RE: RE: Porting C Linux to ILE C: Performance probleme

It took some time to figure out that one...

It is not memory access, which has uniform speed.
This is index arithmetic, when "p[i]" is converted to address.
When index crosses memory segments, it takes longer for index arithmetic to
verify that segment being addressed is valid.

If you replace          p[i] = 'x';
with              *p++ = 'x';
you will see that it takes slightly longer for first 16 meg, but it will be
uniform all over.

Above only applies to SLS pointers.
If you compile the original program to use TS pointers
   CRTCMOD  STGMDL(*TERASPACE) DTAMDL(*LLP64)
it is also uniform all over, because index arithmetic no longer cares which
segment you are addressing.

    Alexei Pytel
always speaking for myself


                                                                           
             "Mike Mills"                                                  
             <mikem@xxxxxxxx>                                              
             Sent by:                                                   To 
             c400-l-bounces@mi         "'C programming iSeries / AS400'"   
             drange.com                <c400-l@xxxxxxxxxxxx>               
                                                                        cc 
                                                                           
             07/01/2004 03:01                                      Subject 
             PM                        RE: [C400-L] RE: Porting C Linux to 
                                       ILE C: Performance problem          
                                                                           
             Please respond to                                             
               C programming                                               
              iSeries / AS400                                              
                                                                           
                                                                           




Just following up my other message with some code you can use to measure
memory access performance.  It takes one required parameter - the amount of
memory to manipulate, and one optional parameter - a flag indicating
whether
to print out the 'touch rate' for every megabyte (by default this is 'on';
specify '0' as the second parm to turn it off.)

Anyway, the program measures three things:

1. How long it takes to perform the malloc().
2. The 'touch rate' per million bytes (optional; on by default).
3. The total time it takes to touch all allocated bytes.

The program uses 'gettimeofday()' in order to get microsecond resolution.

Summarizing the results I get on my V5R2 machine, for 24,000,000 bytes:

malloc() time:  27,392 microseconds
Typical time to touch 1 million bytes BELOW 16MB:  around 225,000
microseconds
Typical time to touch 1 million bytes ABOVE 16MB:  around 3 seconds +
300,000 microseconds

By the way, this affects Java applications as well (try to use byte[]
arrays
>16MB.)

Following is the code.  Note that you have to specify the 'teraspace'
options on the C compiler in order to allocate >16MB of contiguous memory.

/*** BEGIN PROGRAM ***/

#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>

void computeTimeDiff(struct timeval *start, struct timeval *finish,
                     struct timeval *difference)
{
    difference->tv_sec = finish->tv_sec - start->tv_sec;
    difference->tv_usec = finish->tv_usec - start->tv_usec;

    while(difference->tv_usec < 0)
    {
        difference->tv_sec--;
        difference->tv_usec += 1000000L;
    }
    return;
}



int main(int argc, char *argv[])
{
    int mem = 0, i = 0, showRate = 1;
    char *p = NULL;
    struct timeval beforeMalloc, afterMalloc, beforeLoop, afterLoop;
    struct timeval mallocDiff, loopDiff;
    struct timeval mbStart, mbCurrent, mbDiff;

    if(argc < 2)
    {
        printf("Usage: %s <memory size>\n", argv[0]);
        exit(1);
    }
    if(argc > 2)   /* Check 'showRate' flag */
    {
        if((argv[2][0] == 'n') || (argv[2][0] == 'N') ||
           (argv[2][0] == '0'))
        {
            showRate = 0;
        }
    }

    mem = atoi(argv[1]);
    if(mem < 1)
    {
        printf("%s: must specify memory size > 0\n", argv[0]);
        exit(1);
    }

    gettimeofday(&beforeMalloc, NULL);
    p = malloc(mem);
    gettimeofday(&afterMalloc, NULL);

    computeTimeDiff(&beforeMalloc, &afterMalloc, &mallocDiff);
    printf("Attempt to malloc '%d' bytes took: %ld sec and %ld us\n",
           mem, mallocDiff.tv_sec, mallocDiff.tv_usec);

    if(p == NULL)
    {
        printf("%s: failed to malloc '%d' byts of memory\n", argv[0], mem);
        exit(1);
    }

    gettimeofday(&beforeLoop, NULL);
    gettimeofday(&mbStart, NULL);
    for(i = 0; i < mem; i++)
    {
        if(showRate && (i >= 1000000) && ((i % 1000000) == 0))
        {
            gettimeofday(&mbCurrent, NULL);
            computeTimeDiff(&mbStart, &mbCurrent, &mbDiff);
            printf("Current offset: %d, rate: %ld s and %ld us\n",
                   i, mbDiff.tv_sec, mbDiff.tv_usec);
            gettimeofday(&mbStart, NULL);
        }
        p[i] = 'x';
    }
    gettimeofday(&afterLoop, NULL);

    computeTimeDiff(&beforeLoop, &afterLoop, &loopDiff);
    printf("Touching '%d' bytes took: %ld sec and %ld us\n",
           mem, loopDiff.tv_sec, loopDiff.tv_usec);

    printf("\nFinished!\n\n");

    return(0);
}

/*** END PROGRAM ***/

--
Mike Mills
mikem@xxxxxxxx
ConsulTech Information Systems, Inc.


_______________________________________________
This is the C programming iSeries / AS400 (C400-L) mailing list
To post a message email: C400-L@xxxxxxxxxxxx
To subscribe, unsubscribe, or change list options,
visit: http://lists.midrange.com/mailman/listinfo/c400-l
or email: C400-L-request@xxxxxxxxxxxx
Before posting, please take a moment to review the archives
at http://archive.midrange.com/c400-l.