1 <?xml version="1.0" encoding="iso-8859-1"?>
2 <chapter id="profiling">
3 <title>Profiling</title>
4 <indexterm><primary>profiling</primary>
6 <indexterm><primary>cost-centre profiling</primary></indexterm>
8 <para> Glasgow Haskell comes with a time and space profiling
9 system. Its purpose is to help you improve your understanding of
10 your program's execution behaviour, so you can improve it.</para>
12 <para> Any comments, suggestions and/or improvements you have are
13 welcome. Recommended “profiling tricks” would be
14 especially cool! </para>
16 <para>Profiling a program is a three-step process:</para>
20 <para> Re-compile your program for profiling with the
21 <literal>-prof</literal> option, and probably one of the
22 <literal>-auto</literal> or <literal>-auto-all</literal>
23 options. These options are described in more detail in <xref
24 linkend="prof-compiler-options"/> </para>
25 <indexterm><primary><literal>-prof</literal></primary>
27 <indexterm><primary><literal>-auto</literal></primary>
29 <indexterm><primary><literal>-auto-all</literal></primary>
34 <para> Run your program with one of the profiling options, eg.
35 <literal>+RTS -p -RTS</literal>. This generates a file of
36 profiling information.</para>
37 <indexterm><primary><option>-p</option></primary><secondary>RTS
38 option</secondary></indexterm>
42 <para> Examine the generated profiling information, using one of
43 GHC's profiling tools. The tool to use will depend on the kind
44 of profiling information generated.</para>
49 <sect1 id="cost-centres">
50 <title>Cost centres and cost-centre stacks</title>
52 <para>GHC's profiling system assigns <firstterm>costs</firstterm>
53 to <firstterm>cost centres</firstterm>. A cost is simply the time
54 or space required to evaluate an expression. Cost centres are
55 program annotations around expressions; all costs incurred by the
56 annotated expression are assigned to the enclosing cost centre.
57 Furthermore, GHC will remember the stack of enclosing cost centres
58 for any given expression at run-time and generate a call-graph of
59 cost attributions.</para>
61 <para>Let's take a look at an example:</para>
64 main = print (nfib 25)
65 nfib n = if n < 2 then 1 else nfib (n-1) + nfib (n-2)
68 <para>Compile and run this program as follows:</para>
71 $ ghc -prof -auto-all -o Main Main.hs
77 <para>When a GHC-compiled program is run with the
78 <option>-p</option> RTS option, it generates a file called
79 <filename><prog>.prof</filename>. In this case, the file
80 will contain something like this:</para>
83 Fri May 12 14:06 2000 Time and Allocation Profiling Report (Final)
87 total time = 0.14 secs (7 ticks @ 20 ms)
88 total alloc = 8,741,204 bytes (excludes profiling overheads)
90 COST CENTRE MODULE %time %alloc
96 COST CENTRE MODULE entries %time %alloc %time %alloc
98 MAIN MAIN 0 0.0 0.0 100.0 100.0
99 main Main 0 0.0 0.0 0.0 0.0
100 CAF PrelHandle 3 0.0 0.0 0.0 0.0
101 CAF PrelAddr 1 0.0 0.0 0.0 0.0
102 CAF Main 6 0.0 0.0 100.0 100.0
103 main Main 1 0.0 0.0 100.0 100.0
104 nfib Main 242785 100.0 100.0 100.0 100.0
108 <para>The first part of the file gives the program name and
109 options, and the total time and total memory allocation measured
110 during the run of the program (note that the total memory
111 allocation figure isn't the same as the amount of
112 <emphasis>live</emphasis> memory needed by the program at any one
113 time; the latter can be determined using heap profiling, which we
114 will describe shortly).</para>
116 <para>The second part of the file is a break-down by cost centre
117 of the most costly functions in the program. In this case, there
118 was only one significant function in the program, namely
119 <function>nfib</function>, and it was responsible for 100%
120 of both the time and allocation costs of the program.</para>
122 <para>The third and final section of the file gives a profile
123 break-down by cost-centre stack. This is roughly a call-graph
124 profile of the program. In the example above, it is clear that
125 the costly call to <function>nfib</function> came from
126 <function>main</function>.</para>
128 <para>The time and allocation incurred by a given part of the
129 program is displayed in two ways: “individual”, which
130 are the costs incurred by the code covered by this cost centre
131 stack alone, and “inherited”, which includes the costs
132 incurred by all the children of this node.</para>
134 <para>The usefulness of cost-centre stacks is better demonstrated
135 by modifying the example slightly:</para>
138 main = print (f 25 + g 25)
140 g n = nfib (n `div` 2)
141 nfib n = if n < 2 then 1 else nfib (n-1) + nfib (n-2)
144 <para>Compile and run this program as before, and take a look at
145 the new profiling results:</para>
148 COST CENTRE MODULE scc %time %alloc %time %alloc
150 MAIN MAIN 0 0.0 0.0 100.0 100.0
151 main Main 0 0.0 0.0 0.0 0.0
152 CAF PrelHandle 3 0.0 0.0 0.0 0.0
153 CAF PrelAddr 1 0.0 0.0 0.0 0.0
154 CAF Main 9 0.0 0.0 100.0 100.0
155 main Main 1 0.0 0.0 100.0 100.0
156 g Main 1 0.0 0.0 0.0 0.2
157 nfib Main 465 0.0 0.2 0.0 0.2
158 f Main 1 0.0 0.0 100.0 99.8
159 nfib Main 242785 100.0 99.8 100.0 99.8
162 <para>Now although we had two calls to <function>nfib</function>
163 in the program, it is immediately clear that it was the call from
164 <function>f</function> which took all the time.</para>
166 <para>The actual meaning of the various columns in the output is:</para>
172 <para>The number of times this particular point in the call
173 graph was entered.</para>
178 <term>individual %time</term>
180 <para>The percentage of the total run time of the program
181 spent at this point in the call graph.</para>
186 <term>individual %alloc</term>
188 <para>The percentage of the total memory allocations
189 (excluding profiling overheads) of the program made by this
195 <term>inherited %time</term>
197 <para>The percentage of the total run time of the program
198 spent below this point in the call graph.</para>
203 <term>inherited %alloc</term>
205 <para>The percentage of the total memory allocations
206 (excluding profiling overheads) of the program made by this
207 call and all of its sub-calls.</para>
212 <para>In addition you can use the <option>-P</option> RTS option
213 <indexterm><primary><option>-P</option></primary></indexterm> to
214 get the following additional information:</para>
218 <term><literal>ticks</literal></term>
220 <para>The raw number of time “ticks” which were
221 attributed to this cost-centre; from this, we get the
222 <literal>%time</literal> figure mentioned
228 <term><literal>bytes</literal></term>
230 <para>Number of bytes allocated in the heap while in this
231 cost-centre; again, this is the raw number from which we get
232 the <literal>%alloc</literal> figure mentioned
238 <para>What about recursive functions, and mutually recursive
239 groups of functions? Where are the costs attributed? Well,
240 although GHC does keep information about which groups of functions
241 called each other recursively, this information isn't displayed in
242 the basic time and allocation profile, instead the call-graph is
243 flattened into a tree.</para>
245 <sect2><title>Inserting cost centres by hand</title>
247 <para>Cost centres are just program annotations. When you say
248 <option>-auto-all</option> to the compiler, it automatically
249 inserts a cost centre annotation around every top-level function
250 in your program, but you are entirely free to add the cost
251 centre annotations yourself.</para>
253 <para>The syntax of a cost centre annotation is</para>
256 {-# SCC "name" #-} <expression>
259 <para>where <literal>"name"</literal> is an arbitrary string,
260 that will become the name of your cost centre as it appears
261 in the profiling output, and
262 <literal><expression></literal> is any Haskell
263 expression. An <literal>SCC</literal> annotation extends as
264 far to the right as possible when parsing.</para>
268 <sect2 id="prof-rules">
269 <title>Rules for attributing costs</title>
271 <para>The cost of evaluating any expression in your program is
272 attributed to a cost-centre stack using the following rules:</para>
276 <para>If the expression is part of the
277 <firstterm>one-off</firstterm> costs of evaluating the
278 enclosing top-level definition, then costs are attributed to
279 the stack of lexically enclosing <literal>SCC</literal>
280 annotations on top of the special <literal>CAF</literal>
285 <para>Otherwise, costs are attributed to the stack of
286 lexically-enclosing <literal>SCC</literal> annotations,
287 appended to the cost-centre stack in effect at the
288 <firstterm>call site</firstterm> of the current top-level
289 definition<footnote> <para>The call-site is just the place
290 in the source code which mentions the particular function or
291 variable.</para></footnote>. Notice that this is a recursive
296 <para>Time spent in foreign code (see <xref linkend="ffi"/>)
297 is always attributed to the cost centre in force at the
298 Haskell call-site of the foreign function.</para>
302 <para>What do we mean by one-off costs? Well, Haskell is a lazy
303 language, and certain expressions are only ever evaluated once.
304 For example, if we write:</para>
310 <para>then <varname>x</varname> will only be evaluated once (if
311 at all), and subsequent demands for <varname>x</varname> will
312 immediately get to see the cached result. The definition
313 <varname>x</varname> is called a CAF (Constant Applicative
314 Form), because it has no arguments.</para>
316 <para>For the purposes of profiling, we say that the expression
317 <literal>nfib 25</literal> belongs to the one-off costs of
318 evaluating <varname>x</varname>.</para>
320 <para>Since one-off costs aren't strictly speaking part of the
321 call-graph of the program, they are attributed to a special
322 top-level cost centre, <literal>CAF</literal>. There may be one
323 <literal>CAF</literal> cost centre for each module (the
324 default), or one for each top-level definition with any one-off
325 costs (this behaviour can be selected by giving GHC the
326 <option>-caf-all</option> flag).</para>
328 <indexterm><primary><literal>-caf-all</literal></primary>
331 <para>If you think you have a weird profile, or the call-graph
332 doesn't look like you expect it to, feel free to send it (and
333 your program) to us at
334 <email>glasgow-haskell-bugs@haskell.org</email>.</para>
338 <sect1 id="prof-compiler-options">
339 <title>Compiler options for profiling</title>
341 <indexterm><primary>profiling</primary><secondary>options</secondary></indexterm>
342 <indexterm><primary>options</primary><secondary>for profiling</secondary></indexterm>
347 <option>-prof</option>:
348 <indexterm><primary><option>-prof</option></primary></indexterm>
351 <para> To make use of the profiling system
352 <emphasis>all</emphasis> modules must be compiled and linked
353 with the <option>-prof</option> option. Any
354 <literal>SCC</literal> annotations you've put in your source
355 will spring to life.</para>
357 <para> Without a <option>-prof</option> option, your
358 <literal>SCC</literal>s are ignored; so you can compile
359 <literal>SCC</literal>-laden code without changing
365 <para>There are a few other profiling-related compilation options.
366 Use them <emphasis>in addition to</emphasis>
367 <option>-prof</option>. These do not have to be used consistently
368 for all modules in a program.</para>
373 <option>-auto</option>:
374 <indexterm><primary><option>-auto</option></primary></indexterm>
375 <indexterm><primary>cost centres</primary><secondary>automatically inserting</secondary></indexterm>
378 <para> GHC will automatically add
379 <function>_scc_</function> constructs for all
380 top-level, exported functions.</para>
386 <option>-auto-all</option>:
387 <indexterm><primary><option>-auto-all</option></primary></indexterm>
390 <para> <emphasis>All</emphasis> top-level functions,
391 exported or not, will be automatically
392 <function>_scc_</function>'d.</para>
398 <option>-caf-all</option>:
399 <indexterm><primary><option>-caf-all</option></primary></indexterm>
402 <para> The costs of all CAFs in a module are usually
403 attributed to one “big” CAF cost-centre. With
404 this option, all CAFs get their own cost-centre. An
405 “if all else fails” option…</para>
411 <option>-ignore-scc</option>:
412 <indexterm><primary><option>-ignore-scc</option></primary></indexterm>
415 <para>Ignore any <function>_scc_</function>
416 constructs, so a module which already has
417 <function>_scc_</function>s can be compiled
418 for profiling with the annotations ignored.</para>
426 <sect1 id="prof-time-options">
427 <title>Time and allocation profiling</title>
429 <para>To generate a time and allocation profile, give one of the
430 following RTS options to the compiled program when you run it (RTS
431 options should be enclosed between <literal>+RTS...-RTS</literal>
437 <option>-p</option> or <option>-P</option>:
438 <indexterm><primary><option>-p</option></primary></indexterm>
439 <indexterm><primary><option>-P</option></primary></indexterm>
440 <indexterm><primary>time profile</primary></indexterm>
443 <para>The <option>-p</option> option produces a standard
444 <emphasis>time profile</emphasis> report. It is written
446 <filename><replaceable>program</replaceable>.prof</filename>.</para>
448 <para>The <option>-P</option> option produces a more
449 detailed report containing the actual time and allocation
450 data as well. (Not used much.)</para>
457 <indexterm><primary><option>-xc</option></primary><secondary>RTS option</secondary></indexterm>
460 <para>This option makes use of the extra information
461 maintained by the cost-centre-stack profiler to provide
462 useful information about the location of runtime errors.
463 See <xref linkend="rts-options-debugging"/>.</para>
471 <sect1 id="prof-heap">
472 <title>Profiling memory usage</title>
474 <para>In addition to profiling the time and allocation behaviour
475 of your program, you can also generate a graph of its memory usage
476 over time. This is useful for detecting the causes of
477 <firstterm>space leaks</firstterm>, when your program holds on to
478 more memory at run-time that it needs to. Space leaks lead to
479 longer run-times due to heavy garbage collector activity, and may
480 even cause the program to run out of memory altogether.</para>
482 <para>To generate a heap profile from your program:</para>
486 <para>Compile the program for profiling (<xref
487 linkend="prof-compiler-options"/>).</para>
490 <para>Run it with one of the heap profiling options described
491 below (eg. <option>-hc</option> for a basic producer profile).
492 This generates the file
493 <filename><replaceable>prog</replaceable>.hp</filename>.</para>
496 <para>Run <command>hp2ps</command> to produce a Postscript
498 <filename><replaceable>prog</replaceable>.ps</filename>. The
499 <command>hp2ps</command> utility is described in detail in
500 <xref linkend="hp2ps"/>.</para>
503 <para>Display the heap profile using a postscript viewer such
504 as <application>Ghostview</application>, or print it out on a
505 Postscript-capable printer.</para>
509 <sect2 id="rts-options-heap-prof">
510 <title>RTS options for heap profiling</title>
512 <para>There are several different kinds of heap profile that can
513 be generated. All the different profile types yield a graph of
514 live heap against time, but they differ in how the live heap is
515 broken down into bands. The following RTS options select which
516 break-down to use:</para>
522 <indexterm><primary><option>-hc</option></primary><secondary>RTS option</secondary></indexterm>
525 <para>Breaks down the graph by the cost-centre stack which
526 produced the data.</para>
533 <indexterm><primary><option>-hm</option></primary><secondary>RTS option</secondary></indexterm>
536 <para>Break down the live heap by the module containing
537 the code which produced the data.</para>
544 <indexterm><primary><option>-hd</option></primary><secondary>RTS option</secondary></indexterm>
547 <para>Breaks down the graph by <firstterm>closure
548 description</firstterm>. For actual data, the description
549 is just the constructor name, for other closures it is a
550 compiler-generated string identifying the closure.</para>
557 <indexterm><primary><option>-hy</option></primary><secondary>RTS option</secondary></indexterm>
560 <para>Breaks down the graph by
561 <firstterm>type</firstterm>. For closures which have
562 function type or unknown/polymorphic type, the string will
563 represent an approximation to the actual type.</para>
570 <indexterm><primary><option>-hr</option></primary><secondary>RTS option</secondary></indexterm>
573 <para>Break down the graph by <firstterm>retainer
574 set</firstterm>. Retainer profiling is described in more
575 detail below (<xref linkend="retainer-prof"/>).</para>
582 <indexterm><primary><option>-hb</option></primary><secondary>RTS option</secondary></indexterm>
585 <para>Break down the graph by
586 <firstterm>biography</firstterm>. Biographical profiling
587 is described in more detail below (<xref
588 linkend="biography-prof"/>).</para>
593 <para>In addition, the profile can be restricted to heap data
594 which satisfies certain criteria - for example, you might want
595 to display a profile by type but only for data produced by a
596 certain module, or a profile by retainer for a certain type of
597 data. Restrictions are specified as follows:</para>
602 <option>-hc</option><replaceable>name</replaceable>,...
603 <indexterm><primary><option>-hc</option></primary><secondary>RTS option</secondary></indexterm>
606 <para>Restrict the profile to closures produced by
607 cost-centre stacks with one of the specified cost centres
614 <option>-hC</option><replaceable>name</replaceable>,...
615 <indexterm><primary><option>-hC</option></primary><secondary>RTS option</secondary></indexterm>
618 <para>Restrict the profile to closures produced by
619 cost-centre stacks with one of the specified cost centres
620 anywhere in the stack.</para>
626 <option>-hm</option><replaceable>module</replaceable>,...
627 <indexterm><primary><option>-hm</option></primary><secondary>RTS option</secondary></indexterm>
630 <para>Restrict the profile to closures produced by the
631 specified modules.</para>
637 <option>-hd</option><replaceable>desc</replaceable>,...
638 <indexterm><primary><option>-hd</option></primary><secondary>RTS option</secondary></indexterm>
641 <para>Restrict the profile to closures with the specified
642 description strings.</para>
648 <option>-hy</option><replaceable>type</replaceable>,...
649 <indexterm><primary><option>-hy</option></primary><secondary>RTS option</secondary></indexterm>
652 <para>Restrict the profile to closures with the specified
659 <option>-hr</option><replaceable>cc</replaceable>,...
660 <indexterm><primary><option>-hr</option></primary><secondary>RTS option</secondary></indexterm>
663 <para>Restrict the profile to closures with retainer sets
664 containing cost-centre stacks with one of the specified
665 cost centres at the top.</para>
671 <option>-hb</option><replaceable>bio</replaceable>,...
672 <indexterm><primary><option>-hb</option></primary><secondary>RTS option</secondary></indexterm>
675 <para>Restrict the profile to closures with one of the
676 specified biographies, where
677 <replaceable>bio</replaceable> is one of
678 <literal>lag</literal>, <literal>drag</literal>,
679 <literal>void</literal>, or <literal>use</literal>.</para>
684 <para>For example, the following options will generate a
685 retainer profile restricted to <literal>Branch</literal> and
686 <literal>Leaf</literal> constructors:</para>
689 <replaceable>prog</replaceable> +RTS -hr -hdBranch,Leaf
692 <para>There can only be one "break-down" option
693 (eg. <option>-hr</option> in the example above), but there is no
694 limit on the number of further restrictions that may be applied.
695 All the options may be combined, with one exception: GHC doesn't
696 currently support mixing the <option>-hr</option> and
697 <option>-hb</option> options.</para>
699 <para>There are three more options which relate to heap
705 <option>-i<replaceable>secs</replaceable></option>:
706 <indexterm><primary><option>-i</option></primary></indexterm>
709 <para>Set the profiling (sampling) interval to
710 <replaceable>secs</replaceable> seconds (the default is
711 0.1 second). Fractions are allowed: for example
712 <option>-i0.2</option> will get 5 samples per second.
713 This only affects heap profiling; time profiles are always
714 sampled on a 1/50 second frequency.</para>
721 <indexterm><primary><option>-xt</option></primary><secondary>RTS option</secondary></indexterm>
724 <para>Include the memory occupied by threads in a heap
725 profile. Each thread takes up a small area for its thread
726 state in addition to the space allocated for its stack
727 (stacks normally start small and then grow as
730 <para>This includes the main thread, so using
731 <option>-xt</option> is a good way to see how much stack
732 space the program is using.</para>
734 <para>Memory occupied by threads and their stacks is
735 labelled as “TSO” when displaying the profile
736 by closure description or type description.</para>
742 <option>-L<replaceable>num</replaceable></option>
743 <indexterm><primary><option>-L</option></primary><secondary>RTS option</secondary></indexterm>
747 Sets the maximum length of a cost-centre stack name in a
748 heap profile. Defaults to 25.
756 <sect2 id="retainer-prof">
757 <title>Retainer Profiling</title>
759 <para>Retainer profiling is designed to help answer questions
760 like <quote>why is this data being retained?</quote>. We start
761 by defining what we mean by a retainer:</para>
764 <para>A retainer is either the system stack, or an unevaluated
765 closure (thunk).</para>
768 <para>In particular, constructors are <emphasis>not</emphasis>
771 <para>An object B retains object A if (i) B is a retainer object and
772 (ii) object A can be reached by recursively following pointers
773 starting from object B, but not meeting any other retainer
774 objects on the way. Each live object is retained by one or more
775 retainer objects, collectively called its retainer set, or its
776 <firstterm>retainer set</firstterm>, or its
777 <firstterm>retainers</firstterm>.</para>
779 <para>When retainer profiling is requested by giving the program
780 the <option>-hr</option> option, a graph is generated which is
781 broken down by retainer set. A retainer set is displayed as a
782 set of cost-centre stacks; because this is usually too large to
783 fit on the profile graph, each retainer set is numbered and
784 shown abbreviated on the graph along with its number, and the
785 full list of retainer sets is dumped into the file
786 <filename><replaceable>prog</replaceable>.prof</filename>.</para>
788 <para>Retainer profiling requires multiple passes over the live
789 heap in order to discover the full retainer set for each
790 object, which can be quite slow. So we set a limit on the
791 maximum size of a retainer set, where all retainer sets larger
792 than the maximum retainer set size are replaced by the special
793 set <literal>MANY</literal>. The maximum set size defaults to 8
794 and can be altered with the <option>-R</option> RTS
799 <term><option>-R</option><replaceable>size</replaceable></term>
801 <para>Restrict the number of elements in a retainer set to
802 <replaceable>size</replaceable> (default 8).</para>
808 <title>Hints for using retainer profiling</title>
810 <para>The definition of retainers is designed to reflect a
811 common cause of space leaks: a large structure is retained by
812 an unevaluated computation, and will be released once the
813 computation is forced. A good example is looking up a value in
814 a finite map, where unless the lookup is forced in a timely
815 manner the unevaluated lookup will cause the whole mapping to
816 be retained. These kind of space leaks can often be
817 eliminated by forcing the relevant computations to be
818 performed eagerly, using <literal>seq</literal> or strictness
819 annotations on data constructor fields.</para>
821 <para>Often a particular data structure is being retained by a
822 chain of unevaluated closures, only the nearest of which will
823 be reported by retainer profiling - for example A retains B, B
824 retains C, and C retains a large structure. There might be a
825 large number of Bs but only a single A, so A is really the one
826 we're interested in eliminating. However, retainer profiling
827 will in this case report B as the retainer of the large
828 structure. To move further up the chain of retainers, we can
829 ask for another retainer profile but this time restrict the
830 profile to B objects, so we get a profile of the retainers of
834 <replaceable>prog</replaceable> +RTS -hr -hcB
837 <para>This trick isn't foolproof, because there might be other
838 B closures in the heap which aren't the retainers we are
839 interested in, but we've found this to be a useful technique
840 in most cases.</para>
844 <sect2 id="biography-prof">
845 <title>Biographical Profiling</title>
847 <para>A typical heap object may be in one of the following four
848 states at each point in its lifetime:</para>
852 <para>The <firstterm>lag</firstterm> stage, which is the
853 time between creation and the first use of the
857 <para>the <firstterm>use</firstterm> stage, which lasts from
858 the first use until the last use of the object, and</para>
861 <para>The <firstterm>drag</firstterm> stage, which lasts
862 from the final use until the last reference to the object
866 <para>An object which is never used is said to be in the
867 <firstterm>void</firstterm> state for its whole
872 <para>A biographical heap profile displays the portion of the
873 live heap in each of the four states listed above. Usually the
874 most interesting states are the void and drag states: live heap
875 in these states is more likely to be wasted space than heap in
876 the lag or use states.</para>
878 <para>It is also possible to break down the heap in one or more
879 of these states by a different criteria, by restricting a
880 profile by biography. For example, to show the portion of the
881 heap in the drag or void state by producer: </para>
884 <replaceable>prog</replaceable> +RTS -hc -hbdrag,void
887 <para>Once you know the producer or the type of the heap in the
888 drag or void states, the next step is usually to find the
892 <replaceable>prog</replaceable> +RTS -hr -hc<replaceable>cc</replaceable>...
895 <para>NOTE: this two stage process is required because GHC
896 cannot currently profile using both biographical and retainer
897 information simultaneously.</para>
900 <sect2 id="mem-residency">
901 <title>Actual memory residency</title>
903 <para>How does the heap residency reported by the heap profiler relate to
904 the actual memory residency of your program when you run it? You might
905 see a large discrepancy between the residency reported by the heap
906 profiler, and the residency reported by tools on your system
907 (eg. <literal>ps</literal> or <literal>top</literal> on Unix, or the
908 Task Manager on Windows). There are several reasons for this:</para>
912 <para>There is an overhead of profiling itself, which is subtracted
913 from the residency figures by the profiler. This overhead goes
914 away when compiling without profiling support, of course. The
915 space overhead is currently 2 extra
916 words per heap object, which probably results in
917 about a 30% overhead.</para>
921 <para>Garbage collection requires more memory than the actual
922 residency. The factor depends on the kind of garbage collection
923 algorithm in use: a major GC in the standard
924 generation copying collector will usually require 3L bytes of
925 memory, where L is the amount of live data. This is because by
926 default (see the <option>+RTS -F</option> option) we allow the old
927 generation to grow to twice its size (2L) before collecting it, and
928 we require additionally L bytes to copy the live data into. When
929 using compacting collection (see the <option>+RTS -c</option>
930 option), this is reduced to 2L, and can further be reduced by
931 tweaking the <option>-F</option> option. Also add the size of the
932 allocation area (currently a fixed 512Kb).</para>
936 <para>The stack isn't counted in the heap profile by default. See the
937 <option>+RTS -xt</option> option.</para>
941 <para>The program text itself, the C stack, any non-heap data (eg. data
942 allocated by foreign libraries, and data allocated by the RTS), and
943 <literal>mmap()</literal>'d memory are not counted in the heap profile.</para>
951 <title><command>hp2ps</command>––heap profile to PostScript</title>
953 <indexterm><primary><command>hp2ps</command></primary></indexterm>
954 <indexterm><primary>heap profiles</primary></indexterm>
955 <indexterm><primary>postscript, from heap profiles</primary></indexterm>
956 <indexterm><primary><option>-h<break-down></option></primary></indexterm>
961 hp2ps [flags] [<file>[.hp]]
965 <command>hp2ps</command><indexterm><primary>hp2ps
966 program</primary></indexterm> converts a heap profile as produced
967 by the <option>-h<break-down></option> runtime option into a
968 PostScript graph of the heap profile. By convention, the file to
969 be processed by <command>hp2ps</command> has a
970 <filename>.hp</filename> extension. The PostScript output is
971 written to <filename><file>@.ps</filename>. If
972 <filename><file></filename> is omitted entirely, then the
973 program behaves as a filter.</para>
975 <para><command>hp2ps</command> is distributed in
976 <filename>ghc/utils/hp2ps</filename> in a GHC source
977 distribution. It was originally developed by Dave Wakeling as part
978 of the HBC/LML heap profiler.</para>
980 <para>The flags are:</para>
985 <term><option>-d</option></term>
987 <para>In order to make graphs more readable,
988 <command>hp2ps</command> sorts the shaded bands for each
989 identifier. The default sort ordering is for the bands with
990 the largest area to be stacked on top of the smaller ones.
991 The <option>-d</option> option causes rougher bands (those
992 representing series of values with the largest standard
993 deviations) to be stacked on top of smoother ones.</para>
998 <term><option>-b</option></term>
1000 <para>Normally, <command>hp2ps</command> puts the title of
1001 the graph in a small box at the top of the page. However, if
1002 the JOB string is too long to fit in a small box (more than
1003 35 characters), then <command>hp2ps</command> will choose to
1004 use a big box instead. The <option>-b</option> option
1005 forces <command>hp2ps</command> to use a big box.</para>
1010 <term><option>-e<float>[in|mm|pt]</option></term>
1012 <para>Generate encapsulated PostScript suitable for
1013 inclusion in LaTeX documents. Usually, the PostScript graph
1014 is drawn in landscape mode in an area 9 inches wide by 6
1015 inches high, and <command>hp2ps</command> arranges for this
1016 area to be approximately centred on a sheet of a4 paper.
1017 This format is convenient of studying the graph in detail,
1018 but it is unsuitable for inclusion in LaTeX documents. The
1019 <option>-e</option> option causes the graph to be drawn in
1020 portrait mode, with float specifying the width in inches,
1021 millimetres or points (the default). The resulting
1022 PostScript file conforms to the Encapsulated PostScript
1023 (EPS) convention, and it can be included in a LaTeX document
1024 using Rokicki's dvi-to-PostScript converter
1025 <command>dvips</command>.</para>
1030 <term><option>-g</option></term>
1032 <para>Create output suitable for the <command>gs</command>
1033 PostScript previewer (or similar). In this case the graph is
1034 printed in portrait mode without scaling. The output is
1035 unsuitable for a laser printer.</para>
1040 <term><option>-l</option></term>
1042 <para>Normally a profile is limited to 20 bands with
1043 additional identifiers being grouped into an
1044 <literal>OTHER</literal> band. The <option>-l</option> flag
1045 removes this 20 band and limit, producing as many bands as
1046 necessary. No key is produced as it won't fit!. It is useful
1047 for creation time profiles with many bands.</para>
1052 <term><option>-m<int></option></term>
1054 <para>Normally a profile is limited to 20 bands with
1055 additional identifiers being grouped into an
1056 <literal>OTHER</literal> band. The <option>-m</option> flag
1057 specifies an alternative band limit (the maximum is
1060 <para><option>-m0</option> requests the band limit to be
1061 removed. As many bands as necessary are produced. However no
1062 key is produced as it won't fit! It is useful for displaying
1063 creation time profiles with many bands.</para>
1068 <term><option>-p</option></term>
1070 <para>Use previous parameters. By default, the PostScript
1071 graph is automatically scaled both horizontally and
1072 vertically so that it fills the page. However, when
1073 preparing a series of graphs for use in a presentation, it
1074 is often useful to draw a new graph using the same scale,
1075 shading and ordering as a previous one. The
1076 <option>-p</option> flag causes the graph to be drawn using
1077 the parameters determined by a previous run of
1078 <command>hp2ps</command> on <filename>file</filename>. These
1079 are extracted from <filename>file@.aux</filename>.</para>
1084 <term><option>-s</option></term>
1086 <para>Use a small box for the title.</para>
1091 <term><option>-t<float></option></term>
1093 <para>Normally trace elements which sum to a total of less
1094 than 1% of the profile are removed from the
1095 profile. The <option>-t</option> option allows this
1096 percentage to be modified (maximum 5%).</para>
1098 <para><option>-t0</option> requests no trace elements to be
1099 removed from the profile, ensuring that all the data will be
1105 <term><option>-c</option></term>
1107 <para>Generate colour output.</para>
1112 <term><option>-y</option></term>
1114 <para>Ignore marks.</para>
1119 <term><option>-?</option></term>
1121 <para>Print out usage information.</para>
1127 <sect2 id="manipulating-hp">
1128 <title>Manipulating the hp file</title>
1130 <para>(Notes kindly offered by Jan-Willhem Maessen.)</para>
1133 The <filename>FOO.hp</filename> file produced when you ask for the
1134 heap profile of a program <filename>FOO</filename> is a text file with a particularly
1135 simple structure. Here's a representative example, with much of the
1136 actual data omitted:
1139 DATE "Thu Dec 26 18:17 2002"
1140 SAMPLE_UNIT "seconds"
1151 BEGIN_SAMPLE 11695.47
1154 The first four lines (<literal>JOB</literal>, <literal>DATE</literal>, <literal>SAMPLE_UNIT</literal>, <literal>VALUE_UNIT</literal>) form a
1155 header. Each block of lines starting with <literal>BEGIN_SAMPLE</literal> and ending
1156 with <literal>END_SAMPLE</literal> forms a single sample (you can think of this as a
1157 vertical slice of your heap profile). The hp2ps utility should accept
1158 any input with a properly-formatted header followed by a series of
1164 <title>Zooming in on regions of your profile</title>
1167 You can look at particular regions of your profile simply by loading a
1168 copy of the <filename>.hp</filename> file into a text editor and deleting the unwanted
1169 samples. The resulting <filename>.hp</filename> file can be run through <command>hp2ps</command> and viewed
1175 <title>Viewing the heap profile of a running program</title>
1178 The <filename>.hp</filename> file is generated incrementally as your
1179 program runs. In principle, running <command>hp2ps</command> on the incomplete file
1180 should produce a snapshot of your program's heap usage. However, the
1181 last sample in the file may be incomplete, causing <command>hp2ps</command> to fail. If
1182 you are using a machine with UNIX utilities installed, it's not too
1183 hard to work around this problem (though the resulting command line
1184 looks rather Byzantine):
1186 head -`fgrep -n END_SAMPLE FOO.hp | tail -1 | cut -d : -f 1` FOO.hp \
1190 The command <command>fgrep -n END_SAMPLE FOO.hp</command> finds the
1191 end of every complete sample in <filename>FOO.hp</filename>, and labels each sample with
1192 its ending line number. We then select the line number of the last
1193 complete sample using <command>tail</command> and <command>cut</command>. This is used as a
1194 parameter to <command>head</command>; the result is as if we deleted the final
1195 incomplete sample from <filename>FOO.hp</filename>. This results in a properly-formatted
1196 .hp file which we feed directly to <command>hp2ps</command>.
1200 <title>Viewing a heap profile in real time</title>
1203 The <command>gv</command> and <command>ghostview</command> programs
1204 have a "watch file" option can be used to view an up-to-date heap
1205 profile of your program as it runs. Simply generate an incremental
1206 heap profile as described in the previous section. Run <command>gv</command> on your
1209 gv -watch -seascape FOO.ps
1211 If you forget the <literal>-watch</literal> flag you can still select
1212 "Watch file" from the "State" menu. Now each time you generate a new
1213 profile <filename>FOO.ps</filename> the view will update automatically.
1217 This can all be encapsulated in a little script:
1220 head -`fgrep -n END_SAMPLE FOO.hp | tail -1 | cut -d : -f 1` FOO.hp \
1222 gv -watch -seascape FOO.ps &
1224 sleep 10 # We generate a new profile every 10 seconds.
1225 head -`fgrep -n END_SAMPLE FOO.hp | tail -1 | cut -d : -f 1` FOO.hp \
1229 Occasionally <command>gv</command> will choke as it tries to read an incomplete copy of
1230 <filename>FOO.ps</filename> (because <command>hp2ps</command> is still running as an update
1231 occurs). A slightly more complicated script works around this
1232 problem, by using the fact that sending a SIGHUP to gv will cause it
1233 to re-read its input file:
1236 head -`fgrep -n END_SAMPLE FOO.hp | tail -1 | cut -d : -f 1` FOO.hp \
1242 head -`fgrep -n END_SAMPLE FOO.hp | tail -1 | cut -d : -f 1` FOO.hp \
1252 <title>Observing Code Coverage</title>
1253 <indexterm><primary>code coverage</primary></indexterm>
1254 <indexterm><primary>Haskell Program Coverage</primary></indexterm>
1255 <indexterm><primary>hpc</primary></indexterm>
1258 Code coverage tools allow a programer to determine what parts of
1259 their code have been actually executed, and which parts have
1260 never actually been invoked. GHC has an option for generating
1261 instrumented code that records code coverage as part of the
1262 <ulink url="http://www.haskell.org/hpc">Haskell Program Coverage
1263 </ulink>(HPC) toolkit. HPC tools can be used to render the
1264 outputed code coverage infomation into human understandable
1269 HPC provides coverage information of two kinds: source coverage
1270 and boolean-control coverage. Source coverage is the extent to
1271 which every part of the program was used, measured at three
1272 different levels: declarations (both top-level and local),
1273 alternatives (among several equations or case branches) and
1274 expressions (at every level). Boolean coverage is the extent to
1275 which each of the values True and False is obtained in every
1276 syntactic boolean context (ie. guard, condition, qualifier).
1280 HPC displays both kinds of information in two different ways:
1281 textual reports with summary statistics (hpc-report) and sources
1282 with color mark-up (hpc-markup). For boolean coverage, there
1283 are four possible outcomes for each guard, condition or
1284 qualifier: both True and False values occur; only True; only
1285 False; never evaluated. In hpc-markup output, highlighting with
1286 a yellow background indicates a part of the program that was
1287 never evaluated; a green background indicates an always-True
1288 expression and a red background indicates an always-False one.
1291 <sect2><title>A small example: Reciprocation</title>
1294 For an example we have a program which computes exact decimal
1295 representations of reciprocals, with recurring parts indicated in
1296 brackets. We first build an instrumented version using the
1297 hpc-build script. Assuming the source file is Recip.hs.
1300 reciprocal :: Int -> (String, Int)
1301 reciprocal n | n > 1 = ('0' : '.' : digits, recur)
1303 "attempting to compute reciprocal of number <= 1"
1305 (digits, recur) = divide n 1 []
1306 divide :: Int -> Int -> [Int] -> (String, Int)
1307 divide n c cs | c `elem` cs = ([], position c cs)
1308 | r == 0 = (show q, 0)
1309 | r /= 0 = (show q ++ digits, recur)
1311 (q, r) = (c*10) `quotRem` n
1312 (digits, recur) = divide n r (c:cs)
1314 position :: Int -> [Int] -> Int
1315 position n (x:xs) | n==x = 1
1316 | otherwise = 1 + position n xs
1318 showRecip :: Int -> String
1320 "1/" ++ show n ++ " = " ++
1321 if r==0 then d else take p d ++ "(" ++ drop p d ++ ")"
1324 (d, r) = reciprocal n
1328 putStrLn (showRecip number)
1332 ` <para>The HPC intrumentation is enabled using the -fhpc flag.
1336 $ ghc -fhpc Recip.hs --make
1338 <para>HPC index (.mix) files are placed placed in .hpc subdirectory. These can be considered like
1339 the .hi files for HPC. They contain information about what parts of the haskell each modules.
1346 <para>Now for a textual summary of coverage:</para>
1349 80% expressions used (81/101)
1350 12% boolean coverage (1/8)
1351 14% guards (1/7), 3 always True,
1354 0% 'if' conditions (0/1), 1 always False
1355 100% qualifiers (0/0)
1356 55% alternatives used (5/9)
1357 100% local declarations used (9/9)
1358 100% top-level declarations used (5/5)
1360 <para>Finally, we generate a marked-up version of the source.</para>
1363 writing Recip.hs.html
1366 <title>Recip.hs.html</title>
1367 <graphic fileref="images/Recip.png"></graphic>
1373 <sect1 id="ticky-ticky">
1374 <title>Using “ticky-ticky” profiling (for implementors)</title>
1375 <indexterm><primary>ticky-ticky profiling</primary></indexterm>
1377 <para>(ToDo: document properly.)</para>
1379 <para>It is possible to compile Glasgow Haskell programs so that
1380 they will count lots and lots of interesting things, e.g., number
1381 of updates, number of data constructors entered, etc., etc. We
1382 call this “ticky-ticky”
1383 profiling,<indexterm><primary>ticky-ticky
1384 profiling</primary></indexterm> <indexterm><primary>profiling,
1385 ticky-ticky</primary></indexterm> because that's the sound a Sun4
1386 makes when it is running up all those counters
1387 (<emphasis>slowly</emphasis>).</para>
1389 <para>Ticky-ticky profiling is mainly intended for implementors;
1390 it is quite separate from the main “cost-centre”
1391 profiling system, intended for all users everywhere.</para>
1393 <para>To be able to use ticky-ticky profiling, you will need to
1394 have built the ticky RTS. (This should be described in
1395 the building guide, but amounts to building the RTS with way
1396 "t" enabled.)</para>
1398 <para>To get your compiled program to spit out the ticky-ticky
1399 numbers, use a <option>-r</option> RTS
1400 option<indexterm><primary>-r RTS option</primary></indexterm>.
1401 See <xref linkend="runtime-control"/>.</para>
1403 <para>Compiling your program with the <option>-ticky</option>
1404 switch yields an executable that performs these counts. Here is a
1405 sample ticky-ticky statistics file, generated by the invocation
1406 <command>foo +RTS -rfoo.ticky</command>.</para>
1409 foo +RTS -rfoo.ticky
1412 ALLOCATIONS: 3964631 (11330900 words total: 3999476 admin, 6098829 goods, 1232595 slop)
1413 total words: 2 3 4 5 6+
1414 69647 ( 1.8%) function values 50.0 50.0 0.0 0.0 0.0
1415 2382937 ( 60.1%) thunks 0.0 83.9 16.1 0.0 0.0
1416 1477218 ( 37.3%) data values 66.8 33.2 0.0 0.0 0.0
1417 0 ( 0.0%) big tuples
1418 2 ( 0.0%) black holes 0.0 100.0 0.0 0.0 0.0
1419 0 ( 0.0%) prim things
1420 34825 ( 0.9%) partial applications 0.0 0.0 0.0 100.0 0.0
1421 2 ( 0.0%) thread state objects 0.0 0.0 0.0 0.0 100.0
1423 Total storage-manager allocations: 3647137 (11882004 words)
1424 [551104 words lost to speculative heap-checks]
1428 ENTERS: 9400092 of which 2005772 (21.3%) direct to the entry code
1429 [the rest indirected via Node's info ptr]
1430 1860318 ( 19.8%) thunks
1431 3733184 ( 39.7%) data values
1432 3149544 ( 33.5%) function values
1433 [of which 1999880 (63.5%) bypassed arg-satisfaction chk]
1434 348140 ( 3.7%) partial applications
1435 308906 ( 3.3%) normal indirections
1436 0 ( 0.0%) permanent indirections
1439 2137257 ( 36.4%) from entering a new constructor
1440 [the rest from entering an existing constructor]
1441 2349219 ( 40.0%) vectored [the rest unvectored]
1443 RET_NEW: 2137257: 32.5% 46.2% 21.3% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
1444 RET_OLD: 3733184: 2.8% 67.9% 29.3% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
1445 RET_UNBOXED_TUP: 2: 0.0% 0.0%100.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
1447 RET_VEC_RETURN : 2349219: 0.0% 0.0%100.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
1449 UPDATE FRAMES: 2241725 (0 omitted from thunks)
1453 0 ( 0.0%) data values
1454 34827 ( 1.6%) partial applications
1455 [2 in place, 34825 allocated new space]
1456 2206898 ( 98.4%) updates to existing heap objects (46 by squeezing)
1457 UPD_CON_IN_NEW: 0: 0 0 0 0 0 0 0 0 0
1458 UPD_PAP_IN_NEW: 34825: 0 0 0 34825 0 0 0 0 0
1460 NEW GEN UPDATES: 2274700 ( 99.9%)
1462 OLD GEN UPDATES: 1852 ( 0.1%)
1464 Total bytes copied during GC: 190096
1466 **************************************************
1467 3647137 ALLOC_HEAP_ctr
1468 11882004 ALLOC_HEAP_tot
1473 34831 ALLOC_FUN_hst_0
1474 34816 ALLOC_FUN_hst_1
1478 2382937 ALLOC_UP_THK_ctr
1481 0 E!NT_PERM_IND_ctr requires +RTS -Z
1482 [... lots more info omitted ...]
1483 0 GC_SEL_ABANDONED_ctr
1486 0 GC_FAILED_PROMOTION_ctr
1487 47524 GC_WORDS_COPIED_ctr
1490 <para>The formatting of the information above the row of asterisks
1491 is subject to change, but hopefully provides a useful
1492 human-readable summary. Below the asterisks <emphasis>all
1493 counters</emphasis> maintained by the ticky-ticky system are
1494 dumped, in a format intended to be machine-readable: zero or more
1495 spaces, an integer, a space, the counter name, and a newline.</para>
1497 <para>In fact, not <emphasis>all</emphasis> counters are
1498 necessarily dumped; compile- or run-time flags can render certain
1499 counters invalid. In this case, either the counter will simply
1500 not appear, or it will appear with a modified counter name,
1501 possibly along with an explanation for the omission (notice
1502 <literal>ENT_PERM_IND_ctr</literal> appears
1503 with an inserted <literal>!</literal> above). Software analysing
1504 this output should always check that it has the counters it
1505 expects. Also, beware: some of the counters can have
1506 <emphasis>large</emphasis> values!</para>
1513 ;;; Local Variables: ***
1515 ;;; sgml-parent-document: ("users_guide.xml" "book" "chapter") ***