This document discusses transparent hugepages and the huge zero page feature in Linux. It provides 3 key points: 1) Transparent hugepages allow applications to use huge pages without any special configuration, with graceful fallback to traditional 4KB pages. This can reduce TLB and page table walk overhead compared to only using small pages. 2) Huge pages improve performance by reducing TLB misses and faster page table walks compared to using many small 4KB pages. However, they require more memory and clearing huge pages is not cache friendly. 3) The huge zero page feature allows the kernel to satisfy page faults for zero-initialized memory from a single 2MB page of zeroes, improving performance for work

1. Transparent hugepages.
Huge zero page.
Kirill A. Shutemov
29 December 2012
Minsk, Belarus

2. Virtual to physical address translation
TLBVirtual address Physical addressTLB hit
Page tables walk
TLB missTLB write

3. TLB
• TLB is Fast:
– Three virtual to physical address translations every cycle (2 load, 1 store);
– TLB miss, STLB hit penalty is 7 cycles; can usually be hidden by OOO execution;
• Page table walk is Slow:
– Hundreds cycles;
– In the worst case a page table can be in swap;

4. Page table walk, 4k pages
PGD Table PUD Table PMD Table PTE Table
213948
Page
0123063
offsetptepmdpudpgd
mm->pgd
pte
Virtual address

5. Huge pages
• Pros:
– Reduce TLB pressure:
• 1 TLB entry instead of 512 for every 2M;
• 4k and 2M pages have separate TLB entires;
– Faster page tables walk: 3 data access instead of 4;
– Reduce cache pressure;
• Cons:
– Memory footprint;
– clear_page/copy_page is not cache friendly;

6. Page table walk, 2M pages
02130394863
offsetpmdpudpgd
mm->pgd
PGD Table PUD Table PMD Table Page
Virtual address

7. Hugetlbfs
• Cannot be swapped out;
• Need to be reserved on boot time;
• Requires to root access;
• Explicit application support is required;
• No fallback to 4k pages;

8. Transparent Hugepages
• No special enabling is required;
• Graceful fallback to 4k pages;
• Hugepage can be splitted to 4k pages at any time;
• Hugepages can be swapped out as 4k pages;
• MADV_HUGEPAGE and MADV_NOHUGEPAGE for tuning;

9. Zero page
#include <assert.h>
#include <stdlib.h>
#include <unistd.h>
#define MB (1024*1024)
int main(int argc, char **argv)
{
char *p;
int i;
posix_memalign((void **)&p, 2 * MB, 200 * MB);
for (i = 0; i < 200 * MB; i += 4096)
assert(p[i] == 0);
pause();
return 0;
}

10. 0
Huge zero page
• Two designs have been evaluated:
• “physical” huge zero page;
• “virtual” huge zero page;
• Lockless refcounting:
• free huge zero page in shrinker callback;
• ~1% slowdown on 40 parallel read page faulting processes;
• Merged in v3.8;
• LWN: http://lwn.net/Articles/517465/

11. Questions?