2 External-memory Model

In a modern computer, the CPU operates on values in its registers, and can fetch values from main memory, usually via several layers of caches. Large files are stored on disk, and data is transferred from disk to main memory. The memory closest to the CPU is fastest but smallest, and the memory farthest from the CPU is largest but has the highest latency. Moreover, the slower layers of memory have greater parallelism: it takes about as long to fetch a consecutive kilobyte from disk as it does to fetch a single byte, since most of the cost is latency for the disk to find the right location. So we might as well always transfer a full block of data at a time; the slower layers of memory have larger block sizes. This prompts the need for algorithms and data structures that exploit locality of reference, arranging the layout of memory with data that is frequently accessed together placed in the same blocks, in order to minimize the number of blocks that need to be accessed.