Determine process time in microseconds

Hi there,

Can I ask what is the simplest method to determine the time it takes to execute specific process (within a program) with microsecond precision?

I've looked at the "time_val" command but couldn't really understand how it works. An example is given below. Can anyone please advise me?

Code:

---

#include <stdio.h>
int main()
{
    int data[2][3] =
    {
      1, 2, 3,
      4, 5, 6,
    };

    size_t obj_size = sizeof(int);
    size_t obj_cnt = sizeof(data)/sizeof(int);
    char *filename = "output.dat";
    FILE *p_file;

    p_file = fopen(filename, "w");

    // Clock start time (t1)
    fwrite(&data, obj_size, obj_cnt, p_file);
    //Clock end time (t2)

// Do a subtraction (t2-t1)

    fclose(p_file);
 
    return 0;
}

---

Thanks for any help in advance.

Ken

I don't know how you could achieve that, but would it not be possible for you to run a process 1000 times in a row then measure it in millisecs?

It might be difficult with such high precision. Still, there may be a few high performance timers out there. I don't remember their names for Windows, however...

High performance timers? Anything for Linux please!

I've looked at some of the examples running "for" loops for specific number of times and then calculate the delay. However, this may not be ideal for my application.

I'll appreciate if anyone knows of a particular function which can be used to determine the time delay. Thanks.

I'm not sure of it's exact name, but it's something like "QueryPerformanceCounter" or something. I don't recall though and I don't know if it supports Linux. This function has precision to about 0.3 microseconds or so (or 300 nanoseconds). The standard clock seems to have a precision of 1/64 second (which is 15,625 microseconds) and thus not an option for your needs. It can be used with a little thinking - duplication works wonders. The Windows multimedia timer, of which is what I use, gets as precise as about 2000 milliseconds, still a bit much, but by copying a function a few thousand times, it can work this way, but that's not really all that feasible (though it takes about a minute for a test case). If you're just after what's faster between two methods, then you don't need high precision, but just a simple loop with the same criteria for that loop and only one thing being changed. It's part of running experiments.

The only realistic way to get high precision for such a short operation (I'm assuming that data is not hundreds of kilobytes) is to use timestamp counters.

You can use the RDTSC instruction using the "rdtscll()" function - I'm not quite sure which header file it lives in - in the kernel its include/asm-i386/msr.h or include/asm-x86_64/msr.h

--
Mats

Can I ask if anyone tried using the command "ftime"? I've tried to incorporate that in my program and it seems like the command could not be recognised.

Code:

---


#include <sys/timeb.h>

struct timeb        startTime;
struct timeb        endTime;
time_t              deltaSecs;
unsigned            deltaMsecs;

ftime(&startTime);
//
// I tried to calculate the delay in processing this function here.
//
ftime(&endTime);

deltaSecs  = endTime.time - startTime.time;

if ( endTime.millitm < startTime.millitm  ) {
        deltaSecs--;
        deltaMsecs = 1000 + endTime.millitm - startTime.millitm;
    }
    else {
        deltaMsecs = endTime.millitm - startTime.millitm;
    }

---

Errors during compilation:

spi1867h.c:117: error: conflicting types for 'endTime'
spi1867h.c:113: error: previous declaration of 'endTime' was here
spi1867h.c:152: warning: passing arg 1 of `ftime' from incompatible pointer type
spi1867h.c:170: error: request for member `time' in something not a structure or union
spi1867h.c:171: error: request for member `millitm' in something not a structure or union
spi1867h.c:173: error: request for member `millitm' in something not a structure or union
spi1867h.c:176: error: request for member `millitm' in something not a structure or union

Can I ask if I'm missing something here?

Thanks for any help.

And where is line 113?

It says that ftime has been replaced by gettimeofday().

However, you are doing something that is most likely going to take less than 1ms, and whilst getting nominal microseconds value from gettimeofday(), it is not certain to be a precise number [it may well give you microseconds in 1000 microsecond "lumps" between consecutive calls].

--
Mats

Try using what i posted there.
http://cboard.cprogramming.com/showt...ing#post717891