[MPICH] Response too slow! running "mpiexec -n 5 cpi"

[章少衡]

hello, Rusty
Thanks for your comments.
i do some tests as shown below:

1.
The followings were the execution time for each command:

    mpdboot -n 8 -f myhosts       -------> 2 seconds
    mpdtrace -l                   -------> instantaneous 
    mpdringtest                   -------> less than a second

The response to mpdringtest was
    "time for 1 loops = 0.00226........ seconds"
2.
Another program was tested to estimate the network transmission speed.  I "think" the speed was
almost 450M Bits/second: 450000000 Bits per second=8(bits per byte)*4(bytes per
double)*2000000(size in double)*50(rounds)*2(transmitting times per broadcast)/13.649424(seconds)

The running results:
[.....]$ mpiexec -n 4 mytestcodes   
w    ----> 'w' was output 1 seconds after the command keyed in                                   
w
w
w
0
1
3
2
wall clock time = 13.649424
wall clock time = 13.814960

[.....]$ mpiexec -n 8 mytestcodes
w    --------> response very slowly
w
w
w
w
w
w
w
0
1
2
3
5
6
4
7
wall clock time = 19.470830
wall clock time = 20.812201


The followings were the codes:

#include <mpi.h>
#include <math.h>            
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int myid;
#define LLL 2000000
double lotus[LLL];
char processor_name[32];
int main(int argc, char *argv[])
{
	int numprocs,merror;
	int  namelen;

	//MPI=====================================================================================	
	fprintf(stderr,"w\n");
	fflush(stderr);
	merror=MPI_Init(&argc,&argv);
	if(merror!=0)
	{
		printf("Error intializing MPI program. Terminating.\n");
		MPI_Abort(MPI_COMM_WORLD,merror);
		MPI_Finalize();
		return -1;
	}

	MPI_Comm_size(MPI_COMM_WORLD,&numprocs);
	//nodemax=numprocs-1;///////////////////////////
	MPI_Comm_rank(MPI_COMM_WORLD,&myid);
	MPI_Get_processor_name(processor_name,&namelen);
	fprintf(stderr,"%d\n",myid);
	fflush(stderr);

	double startwtime;
	if(myid==0)
		 startwtime= MPI_Wtime();

	for(int asd=0;asd<50;asd++)
	MPI_Bcast(lotus,LLL,MPI_DOUBLE,0,MPI_COMM_WORLD);

	double endwtime;
	if(myid==0)
	{
		endwtime = MPI_Wtime();
		printf("wall clock time = %f\n",endwtime-startwtime);
	}
	
	if(myid==0)
		 startwtime= MPI_Wtime();

	for( asd=0;asd<50;asd++)
	MPI_Bcast(lotus,LLL,MPI_DOUBLE,numprocs-1,MPI_COMM_WORLD);
	if(myid==0)
	{
		endwtime = MPI_Wtime();
		printf("wall clock time = %f\n",endwtime-startwtime);
	}
	
	MPI_Finalize();
	return 0;

}

Thanks for your patience!


--- Rusty Lusk <lusk at mcs.anl.gov> wrote:

> That is unusually slow.  I would expect it to take less than a second.
> How long does it take to do mpdtrace?  You can time how long it takes
> for a message to traverse the mpd ring with mpdringtest.  Use the -help
> argument for each of these commands to find out more about them.
> 
> Rusty Lusk
> 
> From: 章少衡 <shc1515 at yahoo.com.tw>
> Subject: [MPICH] Response too slow! running "mpiexec -n 5 cpi"
> Date: Sun, 13 Aug 2006 18:53:41 +0800 (CST)
> 
> > Is it a normal operation?
> > 
> > After input the command
> > mpiexec -n 5 cpi
> > 
> > it took almost 22 seconds to show the response 
> > "
> > Process 0 of 5 on server
> > Process 1 of 5 on c2
> > Process 2 of 5 on c4
> > Process 3 of 5 on c3
> > Process 4 of 5 on c1
> > .....
> > wall clock time = 0.005223
> > "
> > 
> > Is it normal?
> > 
> > here is my system:
> > 
> > 3Com Baseline 10/100/1000 Switch 8-Port #3C16477
> > AMD Athlon 64 4000+
> > NVIDIA GeForce&#8482;6100 GPU + nForce&#8482;430 MCP
> > 






		
_________________________________________________________ 
捇誥轎煤G蚘眊ㄜ笢弊郔Ч轎煤滅馮毀嶼僵閉湮蚘眊 
http://cn.mail.yahoo.com/?id=77072