A primer on Virtual memory

Virtual Memory

• One of the core functions of an Operating System
• Hides the complexities of memory from the programmer
• Provides isolation between processes
• Allows enforcing permissions
• Allows execution of programs which require large memory

Paging

• Technique used by modern OSes to implement virtual memory
• Physical main memory is divided into frames
• Address space of the process is divided into pages
• Pages are mapped onto frames
• Needs address translation from page number to frame number

Paging

Translation Lookaside Buffer (TLB)

• We need to do address translation fast (as it is needed for every memory access)
• TLB is a small cache of address translation
• Problem is that it is small
• TLB misses are costly (due to page walks)

Huge pages

What are huge pages?

• x86 archtecture supports 4KB, 2MB and 1GB pages
• But 4KB is the default page size
• Increases TLB reach and Reduces TLB misses
• Gives performance benefit for memory intensive applications

Challenges of supporting huge pages

• Fragmentation
• Permission changes
• Need to identify the regions that benefits from using huge pages

Huge pages in Linux

Linux gives two ways to use huge pages

1. Libhugetlbfs
2. Transparent Huge Pages

libhugetlbfs

• Allows us to reserve some space for huge pages
• Program can explicitly request for huge pages using MAP_HUGETLB flag in mmap syscall
• To view reserved space

  
  cat /proc/meminfo | grep Huge
  				

• To reserve space for hugepages

  
  echo 20 > /sys/kernel/mm/hugepages/hugepages-{size}/nr_hugepages
  				

Small demo of libhugetlbfs

Demo code

Transparent Huge pages (THP)

• Kernel tries to allocate Huge pages whenever possible
• A daemon called khugepaged does defragmentation in the background
• Challenge here is to identify address range that benefits from hugepages
• To get more information about THP

  
  cd /sys/kernel/mm/transparent_hugepage/
  				

Huge pages problems

Inspite of the performance benefits, many blogs recommend turning THP off, on systems running production applications

• This is to avoid tail latencies due to page migration
• Huge pages also makes side and covert channels easy to create

Thank you

Questions?

References

• Libhugetlbfs
• THP