Linux MPI集群配置

參考文檔：Linux下MPI並行編程環境搭建配置

 

MPI是一種並行計算架構，MPICH是MPI的一種實現，本集群使用虛擬機安裝，操作系統是ubuntu14.04，使用三台機器，用戶名都是ubuntu，機器名分別是ub0, ub1, ub2

• 安裝MPICH
  • 下載：http://www.mpich.org/static/downloads/3.0.4/mpich-3.0.4.tar.gz
  • 源碼安裝

  1. $ tar -xzvf soft/mpich-3.0.4.tar.gz
     $ cd mpich-3.0.4/
     $ ./configure --prefix=/usr/local/mpich
     $ make && sudo make install

     安裝后加入環境變量到/etc/profile文件，並執行 source /etc/profile，追加內容到/etc/profile

  2. PATH=$PATH:/usr/local/mpich/bin
     MANPATH=$MANPATH:/usr/local/mpich/man
     export PATH MANPATH

• 單節點測試
  • 復制源代碼包下的examples目錄到安裝目錄下

  1. cp -r examples/ /usr/local/mpich

     執行

     mpirun -np 10 ./examples/cpi

     輸出結果如下：

  2. Process 0 of 10 is on ub0
     Process 9 of 10 is on ub0
     Process 1 of 10 is on ub0
     Process 4 of 10 is on ub0
     Process 5 of 10 is on ub0
     Process 7 of 10 is on ub0
     Process 2 of 10 is on ub0
     Process 3 of 10 is on ub0
     Process 6 of 10 is on ub0
     Process 8 of 10 is on ub0
     
     pi is approximately 3.1415926544231256, Error is 0.0000000008333325
     wall clock time = 0.020644

• 集群配置
  • 需要先配置ssh免密碼登錄，把ub0(機器名)當作master node，也即主節點，其他是slave node，也即從節點。配置免密碼ssh登錄的步驟
    • 你需要把主節點的公鑰分別發送給從節點，這樣主節點登錄從節點才是可信任的，就不用密碼，以配置ub0的公鑰給ub1為例
    • ub0產生公鑰

    • $ ssh-keygen -t rsa

      一路enter就行了

    • 把ub0的/home/ubuntu/.ssh/id_rsa.pub 發送到ub1的/home/ubuntu/.ssh/下，如果沒有/home/ubuntu/.ssh/就mkdir .ssh
    • 在ub1的/home/ubuntu/.ssh/下執行

    • $ cat id_rsa.pub >> authorized_keys

    • 嘗試ub0 ssh到ub1試試，看看是否成功設置無密碼登錄，如果成功，就繼續下一個節點吧
  • 復制編譯程序到其他機器上面，這樣就不用在其他機器上進行源碼編譯mpich，節省了時間

  1. scp -r mpich ub1:/usr/local/
     scp -r mpich ub2:/usr/local/

  • 在ub0, ub1, ub2的/etc/hosts上追加

  • 192.168.0.2 ub0
    192.168.0.3 ub1
    192.168.0.4 ub2

    注意，三台機器的/etc/hosts都要追加

  • 把主節點的/usr/local/mpich/example/cpi這個計算圓周率的可執行文件復制到/home/ubuntu目錄下，並且發送到ub1和ub2的/home/ubuntu目錄
  • 在主節點的/home/ubuntu目錄下增加servers文件，記錄集群的機器名和對應的進程數

  • ub0:2
    ub1:2
    ub2:2

  • 在ub0的/home/ubuntu目錄下執行

  • $ mpiexec -n 10 -f servers ./cpi

    你就可以看到下面的結果

  • Process 0 of 10 is on ub0
    Process 1 of 10 is on ub1
    Process 4 of 10 is on ub0
    Process 5 of 10 is on ub2
    Process 6 of 10 is on ub1
    Process 7 of 10 is on ub2
    Process 8 of 10 is on ub0
    Process 9 of 10 is on ub1
    Process 2 of 10 is on ub2
    Process 3 of 10 is on ub1
    pi is approximately 3.1415926544231256, Error is 0.0000000008333325
    wall clock time = 0.018768