1 | #! /usr/bin/env perl |
---|
2 | |
---|
3 | # ***************************************************************************** |
---|
4 | # Copyright (C) 2015, UChicago Argonne, LLC |
---|
5 | # All Rights Reserved |
---|
6 | # memlog (ANL-SF-15-081) |
---|
7 | # Hal Finkel, Argonne National Laboratory |
---|
8 | # |
---|
9 | # OPEN SOURCE LICENSE |
---|
10 | # |
---|
11 | # Under the terms of Contract No. DE-AC02-06CH11357 with UChicago Argonne, LLC, |
---|
12 | # the U.S. Government retains certain rights in this software. |
---|
13 | # |
---|
14 | # Redistribution and use in source and binary forms, with or without |
---|
15 | # modification, are permitted provided that the following conditions are met: |
---|
16 | # |
---|
17 | # 1. Redistributions of source code must retain the above copyright notice, this |
---|
18 | # list of conditions and the following disclaimer. |
---|
19 | # |
---|
20 | # 2. Redistributions in binary form must reproduce the above copyright notice, |
---|
21 | # this list of conditions and the following disclaimer in the documentation |
---|
22 | # and/or other materials provided with the distribution. |
---|
23 | # |
---|
24 | # 3. Neither the names of UChicago Argonne, LLC or the Department of Energy nor |
---|
25 | # the names of its contributors may be used to endorse or promote products |
---|
26 | # derived from this software without specific prior written permission. |
---|
27 | # |
---|
28 | # ***************************************************************************** |
---|
29 | # DISCLAIMER |
---|
30 | # |
---|
31 | # THE SOFTWARE IS SUPPLIED “AS IS” WITHOUT WARRANTY OF ANY KIND. |
---|
32 | # |
---|
33 | # NEITHER THE UNTED STATES GOVERNMENT, NOR THE UNITED STATES DEPARTMENT OF |
---|
34 | # ENERGY, NOR UCHICAGO ARGONNE, LLC, NOR ANY OF THEIR EMPLOYEES, MAKES ANY |
---|
35 | # WARRANTY, EXPRESS OR IMPLIED, OR ASSUMES ANY LEGAL LIABILITY OR RESPONSIBILITY |
---|
36 | # FOR THE ACCURACY, COMPLETENESS, OR USEFULNESS OF ANY INFORMATION, DATA, |
---|
37 | # APPARATUS, PRODUCT, OR PROCESS DISCLOSED, OR REPRESENTS THAT ITS USE WOULD NOT |
---|
38 | # INFRINGE PRIVATELY OWNED RIGHTS. |
---|
39 | # |
---|
40 | # ***************************************************************************** |
---|
41 | |
---|
42 | use strict; |
---|
43 | use File::Basename; |
---|
44 | |
---|
45 | my $memlog_fn = $ARGV[0]; |
---|
46 | my $out_dir = $ARGV[1] || '.'; |
---|
47 | |
---|
48 | my $print_raw_proc_name = 0; |
---|
49 | |
---|
50 | if (! -f $memlog_fn) { |
---|
51 | print "Usage: $0 <memlog file> [<output directory>]\n"; |
---|
52 | exit 1; |
---|
53 | } |
---|
54 | |
---|
55 | # The version of addr2line and friends that you use can make a big difference, |
---|
56 | # especially on BE ppc64, where older versions of addr2line did not account |
---|
57 | # correctly for the function descriptor setup. At ALCF, on the BG/Q, we have |
---|
58 | # newer versions not in the default search path (from bgclang). |
---|
59 | my $alcf_bu_dir = '/soft/compilers/bgclang/current/binutils/bin'; |
---|
60 | if (-d $alcf_bu_dir) { |
---|
61 | $ENV{'PATH'} = $alcf_bu_dir . ':' . $ENV{'PATH'}; |
---|
62 | } |
---|
63 | |
---|
64 | open(MEMLOG, $memlog_fn) || die "Can't open $memlog_fn: $!"; |
---|
65 | |
---|
66 | # The first step is to determine the high-water mark. |
---|
67 | my $max_rss = 0; |
---|
68 | foreach my $line (<MEMLOG>) { |
---|
69 | chomp($line); |
---|
70 | my @parts = split(/\t/, $line); |
---|
71 | |
---|
72 | my $op = shift(@parts); |
---|
73 | my $state = shift(@parts); |
---|
74 | |
---|
75 | my ($time, $then_max_rss, $tid) = split(/\s+/, $state); |
---|
76 | if ($max_rss < $then_max_rss) { |
---|
77 | $max_rss = $then_max_rss; |
---|
78 | } |
---|
79 | } |
---|
80 | |
---|
81 | seek(MEMLOG, 0, 0); |
---|
82 | |
---|
83 | # Scan the log for malloc/free pairings. We're interested only in active |
---|
84 | # allocations at the time when the rss reaches the final maxrss. |
---|
85 | my $max_rss_time = 0; |
---|
86 | my %malloc_lines; |
---|
87 | foreach my $line (<MEMLOG>) { |
---|
88 | chomp($line); |
---|
89 | my @parts = split(/\t/, $line); |
---|
90 | |
---|
91 | my $op = shift(@parts); |
---|
92 | my $state = shift(@parts); |
---|
93 | |
---|
94 | if ($op =~ /^M:/) { |
---|
95 | my ($size, $ptr) = ($op =~ /^M: (\d+) 0x(\w+)/); |
---|
96 | $malloc_lines{$ptr} = $line; |
---|
97 | } elsif ($op =~ /^F:/) { |
---|
98 | my ($ptr) = ($op =~ /^F: 0x(\w+)/); |
---|
99 | delete $malloc_lines{$ptr}; |
---|
100 | } else { |
---|
101 | next; |
---|
102 | } |
---|
103 | |
---|
104 | # If we've reached the max rss, we've seen all we need to see. |
---|
105 | my ($time, $then_max_rss, $tid) = split(/\s+/, $state); |
---|
106 | $max_rss_time = $time; |
---|
107 | if ($then_max_rss == $max_rss) { |
---|
108 | last; |
---|
109 | } |
---|
110 | } |
---|
111 | |
---|
112 | close(MEMLOG); |
---|
113 | |
---|
114 | # Convert maxrss, currently in KB, to bytes. |
---|
115 | $max_rss *= 1024; |
---|
116 | |
---|
117 | my $total_size = 0; |
---|
118 | my %roots; |
---|
119 | my %all_nodes; |
---|
120 | foreach my $line (values %malloc_lines) { |
---|
121 | my @parts = split(/\t/, $line); |
---|
122 | |
---|
123 | my $op = shift(@parts); |
---|
124 | my $state = shift(@parts); |
---|
125 | |
---|
126 | # Only dealing with allocations here... |
---|
127 | if ($op !~ /^M:/) { |
---|
128 | next; |
---|
129 | } |
---|
130 | |
---|
131 | my ($size, $ptr) = ($op =~ /^M: (\d+) 0x(\w+)/); |
---|
132 | my ($time, $then_max_rss, $tid) = split(/\s+/, $state); |
---|
133 | |
---|
134 | $total_size += $size; |
---|
135 | |
---|
136 | sub level_parts($) { |
---|
137 | my $level = @_[0]; |
---|
138 | my ($file_name, $proc_name, $off, $pc, $relpc) = |
---|
139 | ($level =~ /^(.*) \((.*)\+0x(\w+)\) \[0x(\w+) \(0x(\w+)\)\]/); |
---|
140 | |
---|
141 | return ($file_name, $proc_name, $off, $pc, $relpc); |
---|
142 | } |
---|
143 | |
---|
144 | # Put the top of the stack first. |
---|
145 | @parts = reverse(@parts); |
---|
146 | |
---|
147 | my $parent = \%roots; |
---|
148 | for (my $i = 0; $i < scalar(@parts); ++$i) { |
---|
149 | my $level = $parts[$i]; |
---|
150 | my ($file_name, $proc_name, $off, $pc, $relpc) = level_parts($level); |
---|
151 | |
---|
152 | # Skip this level if we don't even know from what file it came. |
---|
153 | if ($file_name eq '?') { |
---|
154 | next; |
---|
155 | } |
---|
156 | |
---|
157 | # print STDERR "parsed: $file_name, $proc_name, $off, $pc, $relpc\n"; |
---|
158 | |
---|
159 | if (!exists $all_nodes{$pc}) { |
---|
160 | $all_nodes{$pc}->{'file_name'} = $file_name; |
---|
161 | $all_nodes{$pc}->{'proc_name'} = $proc_name; |
---|
162 | $all_nodes{$pc}->{'off'} = $off; |
---|
163 | $all_nodes{$pc}->{'pc'} = $pc; |
---|
164 | $all_nodes{$pc}->{'relpc'} = $relpc; |
---|
165 | } |
---|
166 | |
---|
167 | if (!exists $parent->{$pc}) { |
---|
168 | $parent->{$pc} = $all_nodes{$pc}; |
---|
169 | } |
---|
170 | |
---|
171 | $parent->{$pc}->{'size'} += $size; |
---|
172 | |
---|
173 | my ($next_file_name, $next_proc_name, $next_off, $next_pc, $next_relpc); |
---|
174 | if ($i < scalar(@parts)-1) { |
---|
175 | my $next_level = $parts[$i+1]; |
---|
176 | ($next_file_name, $next_proc_name, $next_off, $next_pc, $next_relpc) = |
---|
177 | level_parts($next_level); |
---|
178 | $parent->{$pc}->{'child_sizes'}->{$next_pc} += $size; |
---|
179 | } |
---|
180 | |
---|
181 | if (!exists $parent->{'children'}) { |
---|
182 | $parent->{'children'} = {}; |
---|
183 | } |
---|
184 | |
---|
185 | $parent = $parent->{'children'}; |
---|
186 | } |
---|
187 | } |
---|
188 | |
---|
189 | my $txt_fn = "$out_dir/" . basename($memlog_fn) . ".txt"; |
---|
190 | my $dot_fn = "$out_dir/" . basename($memlog_fn) . ".dot"; |
---|
191 | my $ps_fn = "$out_dir/" . basename($memlog_fn) . ".ps"; |
---|
192 | my $pdf_fn = "$out_dir/" . basename($memlog_fn) . ".pdf"; |
---|
193 | |
---|
194 | open(TXT, ">$txt_fn") || die "Can't open $txt_fn: $!"; |
---|
195 | open(DOT, ">$dot_fn") || die "Can't open $dot_fn: $!"; |
---|
196 | |
---|
197 | sub format_bytes($) { |
---|
198 | my @sizes = qw( B KB MB GB TB PB ); |
---|
199 | my $size = $_[0]; |
---|
200 | |
---|
201 | my $i = 0; |
---|
202 | while ($size > 1024) { |
---|
203 | $size /= 1024; |
---|
204 | ++$i; |
---|
205 | } |
---|
206 | |
---|
207 | return sprintf("%.3f $sizes[$i]", $size); |
---|
208 | } |
---|
209 | |
---|
210 | print DOT ("digraph \"memlog\" {\n"); |
---|
211 | print DOT ("size=\"8,11\";\n"); |
---|
212 | print DOT ("node [width=0.375,height=0.25];\n"); |
---|
213 | |
---|
214 | printf DOT ("Legend [shape=box, fontsize=100, shape=oval," . |
---|
215 | "label=\"Total: %s active at maxrss = %s after %s s\"];\n", |
---|
216 | format_bytes($total_size), format_bytes($max_rss), $max_rss_time); |
---|
217 | |
---|
218 | printf TXT ("memlog: Total: %s active at maxrss = %s after %s s\n\n", |
---|
219 | format_bytes($total_size), format_bytes($max_rss), $max_rss_time); |
---|
220 | |
---|
221 | my %cached_names; |
---|
222 | sub get_name($) { |
---|
223 | my $node = $_[0]; |
---|
224 | my $pc = $node->{'pc'}; |
---|
225 | |
---|
226 | if (exists $cached_names{$pc}) { |
---|
227 | return $cached_names{$pc}; |
---|
228 | } |
---|
229 | |
---|
230 | my $ret = ''; |
---|
231 | |
---|
232 | # Prefer the relative offset (that is what we want for shared libraries), but |
---|
233 | # if is not available, use the full offset (which is what we want for the |
---|
234 | # base executable). |
---|
235 | my $exe_off = $node->{'relpc'}; |
---|
236 | if (!$exe_off) { |
---|
237 | $exe_off = $pc; |
---|
238 | } |
---|
239 | |
---|
240 | my $file_name = $node->{'file_name'}; |
---|
241 | |
---|
242 | # If we don't have an absolute path, this is probably the base dynamic |
---|
243 | # executable, so ask the shell which one it used (not foolproof because we |
---|
244 | # might not have the same PATH now, but hopefully is generally the right |
---|
245 | # thing). |
---|
246 | if ($file_name !~ /^\//) { |
---|
247 | $file_name = `which '$file_name'`; |
---|
248 | chomp($file_name); |
---|
249 | } |
---|
250 | |
---|
251 | my ($func, $loc) = `addr2line -e '$file_name' -f 0x$exe_off`; |
---|
252 | chomp($func); |
---|
253 | chomp($loc); |
---|
254 | |
---|
255 | if ($func !~ /^\?/) { |
---|
256 | # In general, this function name might look something like: |
---|
257 | # 00000329.plt_call.wcsnrtombs@@GLIBC_2.3+0 |
---|
258 | $func =~ s/@.*//; # Remove trailing symbol version strings |
---|
259 | $func =~ s/.*\.//; |
---|
260 | $func = `c++filt '$func'`; |
---|
261 | chomp($func); |
---|
262 | |
---|
263 | # It sometimes happens that addr2line is a bit too smart: when debugging |
---|
264 | # information is available, it might print a local alias for the |
---|
265 | # function instead of the full name (for example, printing 'List' |
---|
266 | # instead of 'Foo::List<int>::List(int, int const&)'). |
---|
267 | if ($node->{'proc_name'} ne '?') { |
---|
268 | my $proc_name = $node->{'proc_name'}; |
---|
269 | $proc_name = `c++filt '$proc_name'`; |
---|
270 | chomp($proc_name); |
---|
271 | |
---|
272 | if (length($proc_name) > length($func)) { |
---|
273 | $func = $proc_name; |
---|
274 | } |
---|
275 | } |
---|
276 | |
---|
277 | $ret .= $func . '\n'; |
---|
278 | |
---|
279 | if ($loc !~ /^\?/) { |
---|
280 | $ret .= $loc . '\n'; |
---|
281 | } |
---|
282 | } elsif ($node->{'proc_name'} ne '?') { |
---|
283 | my $proc_name = $node->{'proc_name'}; |
---|
284 | $proc_name = `c++filt '$proc_name'`; |
---|
285 | chomp($proc_name); |
---|
286 | |
---|
287 | $ret .= $proc_name . '\n'; |
---|
288 | } |
---|
289 | |
---|
290 | $ret .= $node->{'file_name'}; |
---|
291 | if ($print_raw_proc_name and $node->{'proc_name'} ne '?') { |
---|
292 | $ret .= '\n' . $node->{'proc_name'} . '+0x' . $node->{'off'}; |
---|
293 | } |
---|
294 | |
---|
295 | $cached_names{$pc} = $ret; |
---|
296 | return $ret; |
---|
297 | } |
---|
298 | |
---|
299 | my $skip_frac = 0.01; |
---|
300 | my %skipped; |
---|
301 | |
---|
302 | foreach my $pc (keys %all_nodes) { |
---|
303 | my $node = $all_nodes{$pc}; |
---|
304 | |
---|
305 | my $local_size = $node->{'size'}; |
---|
306 | if ($local_size * 1.0 / $total_size < $skip_frac) { |
---|
307 | $skipped{$pc} = 1; |
---|
308 | next; |
---|
309 | } |
---|
310 | |
---|
311 | my $fs = 8.0; |
---|
312 | if ($local_size > 0) { |
---|
313 | $fs = 50.0 * (abs($local_size * 1.0 / $total_size))**0.125; |
---|
314 | } |
---|
315 | |
---|
316 | my $name = get_name($node); |
---|
317 | |
---|
318 | printf DOT ("N%s [label=\"%s\\n%s\", shape=box, fontsize=%.1f%s];\n", |
---|
319 | $pc, $name, format_bytes($local_size), $fs); |
---|
320 | } |
---|
321 | |
---|
322 | foreach my $pc (keys %all_nodes) { |
---|
323 | if ($skipped{$pc}) { |
---|
324 | next; |
---|
325 | } |
---|
326 | |
---|
327 | my $node = $all_nodes{$pc}; |
---|
328 | my $local_size = $node->{'size'}; |
---|
329 | |
---|
330 | foreach my $cpc (keys %{$node->{'child_sizes'}}) { |
---|
331 | if ($skipped{$cpc}) { |
---|
332 | next; |
---|
333 | } |
---|
334 | |
---|
335 | my $child_size = $node->{'child_sizes'}->{$cpc}; |
---|
336 | my $frac = $child_size * 1.0 / $local_size; |
---|
337 | |
---|
338 | my $weight = 100.0 * sqrt($frac); |
---|
339 | my $style = sprintf("setlinewidth(%f)", 8.0 * sqrt($frac)); |
---|
340 | |
---|
341 | my $fs = 40.0 * $frac**0.125; |
---|
342 | |
---|
343 | printf DOT ("N%s -> N%s [label=\"%s\", weight=%d, style=\"%s\", fontsize=%.1f];\n", |
---|
344 | $pc, $cpc, format_bytes($child_size), $weight, $style, $fs); |
---|
345 | } |
---|
346 | } |
---|
347 | |
---|
348 | print DOT ("}\n"); |
---|
349 | |
---|
350 | foreach my $pc (sort { $all_nodes{$::b}->{'size'} <=> |
---|
351 | $all_nodes{$::a}->{'size'} } keys %all_nodes) { |
---|
352 | if ($skipped{$pc}) { |
---|
353 | next; |
---|
354 | } |
---|
355 | |
---|
356 | print TXT ('*' x 80) . "\n\n"; |
---|
357 | |
---|
358 | my $node = $all_nodes{$pc}; |
---|
359 | |
---|
360 | my $local_size = $node->{'size'}; |
---|
361 | printf TXT ("%s - %.1f%%\n", format_bytes($local_size), |
---|
362 | $local_size * 100.0 / $total_size); |
---|
363 | |
---|
364 | my $name = get_name($node); |
---|
365 | $name =~ s/\\n/\n/g; |
---|
366 | |
---|
367 | print TXT "$name\n$pc\n"; |
---|
368 | |
---|
369 | print TXT "\nMEMORY ALLOCATED BY CALLEES:\n"; |
---|
370 | |
---|
371 | foreach my $cpc (sort { $node->{'child_sizes'}->{$::b} <=> |
---|
372 | $node->{'child_sizes'}->{$::a} } |
---|
373 | keys %{$node->{'child_sizes'}}) { |
---|
374 | if ($skipped{$cpc}) { |
---|
375 | next; |
---|
376 | } |
---|
377 | |
---|
378 | my $child_node = $all_nodes{$cpc}; |
---|
379 | my $child_size = $node->{'child_sizes'}->{$cpc}; |
---|
380 | |
---|
381 | printf TXT ("\t%s - %.1f%%\n", format_bytes($child_size), |
---|
382 | $child_size * 100.0 / $local_size); |
---|
383 | |
---|
384 | my $child_name = get_name($child_node); |
---|
385 | $child_name =~ s/\\n/\n\t/g; |
---|
386 | print TXT "\t$child_name\n\t$cpc\n"; |
---|
387 | |
---|
388 | print TXT "\n"; |
---|
389 | } |
---|
390 | |
---|
391 | print TXT "\n"; |
---|
392 | } |
---|
393 | |
---|
394 | close(TXT); |
---|
395 | close(DOT); |
---|
396 | |
---|
397 | system("dot -Tps2 < '$dot_fn' > '$ps_fn'"); |
---|
398 | system("ps2pdf '$ps_fn' '$pdf_fn'"); |
---|
399 | |
---|
400 | exit 0; |
---|
401 | |
---|