To reduce TLB (translation lookaside buffer) misses, modern CPUs offer a feature called "large pages" ("huge pages" on Linux). Large pages allow a page size of 2 MB (on some CPUs even 1 GB) instead of 4 KB. Using a larger page size is especially beneficial when accessing large amounts of memory with a random pattern. X100 supports the use of large pages for certain data structures that would benefit the most.

Linux: Some modern Linux systems (such as RHEL 6) have transparent huge pages functionality. Huge pages are used automatically and do not have to be configured manually. In such cases, do not use the X100 options for huge pages. Consult the documentation for your operating system to see if the operating system supports transparent huge pages and if it needs to be enabled.

You must designate an amount of memory for large pages before starting Ingres. This memory is used for large page allocations only, not for normal allocations. Because this division of memory is static, you should wisely choose the amount of memory available in large pages.

If you encounter problems with large pages, you can switch it off in the vectorwise.conf file by setting [memory] use_huge_tlb (see use_huge_tlb) to FALSE.

To enable the use of huge pages in Ingres the following is needed:

For easier administration of this feature, extra tools are recommended, often found in a package called libhugetlbfs-utils or similar.

The commands and amounts provided here is an example of designating memory for huge pages on Linux. Before issuing these commands, understand what they do, and adapt the examples, as needed. Reserving pages for use in huge page allocations is system wide, so make sure you are not interfering with other users.

To make 2 GB available for 2 MB huge pages, issue the following commands as root on the command line before starting Ingres:

To switch off, enter the following command as root on the command line:

To check if memory is allocated for huge pages and how much of it is in use, type at the command line:

The information about huge pages is shown in the following example lines:

Making memory available for huge pages requires defragmenting the specified amount of memory, so it can take a while. Typically, it is fastest to do this immediately after system startup, when memory is not as fragmented.

For more in-depth information, see the man page of hugeadm, http://linux-mm.org/HugePages, and vm/hugetlbpage.txt in the Linux kernel documentation (http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=blob_plain;f=Documentation/vm/hugetlbpage.txt;hb=HEAD).