Inside: why lavamd is slow on Ubuntu-deer?

[wukefe]

Problem Description

• Figure: See Browser Performance Chrome - Ubuntu Deer on Page 4 Hyper-link
• Raw data (Execution time)

Execution time (Chrome63 used on Ubuntu-deer)

Name	js (s)	wasm (s)	gcc-native (s)
lavamd	1.6998	1.3137	3.4651

Experiments on Ubuntu

Setup

• Run on both sable-deer and sable-intel which have the exact same OS (i.e. Ubuntu)
• Try different compiler versions for C (i.e. gcc 5.4, clang 3.5 and clang 3.8)

Observations

• No much difference between different platforms, though sable-deer and sable-intel have different CPU frequency (2.0 GHz vs. 3.6 GHz, see machine details)
• No much difference between different compilers (between gcc, clang35, and clang38)

benchmark	impl.	compiler	platform	env.	input	mean	std	rep.
lavamd	c	gcc	sable-intel	native	medium	3.1095s	+-0.12%	10
lavamd	c	clang35	sable-intel	native	medium	3.2519s	+-0.14%	10
lavamd	c	clang38	sable-intel	native	medium	3.1092s	+-0.07%	10
lavamd	c	gcc	sable-deer	native	medium	3.4823s	+-0.94%	10
lavamd	c	clang35	sable-deer	native	medium	3.6138s	+-1.03%	10
lavamd	c	clang38	sable-deer	native	medium	3.4792s	+-0.87%	10

Related to Math Library

In lavamd, the math function exp occupies the most of computation time. We observed that the performance of exp on different platforms returned significant varying results.

#define N 1000000
#define op exp   // Operation
double res[N];

int main(){
    clock_t t0 = clock(); // Timing starts
    for(int i=0;i<N;i++){
        res[i] = op(i * 0.2); // Core
    }
    clock_t t1 = clock(); // Timing ends
    double s = 0;
    for(int i=0;i<N;i++){
        s = res[i];
    }
    printf("time: %lf ms\n", (t1-t0)*1.0/CLOCKS_PER_SEC*1000.0);
    return s > 100;
}

There are two versions of the code:

• Ver-op: with a math operation (#define op exp)
• Ver-no-op: without a math operation (#define op)

Performance results (gcc 5.4, -O3)

Platforms	Ver-op (ms)	Ver-no-op (ms)
mbp-2015	6.9	3.4
Ubuntu-deer	103.3	7.4

We can see that the math operation exp is a very expensive operation on Ubuntu-deer. However, on Mac OS, the execution time doesn't go up much with this operation. :)