1

I have a couple processes running a tool I've written that are joined by pipes, and I would like to measure their collected memory usage with valgrind. So far, I have tried something like:

$ valgrind tool=massif trace-children=yes --peak-inaccuracy=0.5 --pages-as-heap=yes --massif-out-file=myProcesses".%p" myProcesses.script

Where myProcesses.script runs the equivalent of my tool foo twice, e.g.:

foo | foo > /dev/null

Valgrind doesn't seem to capture the collected memory usage of this the way I expect. If I use top to track this, I get (for sake of argument) 10% memory usage on the first foo, and then another 10% collects on the second foo before the myProcesses.script completes. This is the sort of thing I want to measure: the usage of both processes. Valgrind instead returns the following error:

Massif: ms_main.c:1891 (ms_new_mem_brk): Assertion 'VG_IS_PAGE_ALIGNED(len)' failed.

Is there a way to collect memory usage data for commands I'm using in a piped fashion (using valgrind)? Or a similar tool that I can use to accurately automate these measurements?

The numbers that top returns while polling seem hand-wavy, to me, and I am seeking accurate and repeatable measurements. If you have suggestions for alternative tools, I would welcome those, as well.

EDIT - Fixed typo with valgrind option.

EDIT 2 - For some reason, it appears that the option --pages-as-heap is giving us troubles with the binaries we're testing. Your examples run fine. A new page is created every time we enter a non-inlined function (stack overflows - heh). We wanted to count those, but they're relatively minor in the scale of memory usage we're testing. (Perhaps there aren't function calls in ls or less?) Removing --pages-as-heap helped get testing working again. Thanks to MrGomez for the great help.

animuson
  • 53,861
  • 28
  • 137
  • 147
Alex Reynolds
  • 95,983
  • 54
  • 240
  • 345
  • As for your page alignment issue, this is obviously related to failed paging behavior when a new memory breakpoint is set in massif. I cannot reproduce this issue on my end, but I speculate `--pages-as-heap=yes` is responsible for this in the context of your invocation on your system. Obviously, seeing why this is being hit in a debugger is the best course of action, should you wish to explore the error condition. (See my answer below.) – MrGomez Mar 24 '12 at 23:03

2 Answers2

3

With the correct valgrind version given in the errata, this seems to just work for me in Valgrind 3.6.1. My invocation:

<me>@harley:/tmp/test$ /usr/local/bin/valgrind --tool=massif \
  --trace-children=yes --peak-inaccuracy=0.5 --pages-as-heap=yes \
  --massif-out-file=myProcesses".%p" ./testscript.sh
==21067== Massif, a heap profiler
==21067== Copyright (C) 2003-2010, and GNU GPL'd, by Nicholas Nethercote
==21067== Using Valgrind-3.6.1 and LibVEX; rerun with -h for copyright info
==21067== Command: ./testscript.sh
==21067==
==21068== Massif, a heap profiler
==21068== Copyright (C) 2003-2010, and GNU GPL'd, by Nicholas Nethercote
==21068== Using Valgrind-3.6.1 and LibVEX; rerun with -h for copyright info
==21068== Command: /bin/ls
==21068==
==21070== Massif, a heap profiler
==21070== Copyright (C) 2003-2010, and GNU GPL'd, by Nicholas Nethercote
==21070== Using Valgrind-3.6.1 and LibVEX; rerun with -h for copyright info
==21069== Massif, a heap profiler
==21069== Copyright (C) 2003-2010, and GNU GPL'd, by Nicholas Nethercote
==21069== Using Valgrind-3.6.1 and LibVEX; rerun with -h for copyright info
==21069== Command: /bin/sleep 5
==21069==
==21070== Command: /usr/bin/less
==21070==
==21068==
(END) ==21069==
==21070==
==21067==

The contents of my test script, testscript.sh:

ls | sleep 5 | less

Sparse contents from one of the files generated by --massif-out-file=myProcesses".%p" (myProcesses.21055):

desc: --peak-inaccuracy=0.5 --pages-as-heap=yes --massif-out-file=myProcesses.%p
cmd: ./testscript.sh
time_unit: i
#-----------
snapshot=0
#-----------
time=0
mem_heap_B=110592
mem_heap_extra_B=0
mem_stacks_B=0
heap_tree=empty
#-----------
snapshot=1
#-----------
time=0
mem_heap_B=118784
mem_heap_extra_B=0
mem_stacks_B=0
heap_tree=empty
...
#-----------
snapshot=18
#-----------
time=108269
mem_heap_B=1708032
mem_heap_extra_B=0
mem_stacks_B=0
heap_tree=peak
n2: 1708032 (page allocation syscalls) mmap/mremap/brk, --alloc-fns, etc.
 n3: 1474560 0x4015E42: mmap (mmap.S:62)
  n1: 1425408 0x4005CAC: _dl_map_object_from_fd (dl-load.c:1209)
   n2: 1425408 0x4007109: _dl_map_object (dl-load.c:2250)
    n1: 1413120 0x400CEEA: openaux (dl-deps.c:65)
     n1: 1413120 0x400D834: _dl_catch_error (dl-error.c:178)
      n1: 1413120 0x400C1E0: _dl_map_object_deps (dl-deps.c:247)
       n1: 1413120 0x4002B59: dl_main (rtld.c:1780)
        n1: 1413120 0x40140C5: _dl_sysdep_start (dl-sysdep.c:243)
         n1: 1413120 0x4000C6B: _dl_start (rtld.c:333)
          n0: 1413120 0x4000855: ??? (in /lib/ld-2.11.1.so)
    n0: 12288 in 1 place, below massif's threshold (01.00%)
  n0: 28672 in 3 places, all below massif's threshold (01.00%)
  n1: 20480 0x4005E0C: _dl_map_object_from_fd (dl-load.c:1260)
   n1: 20480 0x4007109: _dl_map_object (dl-load.c:2250)
    n0: 20480 in 2 places, all below massif's threshold (01.00%)
 n0: 233472 0xFFFFFFFF: ???
#-----------
snapshot=19
#-----------
time=108269
mem_heap_B=1703936
mem_heap_extra_B=0
mem_stacks_B=0
heap_tree=empty
#-----------
snapshot=20
#-----------
time=200236
mem_heap_B=1839104
mem_heap_extra_B=0
mem_stacks_B=0
heap_tree=empty

Massif continues to complain about heap allocations in the remainder of my files. Note this is very similar to your error.

I theorize that your version of valgrind was built in debug mode, causing the asserts to fire. A rebuild from source (I used this with the defaults hanging off ./configure) will fix the issue.

Either way, this seems to be expected with Massif.

MrGomez
  • 23,788
  • 45
  • 72
  • I made a typo in copy-pasting the issue to this SO question. My test script does indeed use `trace-children` and not `track-children`. Thank you for catching this error. I have edited the question to fix the typo. – Alex Reynolds Mar 24 '12 at 23:56
  • No problem. As far as answering the question goes, please include the version of `valgrind` you're using, as well. This seems to be an issue with the disposition of your command line or program `foo`, not the lack of `valgrind`'s featurefulness for handling piped processes. – MrGomez Mar 24 '12 at 23:58
  • @AlexReynolds Your command works for me. I've updated my answer. – MrGomez Mar 25 '12 at 00:49
  • Interesting. Will research this more and hopefully have results soon. Thanks for investigating. – Alex Reynolds Mar 25 '12 at 00:52
  • @AlexReynolds Glad to help. Sometimes, all we need is someone with a different system configuration to help highlight the bugs in relief. – MrGomez Mar 25 '12 at 01:03
0

Some programs allow you to preload the libmemusage.so library and get a report of what memory allocations were allocated recorded:

$ LD_PRELOAD=libmemusage.so less /etc/passwd

Memory usage summary: heap total: 36212, heap peak: 35011, stack peak: 15008
         total calls   total memory   failed calls
 malloc|         39           5985              0
realloc|          3             64              0  (nomove:2, dec:0, free:0)
 calloc|        238          30163              0
   free|         51          11546
Histogram for block sizes:
    0-15            128  45% ==================================================
   16-31             13   4% =====
   32-47            105  37% =========================================
   48-63              2  <1% 
   64-79              4   1% =
   80-95              5   1% =
   96-111             3   1% =
  112-127             3   1% =
  160-175             1  <1% 
  192-207             1  <1% 
  208-223             2  <1% 
  256-271             1  <1% 
  432-447             1  <1% 
  560-575             1  <1% 
  656-671             1  <1% 
  768-783             1  <1% 
  944-959             1  <1% 
 1024-1039            2  <1% 
 1328-1343            1  <1% 
 2128-2143            1  <1% 
 3312-3327            1  <1% 
 7952-7967            1  <1% 
 8240-8255            1  <1%

Though I must admit that it doesn't always work -- LD_PRELOAD=libmemusage.so ls never reports anything, for example -- and I wish I knew the conditions that allow it to work or not work.

sarnold
  • 102,305
  • 22
  • 181
  • 238