vocheretnyi / parallel-sum-MPI

Implementation of sum of two vectors using MPI

parallel-sum-MPI

Implementation of sum of two vectors using MPI.

I've added generated output.csv file with size of vectors, number of processes and run-time.

Here is command to run it, if you use Linux:

for j in 2 4 8 12 16 24; do for i in `seq 3`; do /usr/bin/mpirun -np $j /your/path/to/project/cmake-build-release/MPI >> output.csv; done done

Results of sum

Size	Processes	Run-time
200000000	1	1.3044
200000000	1	1.3092
200000000	1	1.2215
200000000	2	2.3942
200000000	2	2.4517
200000000	2	2.3716
200000000	4	3.2577
200000000	4	2.6539
200000000	4	2.3973
200000000	8	2.3643
200000000	8	2.5601
200000000	8	2.5600
200000000	12	2.4458
200000000	12	2.5164
200000000	12	2.5755
200000000	16	3.4011
200000000	16	3.2230
200000000	16	2.8996
200000000	24	3.1183
200000000	24	3.1636
200000000	24	3.1457

As you see, we haven't got speed acceleration, parallel version working longer than sequential. That's because cost of MPI_Gather/MPI_Scatter operations bigger that cost of sum.

But if we change sum to another function we can get better results of parallel optimizations. So I choose such function:

double f(double a, double b) {
    if (a > b) {
        return sqrt(a) + cbrt(b);
    }
    return cbrt(b) + sqrt(a);
}

And now get better results:

Results of new function

Size	Processes	Run-time
200000000	1	10.9111
200000000	1	10.7411
200000000	1	9.9409
200000000	2	8.0377
200000000	2	9.7273
200000000	2	7.4046
200000000	4	5.3152
200000000	4	5.3474
200000000	4	4.8490
200000000	8	3.7550
200000000	8	3.7938
200000000	8	3.5647
200000000	12	3.6919
200000000	12	4.4344
200000000	12	3.4221
200000000	16	4.2836
200000000	16	3.9359
200000000	16	3.9416
200000000	24	4.0231
200000000	24	3.8237
200000000	24	3.8215

So now we see better results, because new function not so easy, and cost of it bigger that simple sum.

OS: Ubuntu 16.04.

Processor: Intel® Xeon(R) CPU X5650 @ 2.67GHz × 24.