How can I calculate L1 data and L2 cache sizes?

[steve lorimer]

Hi

I want to write a small application to estimate the size of my L1 data and L2 caches.

I don't want to use OS specific calls, such as Windows' GetLogicalProcessorInformation or on Linux looking in /proc/cpuinfo or dmesg etc

The logic behind my first attempt is quite simple - write to increasingly larger arrays, calculating how long it takes to do this, and you should notice an increase in access time at the 2 cache boundaries.

So here's my code:

#include <iostream>
#include <time.h>
#include <cstdio>

clock_t whack_cache(const int sz)
{
    char* buf = new char[sz];

    clock_t start = clock();

    for (int64_t i = 0; i < 64 * 1024 * 1024; i++)
        ++buf[(i * 64) % sz]; // writing in increments hopefully means we write to a new cache-line every time

    clock_t elapsed = clock() - start;

    delete [] buf;
    return elapsed;
}

int main()
{
    std::cout << "writing timing results to \"results.csv\"" << std::endl;

    FILE* f = fopen("results.csv", "w");
    if (!f)
        return 1;

    for (int sz = 1024; sz <= 16 * 1024 * 1024; sz <<= 1)
    {
        fprintf(f, "%d, %lu\n", sz / 1024, whack_cache(sz));
        std::cout << ".";
        fflush(stdout);
    }

    fclose(f);

    std::cout << "done" << std::endl;
    return 0;
}

It produces the following output:

1, 1013997
2, 1012313
4, 1008917
8, 1004319
16, 1003130
32, 1002857
64, 1001270
128, 994797
256, 994616
512, 994171
1024, 995780
2048, 1059103
4096, 1975768
8192, 2142419
16384, 2146428

As you can see, there is a noticeable increase in clock-ticks between 2MB and 4MB. As it turns out, I have a 3MB L2 cache, so this logic is working.

However, there is nothing to show where the L1 data cache boundary is. Running cpu-x (I'm on a mac), I am told I have 2 x 32K L1 data caches (1 per core, Core 2 Duo processor).

So my questions are:

1. Why am I not getting results that show an increase at 64K
2. How can I estimate the size of my L1 data cache?
3. How do applications like CPU-X and CPU-Z etc calculate the cache sizes?

Many thanks in advance
Steve

[jsmith]

I think there are a lot of variables to consider. First, in a multitasking OS, other processes
are competing for both caches, and second, in any OS hardware interrupts will cause code
to execute that will also compete for said resources.

AFAIK, and maybe the way those programs work, is, unfortunately, a lookup table. If you
know exactly what processor is in the machine, then you can use a lookup table to
determine the cache sizes.

(This is OTOMH, I didn't research it).

[steve lorimer]

Thanks for the response.

That did occur to me, swapping, paging etc...

What confuses me is that Igor got plausible results using an application that is much the same: http://igoro.com/archive/gallery-of-processor-cache-effects/ (see example 3)