I'm new to parallel programming and I want to do it in java.
I am wondering if it is possible to send and receive more complex ojects via MPI. I'm using MPJ express. However whenever I want to send a object I get a ClassCastException.
MPI.Init(args);
myrank = MPI.COMM_WORLD.Rank();
numprocs = MPI.COMM_WORLD.Size();
Vector<CustomClass> chr = new Vector<CustomClass>();
if (myrank == 0 ) { //am I the master?
for (int i = 1; i < numprocs; i++) {
MPI.COMM_WORLD.Send(chr, 0, chr.size(), MPI.OBJECT, i, 99); //Here's where the
exception occurs
}
}
else {
Vector<BasicRegion> chr_received = new Vector<BasicRegion>();
MPI.COMM_WORLD.Recv(chr_received, 0, 1, MPI.OBJECT, 0, 99 );
}
Exception:
mpi.MPIException: mpi.MPIException: java.lang.ClassCastException: java.util.Vector cannot be cast to [Ljava.lang.Object;
So my questions are:
- is it possible to send/receive more complex objects with MPJ Express?
- if so: what am I doing wrong?
I am new to MPJ express as well, but seems the enclosing object needs to be primitive type - array of something. (like as you do with the C/C++ implementation in OpenMPI).
This kind of code worked for me well:
Node t[] = new Node[4];
...
count[0] = t.length;
MPI.COMM_WORLD.Send(count, 0, 1, MPI.INT, 1, 98);
MPI.COMM_WORLD.Send(t, 0, t.length, MPI.OBJECT, 1, 99);
} else if( myRank == 1 ) {
int count[] = new int[1];
MPI.COMM_WORLD.Recv( count, 0, 1, MPI.INT, 0, 98);
Status mps = MPI.COMM_WORLD.Recv( t, 0, count[0], MPI.OBJECT, 0, 99 );
...
And of course, you have to have that custom class implementing Serializable interface.
You want to serialize it prior to sending.
import mpi.*;
import java.io.*;
import java.nio.ByteBuffer;
public class MPITest
{
public static void main(String[] args)
{
MPI.Init(args);
int me = MPI.COMM_WORLD.Rank();
int tasks = MPI.COMM_WORLD.Size();
MPI.COMM_WORLD.Barrier();
if(me == 0)
{
Cat cat = new Cat("Tom", 15);
cat.Speak();
ByteBuffer byteBuff = ByteBuffer.allocateDirect(2000 + MPI.SEND_OVERHEAD);
MPI.Buffer_attach(byteBuff);
try
{
ByteArrayOutputStream bos = new ByteArrayOutputStream();
ObjectOutput out = null;
out = new ObjectOutputStream(bos);
out.writeObject(cat);
byte[] bytes = bos.toByteArray();
System.out.println("Serialized to " + bytes.length);
MPI.COMM_WORLD.Isend(bytes, 0, bytes.length, MPI.BYTE, 1, 0);
}
catch(IOException ex)
{
}
}
else
{
byte[] bytes = new byte[2000];
Cat recv = null;
MPI.COMM_WORLD.Recv(bytes, 0, 2000, MPI.BYTE, MPI.ANY_SOURCE, 0);
ByteArrayInputStream bis = new ByteArrayInputStream(bytes);
ObjectInput in = null;
try
{
in = new ObjectInputStream(bis);
Object obj = in.readObject();
recv = (Cat)obj;
recv.Speak();
}
catch(IOException ex)
{
}
catch(ClassNotFoundException cnf)
{
}
}
MPI.COMM_WORLD.Barrier();
MPI.Finalize();
}
}
This works however you may want to use Externalization and do it manually to avoid some of the extra garbage that the serialize routine will send.
HTH
Brian
import mpi.*;
/**
* Compile: javac -cp $MPJ_HOME/lib/mpj.jar:. ObjSend.java
* Execute: mpjrun.sh -np 2 -dport 11000 ObjSend
*/
public class ObjSend {
public static void main(String[] args) throws Exception {
int peer ;
MPI.Init(args);
int rank = MPI.COMM_WORLD.Rank() ;
int size = MPI.COMM_WORLD.Size() ;
int tag = 100 ;
if(rank == 0) {
String [] smsg = new String[1] ;
smsg[0] = "Hi from proc 0" ;
peer = 1 ;
MPI.COMM_WORLD.Send(smsg, 0, smsg.length, MPI.OBJECT,
peer, tag);
System.out.println("proc <"+rank+"> sent a msg to "+
"proc <"+peer+">") ;
} else if(rank == 1) {
String[] rmsg = new String[1] ;
peer = 0 ;
MPI.COMM_WORLD.Recv(rmsg, 0, rmsg.length , MPI.OBJECT,
peer, tag);
System.out.println("proc <"+rank+"> received a msg from "+
"proc <"+peer+">") ;
System.out.println("proc <"+rank+"> received the following "+
"message: \""+rmsg[0]+"\"") ;
}
MPI.Finalize();
}
}