I wonder if the problem is that when you call sleep;3 with an argument of one second, the processor considers this long enough to switch context and continue with the next threads, which is not the case when the argument is only one ten-thousandth of a second. This is exactly what happens when the amount of time to sleep for is less than the clock resolution.

The sleep functions (meaning both the C library and system functions, on both Linux and FreeBSD) all require an integral argument as their sole calling parameter, so I think what's happening is that the argument is being coerced into a zero, the thread is calling sleep with an argument of zero, and thus the process is sequentially calling sleep two threads at a time, as that is its upper limit on an Intel processor with hyperthreading.

#include <unistd.h>

#include <stddef.h>
#include <stdint.h>
#include <stdlib.h>
#include <stdio.h>

int main(void)
{
    sleep(0.001);

    return EXIT_SUCCESS;
}
[download]

main:
        sub     rsp, 8
        xor     edi, edi
        call    sleep
        xor     eax, eax
        add     rsp, 8
        ret
[download]