset;
         this.length = length;
         this.maxLength = maxLength;
-        tmpNioBuf = null;
     }
 
     /**
@@ -166,7 +166,11 @@ abstract class PooledByteBuf<T> extends AbstractReferenceCountedByteBuf {
             final long handle = this.handle;
             this.handle = -1;
             memory = null;
-            tmpNioBuf = null;
+            if (tmpNioBuf != null) {
+                // Try to put back into the cache for later usage.
+                chunk.offerCachedNioBuffer(tmpNioBuf);
+                tmpNioBuf = null;
+            }
             chunk.arena.free(chunk, handle, maxLength, cache);
             chunk = null;
             recycle();
diff --git a/buffer/src/main/java/io/netty/buffer/PooledByteBufAllocator.java b/buffer/src/main/java/io/netty/buffer/PooledByteBufAllocator.java
index aa9ee28b94..92bb476fe7 100644
--- a/buffer/src/main/java/io/netty/buffer/PooledByteBufAllocator.java
+++ b/buffer/src/main/java/io/netty/buffer/PooledByteBufAllocator.java
@@ -45,6 +45,7 @@ public class PooledByteBufAllocator extends AbstractByteBufAllocator implements
     private static final int DEFAULT_CACHE_TRIM_INTERVAL;
     private static final boolean DEFAULT_USE_CACHE_FOR_ALL_THREADS;
     private static final int DEFAULT_DIRECT_MEMORY_CACHE_ALIGNMENT;
+    static final int DEFAULT_MAX_CACHED_BYTEBUFFERS_PER_CHUNK;
 
     private static final int MIN_PAGE_SIZE = 4096;
     private static final int MAX_CHUNK_SIZE = (int) (((long) Integer.MAX_VALUE + 1) / 2);
@@ -116,6 +117,9 @@ public class PooledByteBufAllocator extends AbstractByteBufAllocator implements
         DEFAULT_DIRECT_MEMORY_CACHE_ALIGNMENT = SystemPropertyUtil.getInt(
                 "io.netty.allocator.directMemoryCacheAlignment", 0);
 
+        DEFAULT_MAX_CACHED_BYTEBUFFERS_PER_CHUNK = SystemPropertyUtil.getInt(
+                "io.netty.allocator.maxCachedByteBuffersPerChunk", 1024);
+
         if (logger.isDebugEnabled()) {
             logger.debug("-Dio.netty.allocator.numHeapArenas: {}", DEFAULT_NUM_HEAP_ARENA);
             logger.debug("-Dio.netty.allocator.numDirectArenas: {}", DEFAULT_NUM_DIRECT_ARENA);
@@ -136,6 +140,8 @@ public class PooledByteBufAllocator extends AbstractByteBufAllocator implements
             logger.debug("-Dio.netty.allocator.maxCachedBufferCapacity: {}", DEFAULT_MAX_CACHED_BUFFER_CAPACITY);
             logger.debug("-Dio.netty.allocator.cacheTrimInterval: {}", DEFAULT_CACHE_TRIM_INTERVAL);
             logger.debug("-Dio.netty.allocator.useCacheForAllThreads: {}", DEFAULT_USE_CACHE_FOR_ALL_THREADS);
+            logger.debug("-Dio.netty.allocator.maxCachedByteBuffersPerChunk: {}",
+                    DEFAULT_MAX_CACHED_BYTEBUFFERS_PER_CHUNK);
         }
     }