1 NOTES ON DICTIONARIES 2 ================================ 3 4 Principal Use Cases for Dictionaries 5 ------------------------------------ 6 7 Passing keyword arguments 8 Typically, one read and one write for 1 to 3 elements. 9 Occurs frequently in normal python code. 10 11 Class method lookup 12 Dictionaries vary in size with 8 to 16 elements being common. 13 Usually written once with many lookups. 14 When base classes are used, there are many failed lookups 15 followed by a lookup in a base class. 16 17 Instance attribute lookup and Global variables 18 Dictionaries vary in size. 4 to 10 elements are common. 19 Both reads and writes are common. 20 21 Builtins 22 Frequent reads. Almost never written. 23 About 150 interned strings (as of Py3.3). 24 A few keys are accessed much more frequently than others. 25 26 Uniquification 27 Dictionaries of any size. Bulk of work is in creation. 28 Repeated writes to a smaller set of keys. 29 Single read of each key. 30 Some use cases have two consecutive accesses to the same key. 31 32 * Removing duplicates from a sequence. 33 dict.fromkeys(seqn).keys() 34 35 * Counting elements in a sequence. 36 for e in seqn: 37 d[e] = d.get(e,0) + 1 38 39 * Accumulating references in a dictionary of lists: 40 41 for pagenumber, page in enumerate(pages): 42 for word in page: 43 d.setdefault(word, []).append(pagenumber) 44 45 Note, the second example is a use case characterized by a get and set 46 to the same key. There are similar use cases with a __contains__ 47 followed by a get, set, or del to the same key. Part of the 48 justification for d.setdefault is combining the two lookups into one. 49 50 Membership Testing 51 Dictionaries of any size. Created once and then rarely changes. 52 Single write to each key. 53 Many calls to __contains__() or has_key(). 54 Similar access patterns occur with replacement dictionaries 55 such as with the % formatting operator. 56 57 Dynamic Mappings 58 Characterized by deletions interspersed with adds and replacements. 59 Performance benefits greatly from the re-use of dummy entries. 60 61 Data Layout 62 ----------- 63 64 Dictionaries are composed of 3 components: 65 The dictobject struct itself 66 A dict-keys object (keys & hashes) 67 A values array 68 69 70 Tunable Dictionary Parameters 71 ----------------------------- 72 73 See comments for PyDict_MINSIZE_SPLIT, PyDict_MINSIZE_COMBINED, 74 USABLE_FRACTION and GROWTH_RATE in dictobject.c 75 76 Tune-ups should be measured across a broad range of applications and 77 use cases. A change to any parameter will help in some situations and 78 hurt in others. The key is to find settings that help the most common 79 cases and do the least damage to the less common cases. Results will 80 vary dramatically depending on the exact number of keys, whether the 81 keys are all strings, whether reads or writes dominate, the exact 82 hash values of the keys (some sets of values have fewer collisions than 83 others). Any one test or benchmark is likely to prove misleading. 84 85 While making a dictionary more sparse reduces collisions, it impairs 86 iteration and key listing. Those methods loop over every potential 87 entry. Doubling the size of dictionary results in twice as many 88 non-overlapping memory accesses for keys(), items(), values(), 89 __iter__(), iterkeys(), iteritems(), itervalues(), and update(). 90 Also, every dictionary iterates at least twice, once for the memset() 91 when it is created and once by dealloc(). 92 93 Dictionary operations involving only a single key can be O(1) unless 94 resizing is possible. By checking for a resize only when the 95 dictionary can grow (and may *require* resizing), other operations 96 remain O(1), and the odds of resize thrashing or memory fragmentation 97 are reduced. In particular, an algorithm that empties a dictionary 98 by repeatedly invoking .pop will see no resizing, which might 99 not be necessary at all because the dictionary is eventually 100 discarded entirely. 101 102 The key differences between this implementation and earlier versions are: 103 1. The table can be split into two parts, the keys and the values. 104 105 2. There is an additional key-value combination: (key, NULL). 106 Unlike (<dummy>, NULL) which represents a deleted value, (key, NULL) 107 represented a yet to be inserted value. This combination can only occur 108 when the table is split. 109 110 3. No small table embedded in the dict, 111 as this would make sharing of key-tables impossible. 112 113 114 These changes have the following consequences. 115 1. General dictionaries are slightly larger. 116 117 2. All object dictionaries of a single class can share a single key-table, 118 saving about 60% memory for such cases. 119 120 Results of Cache Locality Experiments 121 -------------------------------------- 122 123 Experiments on an earlier design of dictionary, in which all tables were 124 combined, showed the following: 125 126 When an entry is retrieved from memory, several adjacent entries are also 127 retrieved into a cache line. Since accessing items in cache is *much* 128 cheaper than a cache miss, an enticing idea is to probe the adjacent 129 entries as a first step in collision resolution. Unfortunately, the 130 introduction of any regularity into collision searches results in more 131 collisions than the current random chaining approach. 132 133 Exploiting cache locality at the expense of additional collisions fails 134 to payoff when the entries are already loaded in cache (the expense 135 is paid with no compensating benefit). This occurs in small dictionaries 136 where the whole dictionary fits into a pair of cache lines. It also 137 occurs frequently in large dictionaries which have a common access pattern 138 where some keys are accessed much more frequently than others. The 139 more popular entries *and* their collision chains tend to remain in cache. 140 141 To exploit cache locality, change the collision resolution section 142 in lookdict() and lookdict_string(). Set i^=1 at the top of the 143 loop and move the i = (i << 2) + i + perturb + 1 to an unrolled 144 version of the loop. 145 146 For split tables, the above will apply to the keys, but the value will 147 always be in a different cache line from the key. 148 149 150
