krynecouk / let-it-slow

One hundred thousand falling snowflakes.

Let it Slow

One hundred thousand falling snowflakes.

Application was programmed without focusing on the performance of the rendering. Aim of this project is to measure and analyze performance in such an application. Obtained results will serve as a benchmark comparison with an optimized solution (TBD).

Performance

As performance indicators were measured CPU cache misses and branche mispredictions from three different implementations. Each of them differes in used data type for stored snowflakes:

• vector
• vector modified
• list

For cache and branch profiling was used Cachegrind.

std::vector

Elements of a vector are stored contiguously in a memory and therefore most suitable for CPU caches [Drepper, 2007]. Best cache miss results. But bad branch prediction results because of the unpredicted if in the Game::update function. Definitely place for an improvement.

I   refs:      11,380,565,597
I1  misses:         1,720,676
LLi misses:           531,879
I1  miss rate:           0.02%
LLi miss rate:           0.00%

D   refs:       6,235,673,352  (4,171,929,516 rd   + 2,063,743,836 wr)
D1  misses:        33,433,992  (   27,942,476 rd   +     5,491,516 wr)
LLd misses:        31,596,451  (   26,499,864 rd   +     5,096,587 wr)
D1  miss rate:            0.5% (          0.7%     +           0.3%  )
LLd miss rate:            0.5% (          0.6%     +           0.2%  )

LL refs:           35,154,668  (   29,663,152 rd   +     5,491,516 wr)
LL misses:         32,128,330  (   27,031,743 rd   +     5,096,587 wr)
LL miss rate:             0.2% (          0.2%     +           0.2%  )

Branches:       1,453,407,417  (1,260,642,758 cond +   192,764,659 ind)
Mispredicts:       31,834,170  (   13,412,868 cond +    18,421,302 ind)
Mispred rate:             2.2% (          1.1%     +           9.6%   )

std::vector (modified)

Every fallen snowflake was removed from the vector of snowflakes. Worst result because of the O(n) complexity of v.erase(iterator) and modified vector size.

I   refs:      217,986,584,089
I1  misses:          1,764,113
LLi misses:            257,651
I1  miss rate:            0.00%
LLi miss rate:            0.00%

D   refs:      157,450,187,541  (97,455,277,690 rd   + 59,994,909,851 wr)
D1  misses:      2,239,807,107  ( 2,234,159,053 rd   +      5,648,054 wr)
LLd misses:        236,059,097  (   231,605,745 rd   +      4,453,352 wr)
D1  miss rate:             1.4% (           2.3%     +            0.0%  )
LLd miss rate:             0.1% (           0.2%     +            0.0%  )

LL refs:         2,241,571,220  ( 2,235,923,166 rd   +      5,648,054 wr)
LL misses:         236,316,748  (   231,863,396 rd   +      4,453,352 wr)
LL miss rate:              0.1% (           0.1%     +            0.0%  )

Branches:        3,926,626,287  ( 3,733,811,844 cond +    192,814,443 ind)
Mispredicts:        11,446,855  (    11,164,516 cond +        282,339 ind)
Mispred rate:              0.3% (           0.3%     +            0.1%   )

std::list

Worse results compared to the vector with a static size. Elements of list are scattered across the main memory which is not CPU cache friendly. [Stroustrup, 2012]

I   refs:      11,650,966,650
I1  misses:         1,671,132
LLi misses:           537,711
I1  miss rate:           0.01%
LLi miss rate:           0.00%

D   refs:       6,394,292,664  (4,266,441,223 rd   + 2,127,851,441 wr)
D1  misses:        44,782,629  (   39,361,099 rd   +     5,421,530 wr)
LLd misses:        42,974,182  (   37,944,111 rd   +     5,030,071 wr)
D1  miss rate:            0.7% (          0.9%     +           0.3%  )
LLd miss rate:            0.7% (          0.9%     +           0.2%  )

LL refs:           46,453,761  (   41,032,231 rd   +     5,421,530 wr)
LL misses:         43,511,893  (   38,481,822 rd   +     5,030,071 wr)
LL miss rate:             0.2% (          0.2%     +           0.2%  )

Branches:       1,458,409,526  (1,265,186,899 cond +   193,222,627 ind)
Mispredicts:       13,694,487  (   13,405,127 cond +       289,360 ind)
Mispred rate:             0.9% (          1.1%     +           0.1%   )

Sources

• [Drepper, 2007] : What Every Programmer Should Know About Memory
• [Stroustrup, 2012] : Why you should avoid Linked Lists