|
|
|
|
|
|
by devbug
4132 days ago
|
|
|
Unfortunately it is not as simple when you have to account for different threads on different cores accessing the same structure, as mentioned in the article. [1] Or the performance implications of managing multiple instances of structures. [2] Or the alignment requirements for doing SIMD. [3] Additionally, it sometimes makes sense to duplicate data to minimize cache misses. For example, translating/rotating/scaling entities in game engines. It makes sense to copy entity transforms (usually a matrix computed once) from a central manager to sub-managers for certain components. The memory overhead is traded for more efficient processing in sub-systems like rendering, animation, physics, etc. --- [1]: http://en.wikipedia.org/wiki/False_sharing [2]: https://software.intel.com/en-us/articles/a-case-study-compa... [3]: https://software.intel.com/sites/products/documentation/docl... |
|
|