20

I have a numpy script that -- according to top -- is using about 5GB of RAM:

  PID USER   PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
16994 aix    25   0 5813m 5.2g 5.1g S  0.0 22.1  52:19.66 ipython

Is there a memory profiler that would enable me to get some idea about the objects that are taking most of that memory?

I've tried heapy, but guppy.hpy().heap() is giving me this:

Partition of a set of 90956 objects. Total size = 12511160 bytes.
 Index  Count   %     Size   % Cumulative  % Kind (class / dict of class)
     0  42464  47  4853112  39   4853112  39 str
     1  22147  24  1928768  15   6781880  54 tuple
     2    287   0  1093352   9   7875232  63 dict of module
     3   5734   6   733952   6   8609184  69 types.CodeType
     4    498   1   713904   6   9323088  75 dict (no owner)
     5   5431   6   651720   5   9974808  80 function
     6    489   1   512856   4  10487664  84 dict of type
     7    489   1   437704   3  10925368  87 type
     8    261   0   281208   2  11206576  90 dict of class
     9   1629   2   130320   1  11336896  91 __builtin__.wrapper_descriptor
<285 more rows. Type e.g. '_.more' to view.>

For some reason, it's only accounting for 12MB of the 5GB (the bulk of the memory is almost certainly used by numpy arrays).

Any suggestions as to what I might be doing wrong with heapy or what other tools I should try (other than those already mentioned in this thread)?

Community
  • 1
  • 1
NPE
  • 486,780
  • 108
  • 951
  • 1,012

1 Answers1

11

Numpy (and its library bindings, more on that in a minute) use C malloc to allocate space, which is why memory used by big numpy allocations doesn't show up in the profiling of things like heapy and never gets cleaned up by the garbage collector.

The usual suspects for big leaks are actually scipy or numpy library bindings, rather than python code itself. I got burned badly last year by the default scipy.linalg interface to umfpack, which leaked memory at the rate of about 10Mb a call. You might want to try something like valgrind to profile the code. It can often give some hints as to where to look at where there might be leaks.

talonmies
  • 70,661
  • 34
  • 192
  • 269
  • Yeah but there are still instances where it'd be nice to know the size of numpy/scipy arrays, for instance, in debugging memory leaks in your own 150,000 application where you're passing around and copying arrays left and right and you know the problem is in your own code. It'd be nice if pympler, etc. could just use the `array.nbytes` attribute to give meaningful results. – partofthething Feb 05 '16 at 23:34