hadoop-mapreduce-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Leo Leung <lle...@ddn.com>
Subject RE: how to turn on NativeIO.posixFadviseIfPossible
Date Fri, 21 Jun 2013 19:35:25 GMT
Hi Jun:

Looks like it is not in the apache hadoop 1.x release binaries,  this probably will have to
be answered by the apache hadoop 1.x release team

To build the native libs:


1)      Get the source code and install the dependencies to compile hadoop

2)      You'll have to setup the env for it,  such as JAVA_HOME. Gcc installed etc.

3)      Got to <hadoop-source dir>     ant [veryclean] compile-core-native  >compile.out

Verify with
[branch-1]$ objdump -Tt ./build/native/Linux-amd64-64/lib/libhadoop.so | grep fadv
00000000000056a0 g     F .text  00000000000000a3              Java_org_apache_hadoop_io_nativeio_NativeIO_posix_1fadvise
0000000000000000       F *UND*  0000000000000000              posix_fadvise@@GLIBC_2.2.5
0000000000000000      DF *UND*  0000000000000000  GLIBC_2.2.5 posix_fadvise
00000000000056a0 g    DF .text  00000000000000a3  Base        Java_org_apache_hadoop_io_nativeio_NativeIO_posix_1fadvise


If you get an error about  "make distclean" error
Just cd in src/native and run  make distclean (to cleanup )

Example of ant output that shows checking for fadvise
     [exec] checking for fcntl.h... yes
     [exec] checking for stdlib.h... (cached) yes
     [exec] checking for string.h... (cached) yes
     [exec] checking for unistd.h... (cached) yes
     [exec] checking for fcntl.h... (cached) yes
     [exec] checking for posix_fadvise... yes
     [exec] checking for fcntl.h... (cached) yes
     [exec] checking for sync_file_range... yes

If you see "no" for posix_fadvise or "sync_file_range"
And you are sure you have posix_fdavise support (RHEL 6.x or CentOS 6.x)

you may need to move the section on src/native/configure.ac
Try to modify the configure.ac
--- src/native/configure.ac     (revision 1495498)
+++ src/native/configure.ac     (working copy)
@@ -47,6 +47,21 @@
AC_PROG_CC
AC_PROG_LIBTOOL

+dnl check for posix_fadvise
+AC_CHECK_HEADERS(fcntl.h, [AC_CHECK_FUNCS(posix_fadvise)])
+
+dnl check for sync_file_range
+AC_CHECK_HEADERS(fcntl.h, [AC_CHECK_FUNCS(sync_file_range)])
+
+# Checks for typedefs, structures, and compiler characteristics.
+AC_C_CONST
+
+# Checks for library functions.
+AC_CHECK_FUNCS([memset])
+
+# Check for nonstandard STRERROR_R
+AC_FUNC_STRERROR_R
+
# Checks for libraries.
dnl Check for '-ldl'
AC_CHECK_LIB([dl], [dlopen])
@@ -104,21 +119,6 @@
dnl Check for headers needed by the native Group resolution implementation
AC_CHECK_HEADERS([fcntl.h stdlib.h string.h unistd.h], [], AC_MSG_ERROR(Some system headers
not found... please ensure their presence on your platform.))

-dnl check for posix_fadvise
-AC_CHECK_HEADERS(fcntl.h, [AC_CHECK_FUNCS(posix_fadvise)])
-
-dnl check for sync_file_range
-AC_CHECK_HEADERS(fcntl.h, [AC_CHECK_FUNCS(sync_file_range)])
-
-# Checks for typedefs, structures, and compiler characteristics.
-AC_C_CONST
-
-# Checks for library functions.
-AC_CHECK_FUNCS([memset])
-
-# Check for nonstandard STRERROR_R
-AC_FUNC_STRERROR_R
-
AC_CONFIG_FILES([Makefile])
AC_OUTPUT


Hope this helps




From: Jun Li [mailto:jltz922181@gmail.com]
Sent: Friday, June 21, 2013 9:44 AM
To: user@hadoop.apache.org
Subject: Re: how to turn on NativeIO.posixFadviseIfPossible

Hi Leo,

Following your instruction,  the following is what I got:

[junli@mercoop-26 Linux-amd64-64]$ objdump -Tt libhadoop.so | grep -I fadvise
0000000000004d70 g     F .text    000000000000006f              Java_org_apache_hadoop_io_nativeio_NativeIO_posix_1fadvise
0000000000004d70 g    DF .text    000000000000006f  Base        Java_org_apache_hadoop_io_nativeio_NativeIO_posix_1fadvise
Apparently the part of "the GLIBC" is not there
The libhadoop.so is the one that I downloaded as part of the  tar.gz of hadoop.1.1.2 from
the Hadoop Apache web site.
Could you let me know how I can use the source code directory in the downloaded Hadoop Apache
package, to re-compile the libhadoop.so and to make sure that the fadvise call is able to
get referenced?
Regards,

Jun

On Fri, Jun 21, 2013 at 9:19 AM, Leo Leung <lleung@ddn.com<mailto:lleung@ddn.com>>
wrote:
This looks like a compilation problem on the native hadoop libraries.

Please locate the libhadoop.so library on your system and run
[shell]  objdump -Tt libhadoop.so | grep -I fadvise

If you don't see something like the following *the GLIBC* part  (that means the system where
the share lib was compiled did not have it)

00000000000056a0 g     F .text  00000000000000a3              Java_org_apache_hadoop_io_nativeio_NativeIO_posix_1fadvise
0000000000000000       F *UND*  0000000000000000              posix_fadvise@@GLIBC_2.2.5
0000000000000000      DF *UND*  0000000000000000  GLIBC_2.2.5 posix_fadvise
00000000000056a0 g    DF .text  00000000000000a3  Base        Java_org_apache_hadoop_io_nativeio_NativeIO_posix_1fadvise

Note: objdump is from binutils- rpm  (you can use yum install to install it if you don't have
it)


-----Original Message-----
From: Jun Li [mailto:jltz922181@gmail.com<mailto:jltz922181@gmail.com>]
Sent: Friday, June 21, 2013 1:35 AM
To: user@hadoop.apache.org<mailto:user@hadoop.apache.org>
Subject: how to turn on NativeIO.posixFadviseIfPossible

Hi,

I downloaded the current stable version from the Apache Hadoop web site, hadoop-1.1.2. My
machine is an AMD-based machine and Redhat Enterprise 6.1. The detailed Linux kernel version
is:
2.6.32-131.0.15.el6.x86_64

 I ran the TestNativeIO.java under the distribution directory of "test/org/apache/hadoop/io/nativeio/TestNativeIO.java"
and tried to understand how NativeIO.posixFadviseIfPossible behaves, in particular, to check
whether "posix_fadvise" is turned on or not. I am interested in this call as it is used in
Read-Ahead-Pool to cache data to the OS's buffer cache.  The following is the test case that
I ran:

@Test
  public void testPosixFadvise() throws Exception {
    FileInputStream fis = new FileInputStream("/dev/zero");
    try {
      NativeIO.posixFadviseIfPossible(fis.getFD(), 0, 0,
                             NativeIO.POSIX_FADV_SEQUENTIAL);
    } catch (NativeIOException noe) {
      // we should just skip the unit test on machines where we don't
      // have fadvise support
      assumeTrue(false);
    } finally {
      fis.close();
    }

However, when I stepped into the code and reached "NativeIO.java"
under the package of "org.apache.hadoop.io.nativeio",  in the particular call below:

public static void posixFadviseIfPossible(
      FileDescriptor fd, long offset, long len, int flags)
      throws NativeIOException {

    if (nativeLoaded && fadvisePossible) {
      try {
        posix_fadvise(fd, offset, len, flags);
      } catch (UnsupportedOperationException uoe) {
        fadvisePossible = false;
      } catch (UnsatisfiedLinkError ule) {
        fadvisePossible = false;
      }
    }
  }

The call to "posix_fadvise"  threw the "UnsupportedOperationException"
exception.

I further traced to the native library, and in the code "NativeIO.c", I found

JNIEXPORT void JNICALL
Java_org_apache_hadoop_io_nativeio_NativeIO_posix_1fadvise(
  JNIEnv *env, jclass clazz,
  jobject fd_object, jlong offset, jlong len, jint flags) { #ifndef HAVE_POSIX_FADVISE
  THROW(env, "java/lang/UnsupportedOperationException",
        "fadvise support not available"); #else

...
}

I believe that the problem of throwing the exception is because "HAVE_POSIX_FADVISE" is not
defined.  I made sure that the native IO library is loaded properly in the Java code, as I
can successfully run the other test cases in "TestNativeIO.java".

So my question is: should I re-compile the "libhadoop" in order to get the version of the
shared library that can have "HAVE_POSIX_FADVISE"
turned on? Or by default, FADVISE is turned on already?

Thank you!


Mime
View raw message