Implement efficient prefetch for parallel bitmap heap scan

[knizhnik]

No description provided.

[MMeent]

I'm still not convinced that we should be updating the io concurrency limit, but all right...

[knizhnik]

I'm still not convinced that we should be updating the io concurrency limit, but all right...

May be this program will convince you:

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

long sum(unsigned mod) {
        long sum = 0;
        switch (mod) {
        case 1000:
                for (unsigned i = 0; i < 1000000000; i++) {
                        sum += i % 1000;
                }
                break;
        case 1024:
                for (unsigned i = 0; i < 1000000000; i++) {
                        sum += i % 1024;
                }
                break;
        default:
                for (unsigned i = 0; i < 1000000000; i++) {
                        sum += i % mod;
                }
        }
        return sum;
}

int main(int argc, char const* argv[]) {
        unsigned mod = atoi(argv[1]);
        printf("sum=%ld\n", sum(mod));
        return 0;
}

At EC2 node results are the following:

admin@ip-172-31-30-176:~$ time ./a.out 1024
sum=511499868928

real    0m0.307s
user    0m0.307s
sys     0m0.000s
admin@ip-172-31-30-176:~$ time ./a.out 1000
sum=499500000000

real    0m0.625s
user    0m0.625s
sys     0m0.000s
admin@ip-172-31-30-176:~$ time ./a.out 1023
sum=510999950187

real    0m7.096s
user    0m7.097s
sys     0m0.000s

So division by constant 1024 (AND) is about two times faster than division by 1000 (MUL) and ~20 faster than division by variable (DIV).