迭代速度和元素大小

Question

I have a std::vector filled with following structures:

#define ELEMENTSIZE 8

struct Element {
int value;
char size[ELEMENTSIZE - 4]; //1 char is 1B in size - 4B int
}; 

The size of structure depends on defined elementsize, which makes an array of chars of specified size in the structure.

I am benchmarking an average value of these structures in vector and I would love to know the reason why vector filled with bigger structures in size takes longer to iterate over.

For example vector with 1 000 000 8B structures takes roughly 1,7ms and the same test with 128B structures 12,7ms.

Is that big difference because of cache only? If so, could you explain why? Or is there any other aspect that I can not see?

Answer 1

The structure is 16 times bigger, so it should take 16 times longer to iterate through. Mathematically 12,7/1,7 = 7,47 times more, so it almost matches up mathematically.

Now imagine the structure containing the 128B elements was a structure containing 8B elements, but the same size. Do you see now that it really is 16 times larger?

Answer 2

The OS must bring the larger structures in memory, which may take this path:

• From virtual memory to main-memory (L4)
• From main-memory to L3, and to L2 and L1 till the processor (if needs processing)
• At the L1 or processor level, contents have to be copied around the iterator object being used. It largely depends on cache performance.
• Now at each iteration, it depends what operation you are performing with the iterator. If content is being copied, or displayed on screen, some sorting/compression being performed more time (from CPU to L4 if needed) would be required.

If all of this is happening, why would 128 structure not take more time than a 8 byte structure?