diff --git a/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt b/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt
index 6ce1558d04..66efd7e9b4 100644
--- a/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt
+++ b/hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt
@@ -132,6 +132,9 @@ Trunk (unreleased changes)
     HDFS-2878. Fix TestBlockRecovery and move it back into main test directory.
     (todd)
 
+    HDFS-2655. BlockReaderLocal#skip performs unnecessary IO. (Brandon Li
+    via jitendra) 
+
   OPTIMIZATIONS
     HDFS-2477. Optimize computing the diff between a block report and the
     namenode state. (Tomasz Nykiel via hairong)
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/BlockReaderLocal.java b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/BlockReaderLocal.java
index 3090419fb0..cc61697ede 100644
--- a/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/BlockReaderLocal.java
+++ b/hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/BlockReaderLocal.java
@@ -369,26 +369,68 @@ public synchronized long skip(long n) throws IOException {
     if (LOG.isDebugEnabled()) {
       LOG.debug("skip " + n);
     }
+    if (n <= 0) {
+      return 0;
+    }
     if (!verifyChecksum) {
       return dataIn.skip(n);
     }
-    // Skip by reading the data so we stay in sync with checksums.
-    // This could be implemented more efficiently in the future to
-    // skip to the beginning of the appropriate checksum chunk
-    // and then only read to the middle of that chunk.
+  
+    // caller made sure newPosition is not beyond EOF.
+    int remaining = dataBuff.remaining();
+    int position = dataBuff.position();
+    int newPosition = position + (int)n;
+  
+    // if the new offset is already read into dataBuff, just reposition
+    if (n <= remaining) {
+      assert offsetFromChunkBoundary == 0;
+      dataBuff.position(newPosition);
+      return n;
+    }
+  
+    // for small gap, read through to keep the data/checksum in sync
+    if (n - remaining <= bytesPerChecksum) {
+      dataBuff.position(position + remaining);
+      if (skipBuf == null) {
+        skipBuf = new byte[bytesPerChecksum];
+      }
+      int ret = read(skipBuf, 0, (int)(n - remaining));
+      return ret;
+    }
+  
+    // optimize for big gap: discard the current buffer, skip to
+    // the beginning of the appropriate checksum chunk and then
+    // read to the middle of that chunk to be in sync with checksums.
+    this.offsetFromChunkBoundary = newPosition % bytesPerChecksum;
+    long toskip = n - remaining - this.offsetFromChunkBoundary;
+  
+    dataBuff.clear();
+    checksumBuff.clear();
+  
+    long dataSkipped = dataIn.skip(toskip);
+    if (dataSkipped != toskip) {
+      throw new IOException("skip error in data input stream");
+    }
+    long checkSumOffset = (dataSkipped / bytesPerChecksum) * checksumSize;
+    if (checkSumOffset > 0) {
+      long skipped = checksumIn.skip(checkSumOffset);
+      if (skipped != checkSumOffset) {
+        throw new IOException("skip error in checksum input stream");
+      }
+    }
+
+    // read into the middle of the chunk
     if (skipBuf == null) {
       skipBuf = new byte[bytesPerChecksum];
     }
-    long nSkipped = 0;
-    while ( nSkipped < n ) {
-      int toSkip = (int)Math.min(n-nSkipped, skipBuf.length);
-      int ret = read(skipBuf, 0, toSkip);
-      if ( ret <= 0 ) {
-        return nSkipped;
-      }
-      nSkipped += ret;
+    assert skipBuf.length == bytesPerChecksum;
+    assert this.offsetFromChunkBoundary < bytesPerChecksum;
+    int ret = read(skipBuf, 0, this.offsetFromChunkBoundary);
+    if (ret == -1) {  // EOS
+      return toskip;
+    } else {
+      return (toskip + ret);
     }
-    return nSkipped;
   }
 
   @Override
diff --git a/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/TestShortCircuitLocalRead.java b/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/TestShortCircuitLocalRead.java
index d14b05d076..34ed50a9cd 100644
--- a/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/TestShortCircuitLocalRead.java
+++ b/hadoop-hdfs-project/hadoop-hdfs/src/test/java/org/apache/hadoop/hdfs/TestShortCircuitLocalRead.java
@@ -238,7 +238,53 @@ public ClientDatanodeProtocol run() throws Exception {
       cluster.shutdown();
     }
   }
+
+  @Test
+  public void testSkipWithVerifyChecksum() throws IOException {
+    int size = blockSize;
+    Configuration conf = new Configuration();
+    conf.setBoolean(DFSConfigKeys.DFS_CLIENT_READ_SHORTCIRCUIT_KEY, true);
+    conf.setBoolean(DFSConfigKeys.DFS_CLIENT_READ_SHORTCIRCUIT_SKIP_CHECKSUM_KEY, false);
+    conf.set(DFSConfigKeys.DFS_BLOCK_LOCAL_PATH_ACCESS_USER_KEY,
+        UserGroupInformation.getCurrentUser().getShortUserName());
+    if (simulatedStorage) {
+      conf.setBoolean(SimulatedFSDataset.CONFIG_PROPERTY_SIMULATED, true);
+    }
+    MiniDFSCluster cluster = new MiniDFSCluster.Builder(conf).numDataNodes(1)
+        .format(true).build();
+    FileSystem fs = cluster.getFileSystem();
+    try {
+      // check that / exists
+      Path path = new Path("/");
+      assertTrue("/ should be a directory", fs.getFileStatus(path)
+          .isDirectory() == true);
+      
+      byte[] fileData = AppendTestUtil.randomBytes(seed, size*3);
+      // create a new file in home directory. Do not close it.
+      Path file1 = new Path("filelocal.dat");
+      FSDataOutputStream stm = createFile(fs, file1, 1);
   
+      // write to file
+      stm.write(fileData);
+      stm.close();
+      
+      // now test the skip function
+      FSDataInputStream instm = fs.open(file1);
+      byte[] actual = new byte[fileData.length];
+      // read something from the block first, otherwise BlockReaderLocal.skip()
+      // will not be invoked
+      int nread = instm.read(actual, 0, 3);
+      long skipped = 2*size+3;
+      instm.seek(skipped);
+      nread = instm.read(actual, (int)(skipped + nread), 3);
+      instm.close();
+        
+    } finally {
+      fs.close();
+      cluster.shutdown();
+    }
+  }
+     
   /**
    * Test to run benchmarks between shortcircuit read vs regular read with
    * specified number of threads simultaneously reading.