aboutsummaryrefslogtreecommitdiffstats
path: root/system/doc
diff options
context:
space:
mode:
Diffstat (limited to 'system/doc')
-rw-r--r--system/doc/efficiency_guide/myths.xml108
-rw-r--r--system/doc/efficiency_guide/part.xml1
-rw-r--r--system/doc/efficiency_guide/retired_myths.xml63
3 files changed, 106 insertions, 66 deletions
diff --git a/system/doc/efficiency_guide/myths.xml b/system/doc/efficiency_guide/myths.xml
index 5d3ad78b23..778cd06c09 100644
--- a/system/doc/efficiency_guide/myths.xml
+++ b/system/doc/efficiency_guide/myths.xml
@@ -24,7 +24,7 @@
The Initial Developer of the Original Code is Ericsson AB.
</legalnotice>
- <title>The Eight Myths of Erlang Performance</title>
+ <title>The Seven Myths of Erlang Performance</title>
<prepared>Bjorn Gustavsson</prepared>
<docno></docno>
<date>2007-11-10</date>
@@ -35,80 +35,33 @@
<marker id="myths"></marker>
<p>Some truths seem to live on well beyond their best-before date,
perhaps because "information" spreads faster from person-to-person
- than a single release note that says, for example, that funs
- have become faster.</p>
+ than a single release note that says, for example, that body-recursive
+ calls have become faster.</p>
<p>This section tries to kill the old truths (or semi-truths) that have
become myths.</p>
<section>
- <title>Myth: Funs are Slow</title>
- <p>Funs used to be very slow, slower than <c>apply/3</c>.
- Originally, funs were implemented using nothing more than
- compiler trickery, ordinary tuples, <c>apply/3</c>, and a great
- deal of ingenuity.</p>
-
- <p>But that is history. Funs was given its own data type
- in R6B and was further optimized in R7B.
- Now the cost for a fun call falls roughly between the cost for a call
- to a local function and <c>apply/3</c>.</p>
- </section>
-
- <section>
- <title>Myth: List Comprehensions are Slow</title>
-
- <p>List comprehensions used to be implemented using funs, and in the
- old days funs were indeed slow.</p>
-
- <p>Nowadays, the compiler rewrites list comprehensions into an ordinary
- recursive function. Using a tail-recursive function with
- a reverse at the end would be still faster. Or would it?
- That leads us to the next myth.</p>
- </section>
-
- <section>
<title>Myth: Tail-Recursive Functions are Much Faster
Than Recursive Functions</title>
<p><marker id="tail_recursive"></marker>According to the myth,
- recursive functions leave references
- to dead terms on the stack and the garbage collector has to copy
- all those dead terms, while tail-recursive functions immediately
- discard those terms.</p>
-
- <p>That used to be true before R7B. In R7B, the compiler started
- to generate code that overwrites references to terms that will never
- be used with an empty list, so that the garbage collector would not
- keep dead values any longer than necessary.</p>
-
- <p>Even after that optimization, a tail-recursive function is
- still most of the times faster than a body-recursive function. Why?</p>
-
- <p>It has to do with how many words of stack that are used in each
- recursive call. In most cases, a recursive function uses more words
- on the stack for each recursion than the number of words a tail-recursive
- would allocate on the heap. As more memory is used, the garbage
- collector is invoked more frequently, and it has more work traversing
- the stack.</p>
-
- <p>In R12B and later releases, there is an optimization that
- in many cases reduces the number of words used on the stack in
- body-recursive calls. A body-recursive list function and a
- tail-recursive function that calls <seealso
- marker="stdlib:lists#reverse/1">lists:reverse/1</seealso> at
- the end will use the same amount of memory.
- <c>lists:map/2</c>, <c>lists:filter/2</c>, list comprehensions,
- and many other recursive functions now use the same amount of space
- as their tail-recursive equivalents.</p>
-
- <p>So, which is faster?
- It depends. On Solaris/Sparc, the body-recursive function seems to
- be slightly faster, even for lists with a lot of elements. On the x86
- architecture, tail-recursion was up to about 30% faster.</p>
-
- <p>So, the choice is now mostly a matter of taste. If you really do need
- the utmost speed, you must <em>measure</em>. You can no longer be
- sure that the tail-recursive list function always is the fastest.</p>
+ using a tail-recursive function that builds a list in reverse
+ followed by a call to <c>lists:reverse/1</c> is faster than
+ a body-recursive function that builds the list in correct order;
+ the reason being that body-recursive functions use more memory than
+ tail-recursive functions.</p>
+
+ <p>That was true to some extent before R12B. It was even more true
+ before R7B. Today, not so much. A body-recursive function
+ generally uses the same amount of memory as a tail-recursive
+ function. It is generally not possible to predict whether the
+ tail-recursive or the body-recursive version will be
+ faster. Therefore, use the version that makes your code cleaner
+ (hint: it is usually the body-recursive version).</p>
+
+ <p>For a more thorough discussion about tail and body recursion,
+ see <url href="http://ferd.ca/erlang-s-tail-recursion-is-not-a-silver-bullet.html">Erlang's Tail Recursion is Not a Silver Bullet</url>.</p>
<note><p>A tail-recursive function that does not need to reverse the
list at the end is faster than a body-recursive function,
@@ -199,6 +152,29 @@ vanilla_reverse([], Acc) ->
<p>That was once true, but from R6B the BEAM compiler can see
that a variable is not used.</p>
+
+ <p>Similarly, trivial transformations on the source-code level
+ such as converting a <c>case</c> statement to clauses at the
+ top-level of the function seldom makes any difference to the
+ generated code.</p>
+ </section>
+
+ <section>
+ <title>Myth: A NIF Always Speeds Up Your Program</title>
+
+ <p>Rewriting Erlang code to a NIF to make it faster should be
+ seen as a last resort. It is only guaranteed to be dangerous,
+ but not guaranteed to speed up the program.</p>
+
+ <p>Doing too much work in each NIF call will
+ <seealso marker="erts:erl_nif#WARNING">degrade responsiveness
+ of the VM</seealso>. Doing too little work may mean that
+ the gain of the faster processing in the NIF is eaten up by
+ the overhead of calling the NIF and checking the arguments.</p>
+
+ <p>Be sure to read about
+ <seealso marker="erts:erl_nif#lengthy_work">Long-running NIFs</seealso>
+ before writing a NIF.</p>
</section>
</chapter>
diff --git a/system/doc/efficiency_guide/part.xml b/system/doc/efficiency_guide/part.xml
index 6e10a0c031..5673ddd320 100644
--- a/system/doc/efficiency_guide/part.xml
+++ b/system/doc/efficiency_guide/part.xml
@@ -39,5 +39,6 @@
<xi:include href="drivers.xml"/>
<xi:include href="advanced.xml"/>
<xi:include href="profiling.xml"/>
+ <xi:include href="retired_myths.xml"/>
</part>
diff --git a/system/doc/efficiency_guide/retired_myths.xml b/system/doc/efficiency_guide/retired_myths.xml
new file mode 100644
index 0000000000..37f46566cd
--- /dev/null
+++ b/system/doc/efficiency_guide/retired_myths.xml
@@ -0,0 +1,63 @@
+<?xml version="1.0" encoding="utf-8" ?>
+<!DOCTYPE chapter SYSTEM "chapter.dtd">
+
+<chapter>
+ <header>
+ <copyright>
+ <year>2016</year>
+ <year>2016</year>
+ <holder>Ericsson AB, All Rights Reserved</holder>
+ </copyright>
+ <legalnotice>
+ Licensed under the Apache License, Version 2.0 (the "License");
+ you may not use this file except in compliance with the License.
+ You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+ Unless required by applicable law or agreed to in writing, software
+ distributed under the License is distributed on an "AS IS" BASIS,
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ See the License for the specific language governing permissions and
+ limitations under the License.
+
+ The Initial Developer of the Original Code is Ericsson AB.
+ </legalnotice>
+ <marker id="retired_myths"/>
+ <title>Retired Myths</title>
+ <prepared>Bjorn Gustavsson</prepared>
+ <docno></docno>
+ <date>2016-06-07</date>
+ <rev></rev>
+ <file>retired_myths.xml</file>
+ </header>
+
+ <p>We belive that the truth finally has caught with the following,
+ retired myths.</p>
+
+ <section>
+ <title>Myth: Funs are Slow</title>
+ <p>Funs used to be very slow, slower than <c>apply/3</c>.
+ Originally, funs were implemented using nothing more than
+ compiler trickery, ordinary tuples, <c>apply/3</c>, and a great
+ deal of ingenuity.</p>
+
+ <p>But that is history. Funs was given its own data type
+ in R6B and was further optimized in R7B.
+ Now the cost for a fun call falls roughly between the cost for a call
+ to a local function and <c>apply/3</c>.</p>
+ </section>
+
+ <section>
+ <title>Myth: List Comprehensions are Slow</title>
+
+ <p>List comprehensions used to be implemented using funs, and in the
+ old days funs were indeed slow.</p>
+
+ <p>Nowadays, the compiler rewrites list comprehensions into an ordinary
+ recursive function. Using a tail-recursive function with
+ a reverse at the end would be still faster. Or would it?
+ That leads us to the myth that tail-recursive functions are faster
+ than body-recursive functions.</p>
+ </section>
+</chapter>