PEP: 3128 Title: BList: A Faster List-like Type Version: $Revision$ Last-Modified: $Date$ Author: Daniel Stutzbach Discussions-To: Python 3000 List Status: Draft Type: Standards Track Content-Type: text/x-rst Created: 30-Apr-2007 Python-Version: 2.6 and/or 3.0 Post-History: 30-Apr-2007 Abstract ======== The common case for list operations is on small lists. The current array-based list implementation excels at small lists due to the strong locality of reference and infrequency of memory allocation operations. However, an array takes O(n) time to insert and delete elements, which can become problematic as the list gets large. This PEP introduces a new data type, the BList, that has array-like and tree-like aspects. It enjoys the same good performance on small lists as the existing array-based implementation, but offers superior asymptotic performance for most operations. This PEP proposes replacing the makes two mutually exclusive proposals for including the BList type in Python: 1. Add it to the collections module, or 2. Replace the existing list type Motivation ========== The BList grew out of the frustration of needing to rewrite intuitive algorithms that worked fine for small inputs but took O(n**2) time for large inputs due to the underlying O(n) behavior of array-based lists. The deque type, introduced in Python 2.4, solved the most common problem of needing a fast FIFO queue. However, the deque type doesn't help if we need to repeatedly insert or delete elements from the middle of a long list. A wide variety of data structure provide good asymptotic performance for insertions and deletions, but they either have O(n) performance for other operations (e.g., linked lists) or have inferior performance for small lists (e.g., binary trees and skip lists). The BList type proposed in this PEP is based on the principles of B+Trees, which have array-like and tree-like aspects. The BList offers array-like performance on small lists, while offering O(log n) asymptotic performance for all insert and delete operations. Additionally, the BList implements copy-on-write under-the-hood, so even operations like getslice take O(log n) time. The table below compares the asymptotic performance of the current array-based list implementation with the asymptotic performance of the BList. ========= ================ ==================== Operation Array-based list BList ========= ================ ==================== Copy O(n) **O(1)** Append **O(1)** O(log n) Insert O(n) **O(log n)** Get Item **O(1)** O(log n) Set Item **O(1)** **O(log n)** Del Item O(n) **O(log n)** Iteration O(n) O(n) Get Slice O(k) **O(log n)** Del Slice O(n) **O(log n)** Set Slice O(n+k) **O(log k + log n)** Extend O(k) **O(log k + log n)** Sort O(n log n) O(n log n) Multiply O(nk) **O(log k)** ========= ================ ==================== An extensive empirical comparison of Python's array-based list and the BList are available at [2]_. Use Case Trade-offs =================== The BList offers superior performance for many, but not all, operations. Choosing the correct data type for a particular use case depends on which operations are used. Choosing the correct data type as a built-in depends on balancing the importance of different use cases and the magnitude of the performance differences. For the common uses cases of small lists, the array-based list and the BList have similar performance characteristics. For the slightly less common case of large lists, there are two common uses cases where the existing array-based list outperforms the existing BList reference implementation. These are: 1. A large LIFO stack, where there are many .append() and .pop(-1) operations. Each operation is O(1) for an array-based list, but O(log n) for the BList. 2. A large list that does not change size. The getitem and setitem calls are O(1) for an array-based list, but O(log n) for the BList. In performance tests on a 10,000 element list, BLists exhibited a 50% and 5% increase in execution time for these two uses cases, respectively. The performance for the LIFO use case could be improved to O(n) time, by caching a pointer to the right-most leaf within the root node. For lists that do not change size, the common case of sequential access could also be improved to O(n) time via caching in the root node. However, the performance of these approaches has not been empirically tested. Many operations exhibit a tremendous speed-up (O(n) to O(log n)) when switching from the array-based list to BLists. In performance tests on a 10,000 element list, operations such as getslice, setslice, and FIFO-style insert and deletes on a BList take only 1% of the time needed on array-based lists. In light of the large performance speed-ups for many operations, the small performance costs for some operations will be worthwhile for many (but not all) applications. Implementation ============== The BList is based on the B+Tree data structure. The BList is a wide, bushy tree where each node contains an array of up to 128 pointers to its children. If the node is a leaf, its children are the user-visible objects that the user has placed in the list. If node is not a leaf, its children are other BList nodes that are not user-visible. If the list contains only a few elements, they will all be a children of single node that is both the root and a leaf. Since a node is little more than array of pointers, small lists operate in effectively the same way as an array-based data type and share the same good performance characteristics. The BList maintains a few invariants to ensure good (O(log n)) asymptotic performance regardless of the sequence of insert and delete operations. The principle invariants are as follows: 1. Each node has at most 128 children. 2. Each non-root node has at least 64 children. 3. The root node has at least 2 children, unless the list contains fewer than 2 elements. 4. The tree is of uniform depth. If an insert would cause a node to exceed 128 children, the node spawns a sibling and transfers half of its children to the sibling. The sibling is inserted into the node's parent. If the node is the root node (and thus has no parent), a new parent is created and the depth of the tree increases by one. If a deletion would cause a node to have fewer than 64 children, the node moves elements from one of its siblings if possible. If both of its siblings also only have 64 children, then two of the nodes merge and the empty one is removed from its parent. If the root node is reduced to only one child, its single child becomes the new root (i.e., the depth of the tree is reduced by one). In addition to tree-like asymptotic performance and array-like performance on small-lists, BLists support transparent **copy-on-write**. If a non-root node needs to be copied (as part of a getslice, copy, setslice, etc.), the node is shared between multiple parents instead of being copied. If it needs to be modified later, it will be copied at that time. This is completely behind-the-scenes; from the user's point of view, the BList works just like a regular Python list. Memory Usage ============ In the worst case, the leaf nodes of a BList have only 64 children each, rather than a full 128, meaning that memory usage is around twice that of a best-case array implementation. Non-leaf nodes use up a negligible amount of additional memory, since there are at least 63 times as many leaf nodes as non-leaf nodes. The existing array-based list implementation must grow and shrink as items are added and removed. To be efficient, it grows and shrinks only when the list has grow or shrunk exponentially. In the worst case, it, too, uses twice as much memory as the best case. In summary, the BList's memory footprint is not significantly different from the existing array-based implementation. Backwards Compatibility ======================= If the BList is added to the collections module, backwards compatibility is not an issue. This section focuses on the option of replacing the existing array-based list with the BList. For users of the Python interpreter, a BList has an identical interface to the current list-implementation. For virtually all operations, the behavior is identical, aside from execution speed. For the C API, BList has a different interface than the existing list-implementation. Due to its more complex structure, the BList does not lend itself well to poking and prodding by external sources. Thankfully, the existing list-implementation defines an API of functions and macros for accessing data from list objects. Google Code Search suggests that the majority of third-party modules uses the well-defined API rather than relying on the list's structure directly. The table below summarizes the search queries and results: ======================== ================= Search String Number of Results ======================== ================= PyList_GetItem 2,000 PySequence_GetItem 800 PySequence_Fast_GET_ITEM 100 PyList_GET_ITEM 400 \[^a\-zA\-Z\_\]ob_item 100 ======================== ================= This can be achieved in one of two ways: 1. Redefine the various accessor functions and macros in listobject.h to access a BList instead. The interface would be unchanged. The functions can easily be redefined. The macros need a bit more care and would have to resort to function calls for large lists. The macros would need to evaluate their arguments more than once, which could be a problem if the arguments have side effects. A Google Code Search for "PyList_GET_ITEM\(\[^)\]+\(" found only a handful of cases where this occurs, so the impact appears to be low. The few extension modules that use list's undocumented structure directly, instead of using the API, would break. The core code itself uses the accessor macros fairly consistently and should be easy to port. 2. Deprecate the existing list type, but continue to include it. Extension modules wishing to use the new BList type must do so explicitly. The BList C interface can be changed to match the existing PyList interface so that a simple search-replace will be sufficient for 99% of module writers. Existing modules would continue to compile and work without change, but they would need to make a deliberate (but small) effort to migrate to the BList. The downside of this approach is that mixing modules that use BLists and array-based lists might lead to slow down if conversions are frequently necessary. Reference Implementation ======================== A reference implementations of the BList is available for CPython at [1]_. The source package also includes a pure Python implementation, originally developed as a prototype for the CPython version. Naturally, the pure Python version is rather slow and the asymptotic improvements don't win out until the list is quite large. When compiled with Py_DEBUG, the C implementation checks the BList invariants when entering and exiting most functions. An extensive set of test cases is also included in the source package. The test cases include the existing Python sequence and list test cases as a subset. When the interpreter is built with Py_DEBUG, the test cases also check for reference leaks. Porting to Other Python Variants -------------------------------- If the BList is added to the collections module, other Python variants can support it in one of three ways: 1. Make blist an alias for list. The asymptotic performance won't be as good, but it'll work. 2. Use the pure Python reference implementation. The performance for small lists won't be as good, but it'll work. 3. Port the reference implementation. Discussion ========== This proposal has been discussed briefly on the Python-3000 mailing list [3]_. Although a number of people favored the proposal, there were also some objections. Below summarizes the pros and cons as observed by posters to the thread. General comments: - Pro: Will outperform the array-based list in most cases - Pro: "I've implemented variants of this ... a few different times" - Con: Desirability and performance in actual applications is unproven Comments on adding BList to the collections module: - Pro: Matching the list-API reduces the learning curve to near-zero - Pro: Useful for intermediate-level users; won't get in the way of beginners - Con: Proliferation of data types makes the choices for developers harder. Comments on replacing the array-based list with the BList: - Con: Impact on extension modules (addressed in `Backwards Compatibility`_) - Con: The use cases where BLists are slower are important (see `Use Case Trade-Offs`_ for how these might be addressed). - Con: The array-based list code is simple and easy to maintain To assess the desirability and performance in actual applications, Raymond Hettinger suggested releasing the BList as an extension module (now available at [1]_). If it proves useful, he felt it would be a strong candidate for inclusion in 2.6 as part of the collections module. If widely popular, then it could be considered for replacing the array-based list, but not otherwise. Guido van Rossum commented that he opposed the proliferation of data types, but favored replacing the array-based list if backwards compatibility could be addressed and the BList's performance was uniformly better. On-going Tasks ============== - Reduce the memory footprint of small lists - Implement TimSort for BLists, so that best-case sorting is O(n) instead of O(log n). - Implement __reversed__ - Cache a pointer in the root to the rightmost leaf, to make LIFO operation O(n) time. References ========== .. [1] Reference Implementations for C and Python: http://www.python.org/pypi/blist/ .. [2] Empirical performance comparison between Python's array-based list and the blist: http://stutzbachenterprises.com/blist/ .. [3] Discussion on python-3000 starting at post: http://mail.python.org/pipermail/python-3000/2007-April/006757.html Copyright ========= This document has been placed in the public domain. .. Local Variables: mode: indented-text indent-tabs-mode: nil sentence-end-double-space: t fill-column: 70 coding: utf-8 End: