Profile Memory Allocation in Python (with support for Numpy arrays)

I have a program that contains a large number of objects, many of them Numpy arrays. My program is swapping miserably, and I'm trying to reduce the memory usage, because it actually can't finis on my system with the current memory requirements.

I am looking for a nice prof开发者_如何学运维iler that would allow me to check the amount of memory consumed by various objects (I'm envisioning a memory counterpart to cProfile) so that I know where to optimize.

I've heard decent things about Heapy, but Heapy unfortunately does not support Numpy arrays, and most of my program involves Numpy arrays.

One way to tackle the problem if you are calling lots of different functions and you are unsure where the swapping comes from would be to use the new plotting functionality from memory_profiler. First you must decorate the different functions you are using with @profile. For simplicity I'll use the example examples/numpy_example.py shipped with memory_profiler that contains two functions: create_data() and process_data()

To run your script, instead of running it with the Python interpreter, you use the mprof executable, that is

$ mprof run examples/numpy_example.py

This will create a file called mprofile_??????????.dat, where the ? will hold numbers representing the current date. To plot the result, simply type mprof plot and it will generate a plot similar to this (if you have several .dat files it will always take the last one):

Here you see the memory consumption, with brackets indicating when you enter/leave the current function. This way it is easy to see that function process_data() has a peak of memory consumption. To further dig into your function, you could use the line-by-line profiler to see the memory consumption of each line in your function. This is run with

python -m memory_profiler examples/nump_example.py

This would give you an output similar to this:

Line #    Mem usage    Increment   Line Contents
================================================
    13                             @profile
    14  223.414 MiB    0.000 MiB   def process_data(data):
    15  414.531 MiB  191.117 MiB       data = np.concatenate(data)
    16  614.621 MiB  200.090 MiB       detrended = scipy.signal.detrend(data, axis=0)
    17  614.621 MiB    0.000 MiB       return detrended

where it is clear that scipy.signal.detrend is allocating a huge amount of memory.

Have a look at memory profiler. It provides line by line profiling and Ipython integration, which makes it very easy to use it:

In [1]: import numpy as np

In [2]: %memit np.zeros(1e7)
maximum of 3: 70.847656 MB per loop

Update

As mentioned by @WickedGrey there seems to be a bug (see github issue tracker) when calling a function more than one time, which I can reproduce:

In [2]: for i in range(10):
   ...:     %memit np.zeros(1e7)
   ...:     
maximum of 1: 70.894531 MB per loop
maximum of 1: 70.894531 MB per loop
maximum of 1: 70.894531 MB per loop
maximum of 1: 70.894531 MB per loop
maximum of 1: 70.894531 MB per loop
maximum of 1: 70.894531 MB per loop
maximum of 1: 70.902344 MB per loop
maximum of 1: 70.902344 MB per loop
maximum of 1: 70.902344 MB per loop
maximum of 1: 70.902344 MB per loop

However I don't know to what extend the results maybe influenced (seems to be not that much in my example, so depending on your use case it maybe still useful) and when this issue maybe fixed. I asked that at github.

Since numpy 1.7 there exists a semi built-in way to track memory allocations:

https://github.com/numpy/numpy/tree/master/tools/allocation_tracking

Can you just save/pickle some of the arrays to disk in tmp files when not using them? That's what I've had to do in the past with large arrays. Of course this will slow the program down, but at least it'll finish. Unless you need them all at once?

Have you tried valgrind with the massif tool?

valgrind --tool=massif python yourscript.py

it will create a file called massif.out.xxx which you can inspect via

ms_print massif.out.xxx | less

it has all kinds of useful information, but the plot right in the beginning should be what you're looking for. Also check out massif tutorial on the valgrind homepage.

Using valgrind is quite advanced and there might be easier ways to do what you're looking for.