Hi, after I finally got documentation compilation working I updated the patch to be syntactically correct. You will find it attached.
Best regards, -- Christian Kruse http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 1b5f831..68b38f7 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -1128,10 +1128,7 @@ include 'filename'
The use of huge TLB pages results in smaller page tables and
less CPU time spent on memory management, increasing performance. For
more details, see
- <ulink url="https://wiki.debian.org/Hugepages";>the Debian wiki</ulink>.
- Remember that you will need at least shared_buffers / huge page size +
- 1 huge TLB pages. So for example for a system with 6GB shared buffers
- and a hugepage size of 2kb of you will need at least 3156 huge pages.
+ <link linkend="linux-huge-tlb-pages">Linux huge TLB pages</link>.
</para>
<para>
diff --git a/doc/src/sgml/runtime.sgml b/doc/src/sgml/runtime.sgml
index bbb808f..0b98314 100644
--- a/doc/src/sgml/runtime.sgml
+++ b/doc/src/sgml/runtime.sgml
@@ -1307,6 +1307,82 @@ echo -1000 > /proc/self/oom_score_adj
</para>
</note>
</sect2>
+
+ <sect2 id="linux-huge-tlb-pages">
+ <title>Linux huge TLB pages</title>
+
+ <para>
+ Nowadays memory address spaces for processes are virtual. This means, when
+ a process reserves memory, it gets a virtual memory address which has to
+ be translated to a physical memory address by the OS or the CPU. This can
+ be done via calculations, but since memory is accessed very often there is
+ a cache for that, called Translation Lookaside Buffer,
+ short <emphasis>TLB</emphasis>.
+ </para>
+
+ <para>
+ When a process reserves memory, this is done in chunks (often
+ of <literal>4kb</literal>) named pages. This means if a process requires
+ 1GB of RAM, it has <literal>262144</literal> (<literal>1GB</literal>
+ / <literal>4KB</literal>) pages and therefore <literal>262144</literal>
+ entries for the translation table. Since the TLB has a limited number of
+ entries it is obvious that they can't be they can't all be cached, which
+ will lead to loss of performance.
+ </para>
+
+ <para>
+ One way to tune this is to increase the page size. Most platforms allow
+ larger pages, e.g. x86 allows pages of <literal>2MB</literal>. This would
+ reduce the number of pages to <literal>512</literal>
+ (<literal>1GB</literal> / <literal>2MB</literal>). This reduces the number
+ of necessary lookups drastrically.
+ </para>
+
+ <para>
+ To enable this feature in <productname>PostgreSQL</productname> you need a
+ kernel with <varname>CONFIG_HUGETLBFS=y</varname> and
+ <varname>CONFIG_HUGETLB_PAGE=y</varname>. You also have to tune the system
+ setting <varname>vm.nr_hugepages</varname>. To calculate the number of
+ necessary huge pages start <productname>PostgreSQL</productname> without
+ huge pages enabled and check the <varname>VmPeak</varname> value from the
+ proc filesystem:
+<programlisting>
+$ <userinput>head -1 /path/to/data/directory/postmaster.pid</userinput>
+4170
+$ <userinput>grep ^VmPeak /proc/4170/status</userinput>
+VmPeak: 6490428 kB
+</programlisting>
+ <literal>6490428</literal> / <literal>2048</literal>
+ (<varname>PAGE_SIZE</varname> <literal>2MB</literal>) are
+ roughly <literal>3169.154</literal> huge pages, so you will need at
+ least <literal>3170</literal> huge pages:
+<programlisting>
+$ <userinput>sysctl -w vm.nr_hugepages=3170</userinput>
+</programlisting>
+ Sometimes the kernel is not able to allocate the desired number of huge
+ pages, so it might be necessary to repeat that command or to reboot. Don't
+ forget to add an entry to <filename>/etc/sysctl.conf</filename> to persist
+ this setting through reboots.
+ </para>
+
+ <para>
+ The default behavior for huge pages
+ in <productname>PostgreSQL</productname> is to use them when possible and
+ to fallback to normal pages when failing. To enforce the use of huge
+ pages, you can
+ set <link linkend="guc-huge-tlb-pages"><varname>huge_tlb_pages</varname></link>
+ to <literal>on</literal>. Note that in this
+ case <productname>PostgreSQL</productname> will fail to start if not
+ enough huge pages are available.
+ </para>
+
+ <para>
+ For a detailed description of the <productname>Linux</productname> huge
+ pages feature have a look
+ at <ulink url="">https://www.kernel.org/doc/Documentation/vm/hugetlbpage.txt</ulink>.
+ </para>
+
+ </sect2>
</sect1>
pgpuCF6bKmrpI.pgp
Description: PGP signature
