[Python-Dev] Draft of dictnotes.txt [Please Comment]

Raymond Hettinger Raymond Hettinger" <python@rcn.com
Fri, 2 May 2003 01:24:05 -0400


NOTES ON OPTIMIZING DICTIONARIES
================================


Principal Use Cases for Dictionaries
------------------------------------

Passing keyword arguments
    Typically, one read and one write for 1 to 3 elements.
    Occurs frequently in normal python code.

Class method lookup
    Dictionaries vary in size with 8 to 16 elements being common.
    Usually written once with many lookups.
    When base classes are used, there are many failed lookups
        followed by a lookup in a base class.

Instance attribute lookup and Global variables
    Dictionaries vary in size.  4 to 10 elements are common.
    Both reads and writes are common.

Builtins
    Frequent reads.  Almost never written.
    Size 126 interned strings (as of Py2.3b1).
    A few keys are accessed much more frequently than others.

Uniquification
    Dictionaries of any size.  Bulk of work is in creation.
    Repeated writes to a smaller set of keys.
    Single read of each key.

    * Removing duplicates from a sequence.
        dict.fromkeys(seqn).keys()
    * Counting elements in a sequence.
        for e in seqn:  d[e]=d.get(e,0) + 1
    * Accumulating items in a dictionary of lists.
        for k, v in itemseqn:  d.setdefault(k, []).append(v)

Membership Testing
    Dictionaries of any size.  Created once and then rarely changes.
    Single write to each key.
    Many calls to __contains__() or has_key().
    Similar access patterns occur with replacement dictionaries
        such as with the % formatting operator.


Data Layout (assuming a 32-bit box with 64 bytes per cache line)
-----------------------------------------------------------

Smalldicts (8 entries) are attached to the dictobject structure
and the whole group nearly fills two consecutive cache lines.

Larger dicts use the first half of the dictobject structure (one cache
line) and a separate, continuous block of entries (at 12 bytes each
for a total of 5.333 entries per cache line).


Tunable Dictionary Parameters
-----------------------------

* PyDict_MINSIZE.  Currently set to 8.
    Must be a power of two.  New dicts have to zero-out every cell.
    Each additional 8 consumes 1.5 cache lines.  Increasing improves
    the sparseness of small dictionaries but costs time to read in
    the additional cache lines if they are not already in cache.
    That case is common when keyword arguments are passed.

* Maximum dictionary load in PyDict_SetItem.  Currently set to 2/3.
    Increasing this ratio makes dictionaries more dense resulting
    in more collisions.  Decreasing it, improves sparseness at the
    expense of spreading entries over more cache lines and at the
    cost of total memory consumed.

    The load test occurs in highly time sensitive code.  Efforts
    to make the test more complex (for example, varying the load
    for different sizes) have degraded performance.

* Growth rate upon hitting maximum load.  Currently set to *2.
    Raising this to *4 results in half the number of resizes,
    less effort to resize, better sparseness for some (but not
    all dict sizes), and potentially double memory consumption
    depending on the size of the dictionary.  Setting to *4
    eliminated every other resize step.

Tune-ups should be measured across a broad range of applications and
use cases.  A change to any parameter will help in some situations and
hurt in others.  The key is to find settings that help the most common
cases and do the least damage to the less common cases.  Results will
vary dramatically depending on the exact number of keys, whether the
keys are all strings, whether reads or writes dominate, the exact
hash values of the keys (some sets of values have fewer collisions than
others).  Any one test or benchmark is likely to prove misleading.


Results of Cache Locality Experiments
-------------------------------------

When a entry is retrieved from memory, 4.333 adjacent entries are also
retrieved into a cache line.  Since accessing items in cache is *much*
cheaper than a cache miss, an enticing idea is to probe the adjacent
entries as a first step in collision resolution.  Unfortunately, the
introduction of any regularity into collision searches results in more
collisions than the current random chaining approach.

Exploiting cache locality at the expense of additional collisions fails
to payoff when the entries are already loaded in cache (the expense
is paid with no compensating benefit).  This occurs in small dictionaries
where the whole dictionary fits into a pair of cache lines.  It also
occurs frequently in large dictionaries which have a common access pattern
where some keys are accessed much more frequently than others.  The
more popular entries *and* their collision chains tend to remain in cache.

To exploit cache locality, change the collision resolution section
in lookdict() and lookdict_string().  Set i^=1 at the top of the
loop and move the  i = (i << 2) + i + perturb + 1 to an unrolled
version of the loop.

This optimization strategy can be leveraged in several ways:

* If the dictionary is kept sparse (through the tunable parameters),
then the occurrence of additional collisions is lessened.

* If lookdict() and lookdict_string() are specialized for small dicts
and for largedicts, then the versions for large_dicts can be given
the alternate search without increasing the collision in small dicts
which already have the maximum benefit of cache locality.

* If the use case for the dictionary is known to have a random key
access pattern (as opposed to a more common pattern with a Zipf's law
distribution), then there will be more benefit for large dictionaries
because any given key is no more likely than another to already be
in cache.


Optimizing the Search of Small Dictionaries
-------------------------------------------

If lookdict() and lookdict_string() are specialized for smaller dictionaries,
then a custom search approach can be implemented that exploits the small
search space and cache locality.

* The simplest example is a linear search of contiguous entries.  This is
  simple to implement, guaranteed to terminate rapidly, and precludes
  the need to check for dummy entries.

* A more advanced example is a self-organizing search so that the most
  frequently accessed entries get probed first.  The organization
  adapts if the access pattern changes over time.

* Also, small dictionaries may be made more dense, perhaps filling all
  eight cells to take the maximum advantage of two cache lines.


Strategy Pattern
----------------

Consider allowing the user to set the tunable parameters or to select a
particular search method.  Since some dictionary use cases have known
sizes and access patterns, the user may be able to provide useful hints.

1) For example, if membership testing or lookup dominates runtime and memory
   is not at a premium, the user may benefit from setting the maximum load
   ratio at 5% or 10% instead of the usual 66.7%.  This will sharply
   curtail the number of collisions.

2) Dictionary creation time can be shortened in cases where the ultimate
   size of the dictionary is known in advance.  The dictionary can be
   pre-sized so that *no* resize operations are required during creation.
   Not only does this save resizes, but the key insertion will go
   more quickly because the first half of the keys will be inserted into
   a more sparse environment than before.  The preconditions for this
   strategy arise whenever a dictionary is created from a key or item
   sequence of known length.

3) If the key space is large and the access pattern is known to be random,
   then search strategies exploiting cache locality can be fruitful.
   The preconditions for this strategy arise in simulations and
   numerical analysis.

4) If the keys are fixed and the access pattern strongly favors some of
   the keys, then the entries can be stored consecutively and accessed
   with a linear search.  This exploits knowledge of the data, cache
   locality, and a simplified search routine.  It also eliminates the
   need to test for dummy entries on each probe.  The preconditions for
   this strategy arise in symbol tables and in the builtin dictionary.