netmpi.c

fortran并行计算包

		    t0 = When();
		    for (j = 0; j < nrepeat; j++)
		    {
			if (bHeadToHead)
			    SendRecvData(&args);
			else
			{
			    SendData(&args);
			    if (!streamopt)
			    {
				RecvData(&args);
			    }
			}
			if (bNoCache)
			{
			    args.buff += args.bufflen;
			    /* args.buff1 += args.bufflen; */
			}
		    }
		    t = (When() - t0)/((1 + !streamopt) * nrepeat);
		    
		    if (!streamopt)
		    {
			t2 += t*t;
			t1 += t;
			bwdata[n].t = MIN(bwdata[n].t, t);
		    }
		}
		if (!streamopt)
		    SendTime(&args, &bwdata[n].t, &nzero);
		else
		    RecvTime(&args, &bwdata[n].t, &nzero);
		
		if (!streamopt)
		    bwdata[n].variance = t2/TRIALS - t1/TRIALS * t1/TRIALS;
		
	    }
	    else
	    {
		bwdata[n].t = LONGTIME;
		t2 = t1 = 0;
		for (i = 0; i < TRIALS; i++)
		{
		    if (bNoCache)
		    {
			if (bufalign != 0)
			{
			    args.buff = memtmp + ((bufalign - ((MPI_Aint)args.buff % bufalign) + bufoffset) % bufalign);
			    /* args.buff1 = memtmp1 + ((bufalign - ((MPI_Aint)args.buff1 % bufalign) + bufoffset) % bufalign); */
			}
			else
			{
			    args.buff = memtmp;
			    /* args.buff1 = memtmp1; */
			}
		    }
		    
		    Sync(&args);
		    t0 = When();
		    for (j = 0; j < nrepeat; j++)
		    {
			if (bHeadToHead)
			    SendRecvData(&args);
			else
			{
			    RecvData(&args);
			    if (!streamopt)
				SendData(&args);
			}
			if (bNoCache)
			{
			    args.buff += args.bufflen;
			    /* args.buff1 += args.bufflen; */
			}
		    }
		    t = (When() - t0)/((1 + !streamopt) * nrepeat);
		    
		    if (streamopt)
		    {
			t2 += t*t;
			t1 += t;
			bwdata[n].t = MIN(bwdata[n].t, t);
		    }
		}
		if (streamopt)
		    SendTime(&args, &bwdata[n].t, &nzero);
		else
		    RecvTime(&args, &bwdata[n].t, &nzero);
		
		if (streamopt)
		    bwdata[n].variance = t2/TRIALS - t1/TRIALS * t1/TRIALS;
		
	    }
	    tlast = bwdata[n].t;
	    bwdata[n].bits = args.bufflen * CHARSIZE;
	    bwdata[n].bps = bwdata[n].bits / (bwdata[n].t * 1024 * 1024);
	    bwdata[n].repeat = nrepeat;
	    
	    if (args.tr)
	    {
		if (bSavePert)
		{
		/* fprintf(out,"%f\t%f\t%d\t%d\t%f\n", bwdata[n].t, bwdata[n].bps,
		    bwdata[n].bits, bwdata[n].bits / 8, bwdata[n].variance); */
		    if (bUseMegaBytes)
			fprintf(out,"%d\t%f\t%0.9f\n", bwdata[n].bits / 8, bwdata[n].bps / 8, bwdata[n].t);
		    else
			fprintf(out,"%d\t%f\t%0.9f\n", bwdata[n].bits / 8, bwdata[n].bps, bwdata[n].t);
		    fflush(out);
		}
	    }
	    if (!one_buffer)
	    {
		free(memtmp);
		free(memtmp1);
	    }
	    if (args.tr && printopt)
	    {
		if (bUseMegaBytes)
		    fprintf(stdout," %6.2f MBps in %0.9f sec\n", bwdata[n].bps / 8, tlast);
		else
		    fprintf(stdout," %6.2f Mbps in %0.9f sec\n", bwdata[n].bps, tlast);
		fflush(stdout);
	    }
	} /* End of perturbation loop */	
	if (!bSavePert && args.tr)
	{
	    /* if we didn't save all of the perturbation loops, find the max and save it */
	    int index = 1;
	    double dmax = bwdata[n-1].bps;
	    for (; ipert > 1; ipert--)
	    {
		if (bwdata[n-ipert].bps > dmax)
		{
		    index = ipert;
		    dmax = bwdata[n-ipert].bps;
		}
	    }
	    if (bUseMegaBytes)
		fprintf(out,"%d\t%f\t%0.9f\n", bwdata[n-index].bits / 8, bwdata[n-index].bps / 8, bwdata[n-index].t);
	    else
		fprintf(out,"%d\t%f\t%0.9f\n", bwdata[n-index].bits / 8, bwdata[n-index].bps, bwdata[n-index].t);
	    fflush(out);
	}
    } /* End of main loop  */
	
    if (args.tr)
	fclose(out);
    /* THE_END:		 */
    CleanUp(&args);
    free(bwdata);
    return 0;
}


/* Return the current time in seconds, using a double precision number. 	 */
double When()
{
    return MPI_Wtime();
}

int Setup(ArgStruct *p)
{
    int nproc;
    char s[255];
    int len = 255;
    
    MPI_Comm_rank(MPI_COMM_WORLD, &p->prot.iproc);
    MPI_Comm_size(MPI_COMM_WORLD, &nproc);
    
    MPI_Get_processor_name(s, &len);
    /*gethostname(s, len);*/
    printf("%d: %s\n", p->prot.iproc, s);
    fflush(stdout);
    
    if (p->prot.iproc == 0)
	p->prot.nbor = 1;
    else
	p->prot.nbor = 0;
    
    if (nproc < 2)
    {
	printf("Need two processes\n");
	printf("nproc: %i\n", nproc);
	exit(-2);
    }
    
    if (p->prot.iproc == 0)
	p->tr = 1;
    else
	p->tr = 0;
    return 1;	
}	

void Sync(ArgStruct *p)
{
    char ch;
    MPI_Status status;
    if (p->tr)
    {
	MPI_Send(&ch, 0, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD);
	MPI_Recv(&ch, 0, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD, &status);
	MPI_Send(&ch, 0, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD);
    }
    else
    {
	MPI_Recv(&ch, 0, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD, &status);
	MPI_Send(&ch, 0, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD);
	MPI_Recv(&ch, 0, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD, &status);
    }
}

int DetermineLatencyReps(ArgStruct *p)
{
    MPI_Status status;
    double t0, duration = 0;
    int reps = 1, prev_reps = 0;
    int i;

    /* prime the send/receive pipes */
    Sync(p);
    Sync(p);
    Sync(p);

    /* test how long it takes to send n messages 
     * where n = 1, 2, 4, 8, 16, 32, ...
     */
    t0 = When();
    t0 = When();
    t0 = When();
    while ( (duration < 1) ||
	    (duration < 3 && reps < 1000))
    {
	t0 = When();
	for (i=0; i<reps-prev_reps; i++)
	{
	    Sync(p);
	}
	duration += When() - t0;
	prev_reps = reps;
	reps = reps * 2;

	/* use duration from the root only */
	if (p->prot.iproc == 0)
	    MPI_Send(&duration, 1, MPI_DOUBLE, p->prot.nbor, 2, MPI_COMM_WORLD);
	else
	    MPI_Recv(&duration, 1, MPI_DOUBLE, p->prot.nbor, 2, MPI_COMM_WORLD, &status);
    }

    return reps;
}

double TestLatency(ArgStruct *p)
{
    double latency, t0;
    int i;

    g_LATENCYREPS = DetermineLatencyReps(p);
    if (g_LATENCYREPS < 1024 && p->prot.iproc == 0)
    {
	printf("Using %d reps to determine latency\n", g_LATENCYREPS);
	fflush(stdout);
    }

    p->bufflen = 0;
    p->buff = NULL; /*(char *)malloc(p->bufflen);*/
    p->buff1 = NULL; /*(char *)malloc(p->bufflen);*/
    Sync(p);
    t0 = When();
    t0 = When();
    t0 = When();
    t0 = When();
    for (i = 0; i < g_LATENCYREPS; i++)
    {
	if (p->tr)
	{
	    SendData(p);
	    RecvData(p);
	}
	else
	{
	    RecvData(p);
	    SendData(p);
	}
    }
    latency = (When() - t0)/(2 * g_LATENCYREPS);
    /*
    free(p->buff);
    free(p->buff1);
    */

    return latency;
}

double TestSyncTime(ArgStruct *p)
{
    double synctime, t0;
    int i;

    t0 = When();
    t0 = When();
    t0 = When();
    t0 = When();
    t0 = When();
    t0 = When();
    for (i = 0; i < g_LATENCYREPS; i++)
	Sync(p);
    synctime = (When() - t0)/g_LATENCYREPS;

    return synctime;
}

void SendRecvData(ArgStruct *p)
{
    MPI_Status status;
    
    /*MPI_Sendrecv(p->buff, p->bufflen, MPI_BYTE, p->prot.nbor, 1, p->buff1, p->bufflen, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD, &status);*/

    MPI_Request request;
    MPI_Irecv(p->buff1, p->bufflen, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD, &request);
    MPI_Send(p->buff, p->bufflen, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD);
    MPI_Wait(&request, &status);

    /*
    MPI_Send(p->buff, p->bufflen, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD);
    MPI_Recv(p->buff1, p->bufflen, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD, &status);
    */
}

void SendData(ArgStruct *p)
{
    MPI_Send(p->buff, p->bufflen, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD);
}

void RecvData(ArgStruct *p)
{
    MPI_Status status;
    MPI_Recv(p->buff1, p->bufflen, MPI_BYTE, p->prot.nbor, 1, MPI_COMM_WORLD, &status);
}


void SendTime(ArgStruct *p, double *t, int *rpt)
{
    if (*rpt > 0)
	MPI_Send(rpt, 1, MPI_INT, p->prot.nbor, 2, MPI_COMM_WORLD);
    else
	MPI_Send(t, 1, MPI_DOUBLE, p->prot.nbor, 2, MPI_COMM_WORLD);
}

void RecvTime(ArgStruct *p, double *t, int *rpt)
{
    MPI_Status status;
    if (*rpt > 0)
	MPI_Recv(rpt, 1, MPI_INT, p->prot.nbor, 2, MPI_COMM_WORLD, &status);
    else
	MPI_Recv(t, 1, MPI_DOUBLE, p->prot.nbor, 2, MPI_COMM_WORLD, &status);
}

int Establish(ArgStruct *p)
{
    return 1;
}

int  CleanUp(ArgStruct *p)
{
  /*MPI_Barrier(MPI_COMM_WORLD);*/
    MPI_Finalize();
    return 1;
}