HDFS-7291. Persist in-memory replicas with appropriate unbuffered copy API on POSIX and Windows. Contributed by Xiaoyu Yao.

(cherry picked from commit c6f04f391b)

Conflicts:
	hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt
This commit is contained in:
cnauroth 2014-10-28 14:06:27 -07:00
parent f40389ae08
commit c2790932bd
5 changed files with 66 additions and 59 deletions

View File

@ -22,22 +22,20 @@ import java.io.FileDescriptor;
import java.io.FileInputStream;
import java.io.FileOutputStream;
import java.io.IOException;
import java.io.InputStream;
import java.io.OutputStream;
import java.io.RandomAccessFile;
import java.lang.reflect.Field;
import java.nio.ByteBuffer;
import java.nio.MappedByteBuffer;
import java.nio.channels.FileChannel;
import java.util.Map;
import java.util.concurrent.ConcurrentHashMap;
import org.apache.commons.io.FileUtils;
import org.apache.hadoop.classification.InterfaceAudience;
import org.apache.hadoop.classification.InterfaceStability;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.CommonConfigurationKeys;
import org.apache.hadoop.fs.HardLink;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.IOUtils;
import org.apache.hadoop.io.SecureIOUtils.AlreadyExistsException;
import org.apache.hadoop.util.NativeCodeLoader;
import org.apache.hadoop.util.Shell;
@ -662,7 +660,7 @@ public class NativeIO {
* user account name, of the format DOMAIN\UserName. This method
* will remove the domain part of the full logon name.
*
* @param the full principal name containing the domain
* @param Fthe full principal name containing the domain
* @return name with domain removed
*/
private static String stripDomain(String name) {
@ -855,24 +853,66 @@ public class NativeIO {
/**
* Unbuffered file copy from src to dst without tainting OS buffer cache
* In Linux, it uses sendfile() which uses O_DIRECT flag internally
* In Windows, it uses CopyFileEx with COPY_FILE_NO_BUFFERING flag
*
* Note: This does not support FreeBSD/OSX which have a different sendfile()
* semantic. Also, this simple native wrapper does minimal parameter checking
* In POSIX platform:
* It uses FileChannel#transferTo() which internally attempts
* unbuffered IO on OS with native sendfile64() support and falls back to
* buffered IO otherwise.
*
* It minimizes the number of FileChannel#transferTo call by passing the the
* src file size directly instead of a smaller size as the 3rd parameter.
* This saves the number of sendfile64() system call when native sendfile64()
* is supported. In the two fall back cases where sendfile is not supported,
* FileChannle#transferTo already has its own batching of size 8 MB and 8 KB,
* respectively.
*
* In Windows Platform:
* It uses its own native wrapper of CopyFileEx with COPY_FILE_NO_BUFFERING
* flag, which is supported on Windows Server 2008 and above.
*
* Ideally, we should use FileChannel#transferTo() across both POSIX and Windows
* platform. Unfortunately, the wrapper(Java_sun_nio_ch_FileChannelImpl_transferTo0)
* used by FileChannel#transferTo for unbuffered IO is not implemented on Windows.
* Based on OpenJDK 6/7/8 source code, Java_sun_nio_ch_FileChannelImpl_transferTo0
* on Windows simply returns IOS_UNSUPPORTED.
*
* Note: This simple native wrapper does minimal parameter checking before copy and
* consistency check (e.g., size) after copy.
* It is recommended to use wrapper function like
* the Storage#nativeCopyFileUnbuffered() function in hadoop-hdfs.
*
* the Storage#nativeCopyFileUnbuffered() function in hadoop-hdfs with pre/post copy
* checks.
*
* @param src The source path
* @param dst The destination path
* @throws IOException
*/
public static void copyFileUnbuffered(File src, File dst) throws IOException {
if ((nativeLoaded) && (Shell.WINDOWS || Shell.LINUX)) {
if (nativeLoaded && Shell.WINDOWS) {
copyFileUnbuffered0(src.getAbsolutePath(), dst.getAbsolutePath());
} else {
FileUtils.copyFile(src, dst);
FileInputStream fis = null;
FileOutputStream fos = null;
FileChannel input = null;
FileChannel output = null;
try {
fis = new FileInputStream(src);
fos = new FileOutputStream(dst);
input = fis.getChannel();
output = fos.getChannel();
long remaining = input.size();
long position = 0;
long transferred = 0;
while (remaining > 0) {
transferred = input.transferTo(position, remaining, output);
remaining -= transferred;
position += transferred;
}
} finally {
IOUtils.cleanup(LOG, output);
IOUtils.cleanup(LOG, fos);
IOUtils.cleanup(LOG, input);
IOUtils.cleanup(LOG, fis);
}
}
}

View File

@ -1155,46 +1155,8 @@ Java_org_apache_hadoop_io_nativeio_NativeIO_copyFileUnbuffered0(
JNIEnv *env, jclass clazz, jstring jsrc, jstring jdst)
{
#ifdef UNIX
#if (defined(__FreeBSD__) || defined(__MACH__))
THROW(env, "java/io/IOException",
"The function copyFileUnbuffered() is not supported on FreeBSD or Mac OS");
return;
#else
const char *src = NULL, *dst = NULL;
int srcFd = -1;
int dstFd = -1;
struct stat s;
off_t offset = 0;
src = (*env)->GetStringUTFChars(env, jsrc, NULL);
if (!src) goto cleanup; // exception was thrown
dst = (*env)->GetStringUTFChars(env, jdst, NULL);
if (!dst) goto cleanup; // exception was thrown
srcFd = open(src, O_RDONLY);
if (srcFd == -1) {
throw_ioe(env, errno);
goto cleanup;
}
if (fstat(srcFd, &s) == -1){
throw_ioe(env, errno);
goto cleanup;
}
dstFd = open(dst, O_WRONLY | O_CREAT, s.st_mode);
if (dstFd == -1) {
throw_ioe(env, errno);
goto cleanup;
}
if (sendfile(dstFd, srcFd, &offset, s.st_size) == -1) {
throw_ioe(env, errno);
}
cleanup:
if (src) (*env)->ReleaseStringUTFChars(env, jsrc, src);
if (dst) (*env)->ReleaseStringUTFChars(env, jdst, dst);
if (srcFd != -1) close(srcFd);
if (dstFd != -1) close(dstFd);
#endif
THROW(env, "java/lang/UnsupportedOperationException",
"The function copyFileUnbuffered0 should not be used on Unix. Use FileChannel#transferTo instead.");
#endif
#ifdef WINDOWS

View File

@ -632,7 +632,7 @@ public class TestNativeIO {
final String METHOD_NAME = GenericTestUtils.getMethodName();
File srcFile = new File(TEST_DIR, METHOD_NAME + ".src.dat");
File dstFile = new File(TEST_DIR, METHOD_NAME + ".dst.dat");
final int fileSize = 0x8FFFFFF; // 128 MB
final int fileSize = 0x8000000; // 128 MB
final int SEED = 0xBEEF;
final int batchSize = 4096;
final int numBatches = fileSize / batchSize;
@ -650,7 +650,8 @@ public class TestNativeIO {
mapBuf.put(bytesToWrite);
}
NativeIO.copyFileUnbuffered(srcFile, dstFile);
}finally {
Assert.assertEquals(srcFile.length(), dstFile.length());
} finally {
IOUtils.cleanup(LOG, channel);
IOUtils.cleanup(LOG, raSrcFile);
FileUtils.deleteQuietly(TEST_DIR);

View File

@ -802,6 +802,9 @@ Release 2.6.0 - UNRELEASED
HDFS-6934. Move checksum computation off the hot path when writing to RAM
disk. (cnauroth)
HDFS-7291. Persist in-memory replicas with appropriate unbuffered copy API
on POSIX and Windows. (Xiaoyu Yao via cnauroth)
BREAKDOWN OF HDFS-6134 AND HADOOP-10150 SUBTASKS AND RELATED JIRAS
HDFS-6387. HDFS CLI admin tool for creating & deleting an

View File

@ -1003,10 +1003,11 @@ public abstract class Storage extends StorageInfo {
* This method copies the contents of the specified source file
* to the specified destination file using OS specific unbuffered IO.
* The goal is to avoid churning the file system buffer cache when copying
* large files. TheFileUtils#copyLarge function from apache-commons-io library
* can be used to achieve this with an internal memory buffer but is less
* efficient than the native unbuffered APIs such as sendfile() in Linux and
* CopyFileEx() in Windows wrapped in {@link NativeIO#copyFileUnbuffered}.
* large files.
*
* We can't use FileUtils#copyFile from apache-commons-io because it
* is a buffered IO based on FileChannel#transferFrom, which uses MmapByteBuffer
* internally.
*
* The directory holding the destination file is created if it does not exist.
* If the destination file exists, then this method will delete it first.