CODEWITHSUNDEEP
ccachingcpu-cache
super saiyan 4 saiya-jin
super saiyan 4 saiya-jin

Reputation: 99

L1 Cache Line Size

I am trying to determine the L1 cache line size through a C code on a platform where L1 I D cache are 32 KB each and L2 cache is 2MB.

#include<stdio.h>
#include<stdlib.h>
#include<sys/time.h>
#include<time.h>
#define SIZE 100
long long wall_clock_time();
int main()
{
    int *arr=calloc(SIZE,sizeof(int));
    register int r,i;
    long long before,after;
    double time_elapsed;
    for(i=0;i<SIZE;i++)
    {
        before=wall_clock_time();
        r=arr[i];
        after=wall_clock_time();
        time_elapsed=((float)(after - before))/1000000000;
        printf("Element Index = %d, Time Taken = %1.4fn",i,time_elapsed);
    }
    free(arr);
    return 0;
}
long long wall_clock_time() {
    #ifdef __linux__
    struct timespec tp;
    clock_gettime(CLOCK_REALTIME, &tp);
    return (long long)(tp.tv_nsec + (long long)tp.tv_sec * 1000000000ll);
    #else
    struct timeval tv;
    gettimeofday(&tv, NULL);
    return (long long)(tv.tv_usec * 1000 + (long long)tv.tv_sec * 1000000000ll);
    #endif
}

Above is a small code snippet that I am using to access elements of an array and trying to determine the jump in access delay at cache line boundaries. However, when I execute the code I get all the timing outputs as 0.000. I have read several threads on stackoverflow regarding this topic but couldn't understand much, hence attempted to write this code. Can anybody explain to me whether there is an error conceptually or syntactically?

Upvotes: 2

Views: 605

Answers (1)

Leeor
Leeor

Reputation: 19706

The 0.00 should have hinted that you're measuring something too small. The overhead of calling the measurement function is several magnitudes higher than what you measure.

Instead, measure the overall time it takes you to pass the array, and divide by SIZE to amortize it. Since SIZE is also rather small, you should probably repeat this action several hundreds of times and amortize over the entire thing.

Note that this still won't give you the latency, but rather the throughput of accesses. You'll need to come up with a way to measure the line size from that (try reading from the 2nd level cache, and use the fact the reads to the same line would hit in the L1. By increasing your step, you'll be able to see when your BW stops degrading and stays constant).

Upvotes: 2

Related Questions