I have a case where I need to use Java on our research cluster with MPI. One particular function I needed is nicely covered in this question (C++ code included in the linked answer). I built the C++ code and it works exactly as intended.
I tried to build the Java equivalent to this code, but have failed miserably. Even though functionally I have replicated what the C++ code does, the Java version does not consistently return the desired results.
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
(Running with --oversubscribe on my local laptop.)
When I run my Java equivalent, I do not get anywhere near the same result:
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
I would expect that each rank gets one and only one counter. This run, the counter 1 was used twice. Once in a blue moon I can get it to deliver me 1 - 4 (order is unimportant; unique count is).
We run version 2.1.0 on our cluster. On my local laptop I have OpenMPI 2.1.0 and 3.1.0 (current) installed and I can reproduce the proper behavior of the C++ program and the misbehavior of the Java program on either version.
Here is the Counter class I created:
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;
}
}
It should also be noted that the Java code includes something that the C++ code does not:
this.getWin().flushLocalAll();
Without this, counter would be "1" for every rank.
Here also is the first part of the test class:
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;
}
}
I've tried various other techniques, in terms of attempting to fence, using a group in order to use MPI_Win_start() and MPI_Win_complete(), but to no avail. I feel that this is as close to a true representation of the original C++ code I can get.
What am I missing? Why does this not behave the same as the native C++ code?
EDIT: I also find that I need to add this when running this against the actual cluster (it was down for maintenance the last two days):
this.getWin().flush(0);
I think the issue is with these lines
this.getWin().get(valBuff, 1, MPI.INT, this.getHostRank(), i, 1, MPI.INT);
vals.set(i, valBuff.getInt(0));
My understanding is you cannot assume the content of valBuff is correct before MPI_Win_unlock() has been called.
I rewrote the subroutine by using several buffers, and setting vals after MPI_Win_unlock() and was able to get correct output.
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();
}
Note there is no more need for
this.getWin().flushLocalAll();
this.getWin().flush(0);
FWIW, I tried to use a single buffer of this.getSize() integers but was unable to get something work.