Question

我有一个案例需要在MPI的研究集群中使用Java。我需要的一个特定功能很好地涵盖在this question中（链接答案中包含的C ++代码）。我构建了C ++代码，它完全按照预期工作。

我尝试构建与此代码等效的Java，但却失败了。尽管在功能上我已经复制了C ++代码的功能，但Java版本并不能始终如一地返回所需的结果。

mpiexec --oversubscribe -n 4 ./test

0 got counter 1
2 got counter 2
1 got counter 3
3 got counter 4
 1  1  1  1

（在我的本地笔记本电脑上使用--oversubscribe运行。）

当我运行我的Java等价物时，我得不到任何接近相同的结果：

mpirun --oversubscribe -n 4 java -cp .:/usr/local/lib/openmpi/mpi.jar CounterTest

0 got counter 1
3 got counter 1
1 got counter 3
2 got counter 2
 1  1  1  1

我希望每个等级都有一个且只有一个计数器。这次运行，计数器1使用了两次。一旦进入一个蓝色的月亮，我可以得到它给我1 - 4（命令是不重要的;唯一的计数是）。

我们在群集上运行2.1.0版。在我的本地笔记本电脑上安装了OpenMPI 2.1.0和3.1.0（当前版本），我可以在任一版本上重现C ++程序的正确行为和Java程序的错误行为。

这是我创建的Counter类：

import java.nio.ByteBuffer;
import java.util.ArrayList;

import mpi.MPI;
import mpi.MPIException;
import mpi.Win;

public class Counter {
    Win win;
    int hostRank;
    int myVal;
    ByteBuffer data;
    int rank;
    int size;

    public Counter(int hostRank) throws MPIException {
        this.setHostRank(hostRank);
        this.setSize(MPI.COMM_WORLD.getSize());
        this.setRank(MPI.COMM_WORLD.getRank());

        if (this.getRank() == hostRank) {
//          this.setData(MPI.newByteBuffer(this.getSize() * Integer.BYTES));
            this.setData(ByteBuffer.allocateDirect(this.getSize() * Integer.BYTES));
            for (int i = 0; i < this.getData().capacity(); i += Integer.BYTES)
                this.getData().putInt(i, 0);
        } else {
//          this.setData(MPI.newByteBuffer(0));
            this.setData(ByteBuffer.allocateDirect(0));
        }   

        this.setWin(new Win(this.getData(), this.getData().capacity(), Integer.BYTES,
                MPI.INFO_NULL, MPI.COMM_WORLD));

        this.setMyVal(0);
    }

    public int increment(int increment) throws MPIException {

        // A list to store all of the values we pull
        ArrayList<Integer> vals = new ArrayList<Integer>();
        for (int i = 0; i < this.getSize(); i++)
            vals.add(i, 0);

        // Need to convert the increment to a buffer
        ByteBuffer incrbuff = ByteBuffer.allocateDirect(Integer.BYTES);
        incrbuff.putInt(increment);

        // Our values are returned to us in a byte buffer
        ByteBuffer valBuff = ByteBuffer.allocateDirect(Integer.BYTES);

//      System.out.printf("Data for RANK %d: ", this.getRank());
        this.getWin().lock(MPI.LOCK_EXCLUSIVE, this.getHostRank(), 0);
        for (int i = 0; i < this.getSize(); i++) {
            // Always ensure that we're at the top of the buffer
            valBuff.position(0);
            if (i == this.getRank()) {
                this.getWin().accumulate(incrbuff, 1, MPI.INT, this.getHostRank(), i, 1, MPI.INT, MPI.SUM);
                // Without this, it comes back all 1s 
                this.getWin().flushLocalAll();
//              System.out.printf(" [%d] ", this.getMyVal() + increment);
            } else {
                this.getWin().get(valBuff, 1, MPI.INT, this.getHostRank(), i, 1, MPI.INT);
                vals.set(i, valBuff.getInt(0));
//              System.out.printf("  %d  ", vals.get(i))
            }
        }
        this.getWin().unlock(this.getHostRank());

        this.setMyVal(this.getMyVal() + increment);
        vals.set(this.getRank(), this.getMyVal());

//      System.out.printf(" <<%d>> \n", vals.stream().mapToInt(Integer::intValue).sum());
//      this.getWin().unlock(this.getHostRank());

        return vals.stream().mapToInt(Integer::intValue).sum();

    }

    public void printCounter() {
        if (this.getRank() == this.getHostRank()) {
            for (int i = 0; i < this.getSize(); i++) {
                System.out.printf(" %d ", this.getData().getInt());
            }
            System.out.println("");
        }
    }

    public void delete() throws MPIException {
        this.getWin().detach(this.getData());
        this.getWin().free();

        this.setData(null);
        this.setHostRank(0);
        this.setMyVal(0);
        this.setRank(0);
        this.setSize(0);
        this.setWin(null);

    }

    private Win getWin() {
        return win;
    }

    private void setWin(Win win) {
        this.win = win;
    }

    private int getHostRank() {
        return hostRank;
    }

    private void setHostRank(int hostrank) {
        this.hostRank = hostrank;
    }

    private int getMyVal() {
        return myVal;
    }

    private void setMyVal(int myval) {
        this.myVal = myval;
    }

    private ByteBuffer getData() {
        return data;
    }

    private void setData(ByteBuffer data) {
        this.data = data;
    }

    private int getRank() {
        return rank;
    }

    private void setRank(int rank) {
        this.rank = rank;
    }

    private int getSize() {
        return size;
    }

    private void setSize(int size) {
        this.size = size;
    }

}

还应该注意，Java代码包含C ++代码不包含的内容：

this.getWin().flushLocalAll();

如果没有这个，反击将是＆＃34; 1＆＃34;每个级别。

这也是测试类的第一部分：

import java.util.Random;

import mpi.*;

public class CounterTest {

    public static void main(String[] args) {

        try {
            MPI.Init(args);
        } catch (MPIException e1) {
            // TODO Auto-generated catch block
            e1.printStackTrace();
        }

        try {
            test1();
//          test2();
        } catch (MPIException e) {
            // TODO Auto-generated catch block
            e.printStackTrace();
        }

        try {
            MPI.Finalize();
        } catch (MPIException e) {
            // TODO Auto-generated catch block
            e.printStackTrace();
        }

    }

    public static void test1 () throws MPIException {
        Counter c = new Counter(0);
        int rank = MPI.COMM_WORLD.getRank();
        int size = MPI.COMM_WORLD.getSize();

        int result = c.increment(1);
        System.out.printf("%d got counter %d\n", rank, result);

        MPI.COMM_WORLD.barrier();
        c.printCounter();
        c.delete();
        c = null;                       

    }
}

我尝试过各种其他技术，尝试围栏，使用组来使用MPI_Win_start（）和MPI_Win_complete（），但无济于事。我觉得这与我能得到的原始C ++代码的真实表示非常接近。

我错过了什么？为什么这与本机C ++代码的行为不一样？

编辑：我还发现在针对实际群集运行时需要添加此内容（过去两天内维护已停止）：

this.getWin().flush(0);

Answer 1

我认为问题在于这些行

this.getWin().get(valBuff, 1, MPI.INT, this.getHostRank(), i, 1, MPI.INT);
vals.set(i, valBuff.getInt(0));

我的理解是，在调用 valBuff之前，您无法假设MPI_Win_unlock()的内容是正确的。

我通过使用几个缓冲区重写了子程序，并在 vals之后设置MPI_Win_unlock() 并且能够获得正确的输出。

public int increment(int increment) throws MPIException { // A list to store all of the values we pull ArrayList<Integer> vals = new ArrayList<Integer>(); for (int i = 0; i < this.getSize(); i++) vals.add(i, 0); // Need to convert the increment to a buffer ByteBuffer incrbuff = ByteBuffer.allocateDirect(Integer.BYTES); incrbuff.putInt(increment); // Our values are returned to us in several byte buffers ByteBuffer valBuff[] = new ByteBuffer[this.getSize()]; this.getWin().lock(MPI.LOCK_EXCLUSIVE, this.getHostRank(), 0); for (int i = 0; i < this.getSize(); i++) { // Always ensure that we're at the top of the buffer if (i == this.getRank()) { this.getWin().accumulate(incrbuff, 1, MPI.INT, this.getHostRank(), i, 1, MPI.INT, MPI.SUM); } else { valBuff[i] = ByteBuffer.allocateDirect(Integer.BYTES); valBuff[i].position(0); this.getWin().get(valBuff[i], 1, MPI.INT, this.getHostRank(), i, 1, MPI.INT); } } this.getWin().unlock(this.getHostRank()); for (int i = 0; i < this.getSize(); i++) { if (i != this.getRank()) { vals.set(i, valBuff[i].getInt(0)); } } this.setMyVal(this.getMyVal() + increment); vals.set(this.getRank(), this.getMyVal()); return vals.stream().mapToInt(Integer::intValue).sum(); }

请注意，不再需要

this.getWin().flushLocalAll(); this.getWin().flush(0);

FWIW，我试图使用this.getSize()整数的单个缓冲区，但无法得到一些工作。

OpenMPI Java Bindings行为锁，累积，获取的差异

1 个答案: