<?xml version="1.0" encoding="latin1" ?> <!DOCTYPE chapter SYSTEM "chapter.dtd"> <chapter> <header> <copyright> <year>2007</year> <year>2011</year> <holder>Ericsson AB, All Rights Reserved</holder> </copyright> <legalnotice> The contents of this file are subject to the Erlang Public License, Version 1.1, (the "License"); you may not use this file except in compliance with the License. You should have received a copy of the Erlang Public License along with this software. If not, it can be retrieved online at http://www.erlang.org/. Software distributed under the License is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See the License for the specific language governing rights and limitations under the License. The Initial Developer of the Original Code is Ericsson AB. </legalnotice> <title>The Eight Myths of Erlang Performance</title> <prepared>Bjorn Gustavsson</prepared> <docno></docno> <date>2007-11-10</date> <rev></rev> <file>myths.xml</file> </header> <p>Some truths seem to live on well beyond their best-before date, perhaps because "information" spreads more rapidly from person-to-person faster than a single release note that notes, for instance, that funs have become faster.</p> <p>Here we try to kill the old truths (or semi-truths) that have become myths.</p> <section> <title>Myth: Funs are slow</title> <p>Yes, funs used to be slow. 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 ancient history. Funs was given its own data type in the R6B release and was further optimized in the R7B release. Now the cost for a fun call falls roughly between the cost for a call to 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 bad old days funs were really slow.</p> <p>Nowadays the compiler rewrites list comprehensions into an ordinary recursive function. Of course, 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 will have 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 would still most of the time be 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 would use more words on the stack for each recursion than the number of words a tail-recursive would allocate on the heap. Since more memory is used, the garbage collector will be invoked more frequently, and it will have more work traversing the stack.</p> <p>In R12B and later releases, there is an optimization that will in many cases reduces the number of words used on the stack in body-recursive calls, so that a body-recursive list function and tail-recursive function that calls <seealso marker="stdlib:lists#reverse/1">lists:reverse/1</seealso> at the end will use exactly 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?</p> <p>It depends. On Solaris/Sparc, the body-recursive function seems to be slightly faster, even for lists with very many elements. On the x86 architecture, tail-recursion was up to about 30 percent 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 absolutely sure that the tail-recursive list function will be the fastest in all circumstances.</p> <p>Note: A tail-recursive function that does not need to reverse the list at the end is, of course, faster than a body-recursive function, as are tail-recursive functions that do not construct any terms at all (for instance, a function that sums all integers in a list).</p> </section> <section> <title>Myth: '++' is always bad</title> <p>The <c>++</c> operator has, somewhat undeservedly, got a very bad reputation. It probably has something to do with code like</p> <p><em>DO NOT</em></p> <code type="erl"> naive_reverse([H|T]) -> naive_reverse(T)++[H]; naive_reverse([]) -> [].</code> <p>which is the most inefficient way there is to reverse a list. Since the <c>++</c> operator copies its left operand, the result will be copied again and again and again... leading to quadratic complexity.</p> <p>On the other hand, using <c>++</c> like this</p> <p><em>OK</em></p> <code type="erl"> naive_but_ok_reverse([H|T], Acc) -> naive_but_ok_reverse(T, [H]++Acc); naive_but_ok_reverse([], Acc) -> Acc.</code> <p>is not bad. Each list element will only be copied once. The growing result <c>Acc</c> is the right operand for the <c>++</c> operator, and it will <em>not</em> be copied.</p> <p>Of course, experienced Erlang programmers would actually write</p> <p><em>DO</em></p> <code type="erl"> vanilla_reverse([H|T], Acc) -> vanilla_reverse(T, [H|Acc]); vanilla_reverse([], Acc) -> Acc.</code> <p>which is slightly more efficient because you don't build a list element only to directly copy it. (Or it would be more efficient if the the compiler did not automatically rewrite <c>[H]++Acc</c> to <c>[H|Acc]</c>.)</p> </section> <section> <title>Myth: Strings are slow</title> <p>Actually, string handling could be slow if done improperly. In Erlang, you'll have to think a little more about how the strings are used and choose an appropriate representation and use the <seealso marker="stdlib:re">re</seealso> module instead of the obsolete <c>regexp</c> module if you are going to use regular expressions.</p> </section> <section> <title>Myth: Repairing a Dets file is very slow</title> <p>The repair time is still proportional to the number of records in the file, but Dets repairs used to be much, much slower in the past. Dets has been massively rewritten and improved.</p> </section> <section> <title>Myth: BEAM is a stack-based byte-code virtual machine (and therefore slow)</title> <p>BEAM is a register-based virtual machine. It has 1024 virtual registers that are used for holding temporary values and for passing arguments when calling functions. Variables that need to survive a function call are saved to the stack.</p> <p>BEAM is a threaded-code interpreter. Each instruction is word pointing directly to executable C-code, making instruction dispatching very fast.</p> </section> <section> <title>Myth: Use '_' to speed up your program when a variable is not used</title> <p>That was once true, but since R6B the BEAM compiler is quite capable of seeing itself that a variable is not used.</p> </section> </chapter>