I am working, out of passion only, on a simple game engine to try ideas and implement a sample asteroids like game.

The recommended way is component based, streaming data etc. for good cache coherence.

I chose a monolithic object approach with data members ordered in slices as used by the engine sub-systems in the game loop with cache coherence in mind.

In the actual game it doesn’t really matter but I ran some performance tests using 64K cubes in a multi-threaded grid of cells.

To get metrics for cache coherence I ran valgrind --tool=cachegrind --cache-sim=yes and the results are pasted below.

I have no reference for where those numbers are on a scale from “bad” to “decent” to “good”.

What can be expected from an optimized engine?

Kind regards

==1094009== I refs:        16,144,217,054
==1094009== I1  misses:         3,843,166
==1094009== LLi misses:           375,418
==1094009== I1  miss rate:           0.02%
==1094009== LLi miss rate:           0.00%
==1094009== 
==1094009== D refs:         7,066,247,563  (5,233,276,534 rd   + 1,832,971,029 wr)
==1094009== D1  misses:        56,909,087  (   36,616,715 rd   +    20,292,372 wr)
==1094009== LLd misses:        24,687,648  (   18,660,335 rd   +     6,027,313 wr)
==1094009== D1  miss rate:            0.8% (          0.7%     +           1.1%  )
==1094009== LLd miss rate:            0.3% (          0.4%     +           0.3%  )
==1094009== 
==1094009== LL refs:           60,752,253  (   40,459,881 rd   +    20,292,372 wr)
==1094009== LL misses:         25,063,066  (   19,035,753 rd   +     6,027,313 wr)
==1094009== LL miss rate:             0.1% (          0.1%     +           0.3%  )