Page Size Extension - Motivation

Motivation

Imagine the following scenario. An application program requests a 1 MiB memory block. In order to fulfill this request, an operating system that supports paging and that is running on older x86 CPUs will have to allocate 256 pages of 4 KiB each. An overhead of 1 KiB of memory is required for maintaining page directories and page tables.

When accessing this 1 MiB memory, each of the 256 page entries would be cached in the TLB (a cache that remembers virtual address to physical address translations for faster lookup on subsequent memory requests). Cluttering the TLB is possibly one of the largest disadvantages of having several page entries for what could have been allocated in one single memory block. If the TLB gets filled, then a TLB entry would have to be freed, the page directory and page tables would have to be “walked” in memory, and finally, the memory would be accessed and the new entry would be brought into the TLB. This is a severe performance penalty and it is possibly the largest motivation for augmenting the x86 architecture with larger page sizes.

The PSE allows for page sizes of 4 MiB to exist along with 4 KiB pages. The 1 MiB request described previously would easily be fulfilled with a single 4 MiB page, and it would require only one TLB entry. However, the disadvantage of using larger page sizes is internal fragmentation.