Python: Reducing memory usage of dictionary

I'm trying to load a couple of files into the memory. The files have either of the following 3 formats:

  • string TAB int
  • string TAB float
  • int TAB float.

Indeed, they are ngram statics files, in case this helps with the solution. For instance:

i_love TAB 10
love_you TAB 12

Currently, the pseudocode of I'm doing right now is

     data = {}
     for line in file:
        first, second = line.split('\t')
        data[first] = int(second) #or float(second)

     return data

To much of my surprise, while the total size of the files in disk is about 21 mb, when loaded into memory the process takes 120 - 180 mb of memory! (the whole python application doesn't load any other data into memory).

There are less than 10 files, most of them would stay stable at about 50-80k lines, except for one file which currently has millions of lines.

So I would like to ask for a technique/data structure to reduce the memory consumption:

  • Any advice for compression techniques?
  • If I still use dict, is there any way to reduce the memory? Is it possible to set the "load factor" as in Java for Python dict?
  • If you have some other data structures, 'm also willing to trade some of the speed to reduce the memory. Nevertheless, this is a time sensitive application so that once the users input their queries, I think it'd be not quite reasonable to take more than a few seconds to return the result. With regard to this, I'm still amazed by how Google manage to do the Google Translate so fast: they must be using a lot of techniques + lots of servers' power?

Thank you very much. I look forward to your advice.


I cannot offer a complete strategy that would help improve memory footprint, but I believe it may help to analyse what exactly is taking so much memory.

If you look at the Python implementation of dictionary (which is a relatively straight-forward implementation of a hash table), as well as the implementation of the built-in string and integer data types, for example here (specifically object.h, intobject.h, stringobject.h and dictobject.h, as well as the corresponding *.c files in ../Objects), you can calculate with some accuracy the expected space requirements:

  1. An integer is a fixed-sized object, i.e. it contains a reference count, a type pointer and the actual integer, in total typically at least 12 bytes on a 32bit system and 24 bytes on a 64bit system, not taking into account extra space possibly lost through alignment.

  2. A string object is variable-sized, which means it contains

    • reference count
    • type pointer
    • size information
    • space for the lazily calculated hash code
    • state information (e.g. used for interned strings)
    • a pointer to the dynamic content

    in total at least 24 bytes on 32bit or 60 bytes on 64bit, not including space for the string itself.

  3. The dictionary itself consists of a number of buckets, each containing

    • the hash code of the object currently stored (that is not predictable from the position of the bucket due to the collision resolution strategy used)
    • a pointer to the key object
    • a pointer to the value object

    in total at least 12 bytes on 32bit and 24 bytes on 64bit.

  4. The dictionary starts out with 8 empty buckets and is resized by doubling the number of entries whenever its capacity is reached.

I carried out a test with a list of 46,461 unique strings (337,670 bytes concatenated string size), each associated with an integer — similar to your setup, on a 32-bit machine. According to the calculation above, I would expect a minimum memory footprint of

  • 46,461 * (24+12) bytes = 1.6 MB for the string/integer combinations
  • 337,670 = 0.3 MB for the string contents
  • 65,536 * 12 bytes = 1.6 MB for the hash buckets (after resizing 13 times)

in total 2.65 MB. (For a 64-bit system the corresponding calculation yields 5.5 MB.)

When running the Python interpreter idle, its footprint according to the ps-tool is 4.6 MB. So the total expected memory consumption after creating the dictionary is approximately 4.6 + 2.65 = 7.25 MB. The true memory footprint (according to ps) in my test was 7.6 MB. I guess the extra ca. 0.35 MB were consumed by overhead generated through Python's memory allocation strategy (for memory arenas etc.)

Of course many people will now point out that my use of ps to measure the memory footprint is inaccurate and my assumptions about the size of pointer types and integers on 32-bit and 64-bit systems may be wrong on many specific systems. Granted.

But, nevertheless, the key conclusions, I believe, are these:

  • The Python dictionary implementation consumes a surprisingly small amount of memory
  • But the space taken by the many int and (in particular) string objects, for reference counts, pre-calculated hash codes etc., is more than you'd think at first
  • There is hardly a way to avoid the memory overhead, as long as you use Python and want the strings and integers represented as individual objects — at least I don't see how that could be done
  • It may be worthwhile to look for (or implement yourself) a Python-C extension that implements a hash that stores keys and values as C-pointers (rather than Python objects). I don't know if that exists; but I believe it could be done and could reduce the memory footprint by more than half.

1) SQLite in memory sounds like a great solution, it'll let you query your data more easily once it's loaded which is a pleasure


2) you probably want a named tuple - I'm pretty sure they're lighter than dictionaries and you can access properties using dot notation (for which I have an aesthetic preference anyway).

3) you may want to have a look at Redis:

It is FAST and will let you persist things easily, meaning you don't have to load the whole set in every time you want to use it.

4) a trie sounds like a good idea, but adds some theoretical complexity on your end of the work. You can use Redis to implement and store it though, which will boost your speed even further.

But overall, named tuples are probably the trick here.

In disk you have just strings, when loading to Python the interpreter has to create a whole structure for each string and for each dictionary, besides the string itself.

There's no way to reduce the memory used by the dicts, but there are other ways to approach the problem. If you're willing to trade some speed for memory, you should consider storing and querying the strings from an SQLite file instead of loading everything to dictionaries in memory.

Sounds like a Trie ( data structure might better suit your desire for memory efficiency.

Update: the memory efficiency of python dict has been raised as an issue, though it was rejected from the standard lib given the availability of third party libraries. See:

You can replace dict with blist.sorteddict for log(n) access without the memory overhead. It is convenient because it behaves exactly like a dictionary, ie it implements all its methods, so you only have to change the initial type.

If you are trying to store numeric data compactly in python in memory, your best solution is probably Numpy.

Numpy ( allocates data structures using native C structures. Most of its data structures presume you're storing a single datatype, so this isn't for all situations (you might need to store null, etc.), but it can be very, very, very fast and is about as compact as you could ask for. Lots of science gets done with it (see also: SciPy).

Of course, there is another option: zlib, if you have:

  • Ample CPU cycles, and
  • LOTS of data that won't fit into memory

You could just declare a 'page' of data (however big you want) as an array, read in the data, store it in the array, zip it, then read in some more data, until you have all the data in memory you want.

Then, iterate over the pages, unzip, convert back to an array, and do your operations as needed. For instance:

def arrayToBlob(self, inArray):
    a = array.array('f', inArray)
    return a.tostring()

def blobToArray(self, blob, suppressWarning=False):
        out = array.array('f', [])
    except Exception, e:
        if not suppressWarning:
            msg = "Exception: blob2array, err: %s, in: %s" % (e, blob)
            raise Exception, msg
    return out

Once you have data as a blob, you can pass this blob to zlib and compress the data. If you have lots of repeated values, this blob can be vastly compressed.

Of course, it's slower than keeping it all uncompressed, but if you can't fit it in memory your choices are limited to begin with.

Even with compression, it might not all fit in memory, at which point you might have to write out the compressed pages as strings or pickles, etc.

Good luck!

Need Your Help

Use a web mobile framework?

javascript cordova jquery-mobile sencha-touch titanium

I'm currently on a new projet to realize an application for mobile. The client isn't decided and I've to suggest several solutions.

Morphia mongoDB wildcard query

regex mongodb wildcard morphia mongodb-java

Straight forward question, does anybody know how to make a wildcard query using morphia linked to a mongoDB database?