Skip to content
main
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 

Microdict

A high performance python hash table library that is generally faster and consumes significantly less memory than Python Dictionaries. It currently supports Python 3.5+.


Why Microdict? Why create another Hash table library when there is the builtin Python Dictionary?

Python Dictionaries are fast but their memory consumption can also be high at the same time. This is partly due to the nature of Python keeping data, within RAM, in the form of PyObjects, which consume far more memory than native types such as Integers and Character Arrays. As a result, Python Dictionaries can be prohibitive in many cases while building memory intensive python applications. This motivated me to develop a typed python hash table library that consumes significantly (upto 7 times) less memory compared to python dictionaries. It is also faster than python dictionaries. Moreover, it's underlying C implementation can also outperform Google's highly optimized Swiss Table and Facebook's F14 hash tables. See the Performance Section.


Installation and Building

You can install Microdict using pip : pip install microdict.

Microdict is absolutely built using C extensions and as such building it will require Python C api header files. Build and install the package using python setup.py install from the terminal after cloning the repository. Microdict is tested to work on Linux, Mac OSX, and Windows systems. You will need GCC 7+ on linux/mac osx systems and Visual C++ 14+ compiler on Windows systems to build the package. For the best performance use on a 64 bit system.


Run tests

Once installed, type the following code snippet in your python interpreter to run the tests:

from microdict import run_tests
run_tests.run()

Usage

The following code snippet shows common uses of the library.

from microdict import mdict

dict_i32 = mdict.create("i32:i32") # Generates a dictionary with key and value type of signed 32 bit integer.
dict_i32[1] = 2 # Just like python dictionaries, setting a key value pair.
print(4 in dict_i32) # prints False.

print(dict_i32[1]) # prints 2
try:
   print(dict_i32.pop(4)) # prints None.
except KeyError:
   pass
print(dict_i32.pop(1)) # Removes [1,2] from the hashtable and prints 2.

dict_i32[10] = 0
dict_i32[5] = 1
dict_i32[6] = 8

for k in dict_i32:
   print(k) # Will print 10, 5, 6
  
for v in dict_i32.values():
   print(v) # Will print 0, 1, 8
  
for k,v in dict_i32.items():
   print(k,v) # Will print all the items.

d2 = dict_i32.copy() # Creates a deep copy of dict_i32.
dict_i32.clear([10,6]) # Removes the pairs [10,0] and [6,8]
dict_i32.clear() # Makes the dictionary empty.

pydict = d2.to_Pydict() # Returns a python dictionary containing all the items in d2
pydict[120] = 5
pydict[42] = 9

d2.update(pydict) # d2 now will additionally have the pairs [120, 5] and [42, 9]

dict_i32[111] = 89
dict_i32[123] = 1
d2.update(dict_i32) # d2 now will additionally have the pairs [111, 89] and [123, 1].

print(list(d2.items())) # prints all d2 items
"""
Faster approach shown below. d2.get_items() creates and returns a list of all items. 
So if you don't need a list container, iterate using d2.items() for memory efficiency. 
Same applies for other d2.get_* methods shown below.
"""
print(d2.get_items()) # 
print(list(d2.values())) # prints all d2 values
print(d2.get_values()) # Same but faster approach
print(list(d2)) # prints all d2 keys
print(d2.get_keys()) # Same but faster approach

Hash Table types

Currently, Microdict includes 5 types of dictionaries:

  • "i32:i32" -> 32 bit signed keys and 32 bit signed values
  • "i32:i64" -> 32 bit signed keys and 64 bit signed values
  • "i64:i32" -> 64 bit signed keys and 32 bit signed values
  • "i64:i64" -> 64 bit signed keys and 64 bit signed values
  • "str:str" -> string keys and string values.

Method Documentations

  • microdict.mdict.create (dtype, key_len=None, val_len=None)

    : Returns a Microdict hash table of any of the types given above.

    Parameters:

    • dtype: A python string type (str) that sets the hash table type to be created. It can be any one of the above types.
    • key_len: A python Integer type (int). It sets the maximum number of bytes the characters of a key (UTF-8 string) requires. Passing a UTF-8 encoded string key which consumes more bytes than key_len will not be accepted. This argument is only applicable when dtype="str:str". It only accepts a value of at most 65355 and a larger value will raise a TypeError.
    • val_len: A python Integer type(int). It sets the maximum number of bytes the characters of a value (UTF-8 string) requires. Passing a UTF-8 encoded string value which consumes more bytes than val_len will not be accepted. This argument is only applicable when dtype="str:str". It only accepts a value of at most 65355 and a larger value will raise a TypeError.
  • microdict.mdict.listDictionaryTypes ()

    : Prints a series of lines of the form : Key Type: key_t . Value Type: val_t, where key_t:val_t forms a type given above.

The following are the methods that are common to all hash table types returned by mdict.create.

  • clear (key_list = None)

    : Returns None. If key_list is not provided, the clear method will delete all items from the hash table.

    Parameters:

    • key_list: It is an optional argument but not a keyword optional argument (keyword must not be provided). So, it suffices: keys = [1,2]; d.clear(keys). If provided, it must be of type list. The entries within key_list are the keys that will be removed from the hash table. By default, any entry that is not present in the hash table will be skipped. For any of the integer hash table types, any non python int entry will be skipped. It is upto the programmer to pass the proper sized integer to prevent overflows. For str:str type, the entries must be python UTF-8 strings with UTF-8 character bytes upto key_len as set by mdict.create and otherwise, that entry will be skipped.
  • copy ()

    : Returns a deep copy of the Microdict Hash table of the same type as the caller Hash table object.

  • get_items ()

    : Creates and returns a python list containing all the items (key, value) in the hash table.

  • get_keys ()

    : Creates and returns a python list containing all the keys in the hash table.

  • get_values ()

    : Creates and returns a python list containing all the values in the hash table.

  • items ()

    : Used to iterate over items using a for loop. Example : for k,v in d.items() : print(k, v)

  • pop (key)

    : Deletes a key from the hash table and returns its corresponding value. If the key is not present, then a KeyError is raised.

    Parameters:

    • key: For any of the integer hash table types, any non python int entry will raise a TypeError. It is upto the programmer to pass the proper sized integer to prevent overflows. For str:str type, the entries must be python UTF-8 strings with UTF-8 character bytes upto key_len as set by mdict.create and otherwise, a TypeError will be raised.
  • to_Pydict ()

    : Creates and returns a python dictionary containing all items present in the Microdict hash table.

  • update (dictionary)

    : Inserts all the items present in the dictionary into the Microdict hash table.

    Parameters:

    • dictionary: Must be either a python dictionary or a Microdict hash table. If it is a python dictionary, then all its items that are of the same type as the Microdict hash table will be inserted. The rest will be skipped by default. The conditions given in the method documentation of clear regarding type constraints apply here too.
  • values ()

    : Used to iterate over values using a for loop. Example : for v in d.values() : print(v).


Performance

Competing with Python Dictionary

Each of the cells in the tables below are of the format (Speed Gain, Space Gain).

  • Speed Gain is defined as : .

  • Similarly, Space Gain :

Experiments were carried out for the types "i32:i32", "i64:64", "str:str" (key and value length kept to 8 characters). Speed Gain and Space Gain are computed by averaging over the results of 30 experiments carried out using (unique) randomly generated data. Space consumption was computed using the psutil library. Time consumption was computed using the time.perf_counter method. All the experiments were conducted with Python 3.8.2 on a 64 bit Ubuntu 20.04 LTS system. The following table shows the benchmarks against Python Dictionary for retrieving all values using the full set of keys and the [] operator.

#Items i32:i32 i64:i64 str:str
100000 1.48x, 7.13x 1.29x, 4.67x 1.43x, 5.19x
1000000 1.47x, 4.23x 1.48x, 2.64x 1.46x, 4.46x
10000000 1.44x, 4.77x 1.48x, 3.02x 1.53x, 4.97x
3x10000000 1.55x, 4.19x 1.3x , 2.57x 1.36x, 3.93x

Competing with Google's Swiss Table and Facebook's F14

Microdict's underlying C implementation was benchmarked against Swiss Table->abseil::flat_hash_map and F14->folly::F14FastMap to further test its capabilities. The data types were set as above using the same settings. The following tables show the results for retrieving all values using the keys.

abseil::flat_hash_map folly::F14FastMap
#Items i32:i32 i64:i64 str:str
1000000 1.22x, 1.56x 1.02x, 1.08x 1.33x, 1.82x
10000000 1.14x, 1x 1.09x, 1.33x 1.73x, 1.67x
3x10000000 1.11x, 1.27x 1.15x, 1.03x 1.68x, 1.65x
#Items i32:i32 i64:i64 str:str
1000000 2.96x, 1.56x 2.74x, 1x 3.12x, 1.39x
10000000 2.11x, 1x 2.12x, 1.29x 3.45x, 1.4x
3x10000000 1.84x, 1.22x 1.97x, 1x 3.36x, 1.21x

About

A high performance python hash table library that is generally faster and consumes significantly less memory than Python Dictionaries

Resources

License

Packages

No packages published