PEP 412 – Key-Sharing Dictionary | peps.python.org (original) (raw)

Author:

Mark Shannon

Status:

Final

Type:

Standards Track

Created:

08-Feb-2012

Python-Version:

3.3

Post-History:

08-Feb-2012

Table of Contents

Abstract

This PEP proposes a change in the implementation of the builtin dictionary type dict. The new implementation allows dictionaries which are used as attribute dictionaries (the __dict__ attribute of an object) to share keys with other attribute dictionaries of instances of the same class.

Motivation

The current dictionary implementation uses more memory than is necessary when used as a container for object attributes as the keys are replicated for each instance rather than being shared across many instances of the same class. Despite this, the current dictionary implementation is finely tuned and performs very well as a general-purpose mapping object.

By separating the keys (and hashes) from the values it is possible to share the keys between multiple dictionaries and improve memory use. By ensuring that keys are separated from the values only when beneficial, it is possible to retain the high-performance of the current dictionary implementation when used as a general-purpose mapping object.

Behaviour

The new dictionary behaves in the same way as the old implementation. It fully conforms to the Python API, the C API and the ABI.

Performance

Memory Usage

Reduction in memory use is directly related to the number of dictionaries with shared keys in existence at any time. These dictionaries are typically half the size of the current dictionary implementation.

Benchmarking shows that memory use is reduced by 10% to 20% for object-oriented programs with no significant change in memory use for other programs.

Speed

The performance of the new implementation is dominated by memory locality effects. When keys are not shared (for example in module dictionaries and dictionary explicitly created by dict() or{}) then performance is unchanged (within a percent or two) from the current implementation.

For the shared keys case, the new implementation tends to separate keys from values, but reduces total memory usage. This will improve performance in many cases as the effects of reduced memory usage outweigh the loss of locality, but some programs may show a small slow down.

Benchmarking shows no significant change of speed for most benchmarks. Object-oriented benchmarks show small speed ups when they create large numbers of objects of the same class (the gcbench benchmark shows a 10% speed up; this is likely to be an upper limit).

Implementation

Both the old and new dictionaries consist of a fixed-sized dict struct and a re-sizeable table. In the new dictionary the table can be further split into a keys table and values array. The keys table holds the keys and hashes and (for non-split tables) the values as well. It differs only from the original implementation in that it contains a number of fields that were previously in the dict struct. If a table is split the values in the keys table are ignored, instead the values are held in a separate array.

Split-Table dictionaries

When dictionaries are created to fill the __dict__ slot of an object, they are created in split form. The keys table is cached in the type, potentially allowing all attribute dictionaries of instances of one class to share keys. In the event of the keys of these dictionaries starting to diverge, individual dictionaries will lazily convert to the combined-table form. This ensures good memory use in the common case, and correctness in all cases.

When resizing a split dictionary it is converted to a combined table. If resizing is as a result of storing an instance attribute, and there is only instance of a class, then the dictionary will be re-split immediately. Since most OO code will set attributes in the __init__ method, all attributes will be set before a second instance is created and no more resizing will be necessary as all further instance dictionaries will have the correct size. For more complex use patterns, it is impossible to know what is the best approach, so the implementation allows extra insertions up to the point of a resize when it reverts to the combined table (non-shared keys).

A deletion from a split dictionary does not change the keys table, it simply removes the value from the values array.

Combined-Table dictionaries

Explicit dictionaries (dict() or {}), module dictionaries and most other dictionaries are created as combined-table dictionaries. A combined-table dictionary never becomes a split-table dictionary. Combined tables are laid out in much the same way as the tables in the old dictionary, resulting in very similar performance.

Implementation

The new dictionary implementation is available at [1].

Pros and Cons

Pros

Significant memory savings for object-oriented applications. Small improvement to speed for programs which create lots of similar objects.

Cons

Change to data structures: Third party modules which meddle with the internals of the dictionary implementation will break.

Changes to repr() output and iteration order: For most cases, this will be unchanged. However, for some split-table dictionaries the iteration order will change.

Neither of these cons should be a problem. Modules which meddle with the internals of the dictionary implementation are already broken and should be fixed to use the API. The iteration order of dictionaries was never defined and has always been arbitrary; it is different for Jython and PyPy.

Alternative Implementation

An alternative implementation for split tables, which could save even more memory, is to store an index in the value field of the keys table (instead of ignoring the value field). This index would explicitly state where in the value array to look. The value array would then only require 1 field for each usable slot in the key table, rather than each slot in the key table.

This “indexed” version would reduce the size of value array by about one third. The keys table would need an extra “values_size” field, increasing the size of combined dicts by one word. The extra indirection adds more complexity to the code, potentially reducing performance a little.

The “indexed” version will not be included in this implementation, but should be considered deferred rather than rejected, pending further experimentation.

References

This document has been placed in the public domain.