Ubuntu12.04安装OmpSs环境

此篇博客是接着上两篇的内容

目的是为了实现OmpSs环境以及DLB环境的搭建

前两篇的内容在这里

分别是关于在NVIDIA双显卡环境下ubuntu12.04的CUDA以及OpenCL的搭建方法

地址在这里
http://blog.csdn.net/xianbt/article/details/18871851
http://blog.csdn.net/xianbt/article/details/18894597

为了实现DLB的搭建，我们需要

MPI+OmpSS环境

现在先开始安装OpenMPI

看了看README

发现可以先安装hwloc，所以我先安装了hwloc-1.4.2

./configure --prefix=/usr/local/hwloc-1.4.2

make

make install

现在开始安装OpenMPI

www.open-mpi.org/software/ompi/v1.6/

下载源码包

./configure --prefix=/usr/local/openmpi-1.6.5 --with-hwloc=/usr/local/hwloc-1.4.2

make

make install

然后配置一下环境变量就可以使用了

在~/.bashrc中加上这这个命令

export OPENMPI=/usr/local/openmpi-1.6.5

export LD_LIBRARY_PATH=$OPENMPI/lib:$LD_LIBRARY_PATH

export PATH=$OPENMPI/bin:$PATH

然后执行source ~/.bashrc

注意在root和你自己的用户中的bashrc中都要加上才好

安装好了以后

现在测试一下是不是能用

cd example

make

mpirun -np 4 hello_c

可以看到程序正常执行了

到这里OpenMPI安装成功

接下来就是比较困难的了

要安装OMPSs

这里主要是按照这个网站进行配置的
http://pm.bsc.es/ompss-docs/user-guide/installation.html#installation-of-nanos

首先下载需要的安装包

首先安装extrace

下载安装包
http://www.bsc.es/computer-sciences/performance-tools/downloads

并且编译，在这里会遇到非常多的依赖问题

成功了最好

不成功也不影响下面的继续安装

这里是一些依赖关系

Dependencies: libxml2 2.5.0; libunwind for linux x86/x86-64/ia64/ARM.

Optional: PAPI; DynInst; liberty and libbfd; MPI; OpenMP

先安装

libunwind

./configure --prefix=/usr/local/libunwind-1.1

make

make install

再安装libxml2 2.5.0

apt-get install libxml2

接下来安装

dyninst

这个的安装可以安装这个

apt-get isntall vargrind

其它的都不用安装了

是这样配置的

./configure --prefix=$TARGET --without-papi --without-liberty --without-libbfd --without-openmp --with-mpi=/usr/local/openmpi-1.6.5 --without-libz --with-unwind=/usr/local/libunwind-1.1 --without-dyninst --with-cuda=/usr/local/cuda-5.5 --with-opencl-include=/usr/include/CL
--with-opencl-lib=/usr/lib --with-hwloc=/usr/local/hwloc-1.4.2

make

make install

不出意外是可以成功的

但是配置有问题，说是没找到xml2-config

所以

再安装

apt-get install xml

apt-get install libxml2-dev

再进行一次配置

又出问题了

说是要安装这个

apt-get install binutils-dev

再进行一次配置

最后的配置结果

Package configuration for Extrae 2.4.1:

-----------------------

Installation prefix: /usr/local/ompss

Cross compilation: no

CC: gcc

CXX: g++

Binary type: 64 bits

MPI instrumentation: yes

MPI home: /usr/local/openmpi-1.6.5

Fortran decoration: 0 underscores

peruse available? no

mixed C/Fortran libraries? no

shared libraries? yes

1-sided operations? yes

MPI I/O operations? yes

PACX instrumentation: no

OpenMP instrumentation: yes

pThread instrumentation: yes

CUDA instrumentation: yes, through LD_PRELOAD

CUDA home : /usr/local/cuda-5.5

OpenCL instrumentation: no

Performance counters at instrumentation: no

libbfd available: yes (/usr)

libiberty available: yes (/usr)

zlib available: no

libxml2 available: yes (/usr)

BOOST available: no

callstack access: through libunwind (/usr/local/libunwind-1.1)

Dynamic instrumentation: no

Optional features:

------------------

On-line analysis: no

Clock routine: POSIX / clock_gettime, library in /lib/x86_64-linux-gnu/librt-2.15.so

Heterogeneous support: no

Parallel merge: yes

然后

make

但是make这里出了问题

大概原因是

这里/usr/lib/libmpi.so这些库用不了了，所以我们需要重新把他们链接到/usr/local/openmpi-1.6.5/lib里面

的libmpi.so文件上，对应修改一下就好了

修改了一下，现在又可以继续编译了，以后再遇到有关

/usr/lib/libmpi.so等这些问题的时候就到这里来更改就好了

make成功后

执行

make install

到这里就安装成功了extrace

然后是安装

nanos++

注意安装extrace成功以后在接下来的安装中不要忘记传递这个

flags

--with-extrace=$TARGET

首先确定一下环境要求

主要是确定有cuda

按照下面的步骤进行配置

./configure --prefix=$TARGET --with-extrae=$TARGET --with-cuda=/usr/local/cuda-5.5 --with-opencl-include=/usr/include/CL --with-opencl-lib=/usr/lib --with-hwloc=/usr/local/hwloc-1.4.2

make

make install

到此

nanos++也安装成功了

现在开始安装mercurium

首先看看是否满足环境要求

首先安装

apt-get install bison

apt-get install flex

apt-get install gperf

apt-get install sqlite3

apt-get install libsqlite3-dev

现在开始安装mercurium

./configure --prefix=$TARGET --enable-ompss --with-nanox=$TARGET --enable-nanox-fpga-device --with-nanox=$TARGET --with-cuda=/usr/local/cuda-5.5 --with-mpi=/usr/local/openmpi-1.6.5

错误提示：

configure: error: Fortran compiler not found

安装

apt-get install fortran-compiler

apt-get install git

再配置

配置结果如下

* Tools configured:

Flex : flex

GNU bison : bison

GNU gperf : gperf

git content tracker: (not found)

SQLite 3 cflags:

SQLite 3 libs: -lsqlite3

nVidia CUDA compiler : /usr/local/cuda/bin/nvcc

Intel C compiler: (not found)

Intel C++ compiler: (not found)

Intel Fortran compiler: (not found)

IBM XL C compiler : (not found)

IBM XL C++ compiler : (not found)

IBM XL Fortran compiler : (not found)

* Default type environment : linux-x86_64

Host supports __float128 and quadmath.h : yes

Host supports __int128 : yes

SIMD Support: SSE4.2 (SVML not detected)

Intel Xeon Phi compilation support: no

Intel Xeon Phi remote device used for testing: no

Fortran tests enabled: yes

* TL Examples will be built : no

* Nanos++ OpenMP : yes

Nanos++ includes : /usr/local/ompss/include/nanox

Nanos++ libraries: /usr/local/ompss/lib

Nanos++ tests regions support:

* Intel OpenMP RTL : no

MPI testing enabled: no

GPU testing enabled: yes

OpenCL testing enabled: yes

* OpenMP static profile mode: no

* Optional phases :

Vectorization build and install disabled

Analysis build and install enabled

make

make install

修改一些环境变量

export $PATH=$TARGET/bin:$PATH

source ~/.bashrc

对root和自己用户都执行一次就可以了

最后还有DLB的安装

./configure --with-mpi=/usr/local/openmpi-1.6.5 --prefix=/usr/local/dlb-1.0 --with-hwloc=/usr/local/hwloc-1.4.2

这里hwloc是必须的

DLB 1.0 configuration summary

=============================

Installation prefix : /usr/local/dlb-1.0

Debug version : yes

Instrumentation version : yes

Instr + debug version : yes

Extra Packages:

* MPI CONFIGURED: yes

MPI home: /usr/local/openmpi-1.6.5

MPI includes: /usr/local/openmpi-1.6.5/include

MPI libs: /usr/local/openmpi-1.6.5/lib, -lmpi

* HWLOC CONFIGURED: yes

HWLOC home: /usr/local/hwloc-1.4.2

HWLOC includes: /usr/local/hwloc-1.4.2/include

HWLOC libs: /usr/local/hwloc-1.4.2/lib, -lhwloc

make

make install

到这里就算安装完成了，关于环境测试明天再进行