Benchmarks and Speed

Author: Stefan Behnel

lxml.etree is a very fast XML library. Most of this is due to the speed of libxml2, e.g. the parser and serialiser, or the XPath engine. Other areas of lxml were specifically written for high performance in high-level operations, such as the tree iterators.

On the other hand, the simplicity of lxml sometimes hides internal operations that are more costly than the API suggests. If you are not aware of these cases, lxml may not always perform as you expect. A common example in the Python world is the Python list type. New users often expect it to be a linked list, while it actually is implemented as an array, which results in a completely different complexity for common operations.

Similarly, the tree model of libxml2 is more complex than what lxml's ElementTree API projects into Python space, so some operations may show unexpected performance. Rest assured that most lxml users will not notice this in real life, as lxml is very fast in absolute numbers. It is definitely fast enough for most applications, so lxml is probably somewhere between 'fast enough' and 'the best choice' for yours. Read some messages from happy users to see what we mean.

This text describes where lxml.etree (abbreviated to 'lxe') excels, gives hints on some performance traps and compares the overall performance to the original ElementTree (ET) and cElementTree (cET) libraries by Fredrik Lundh. The cElementTree library is a fast C-implementation of the original ElementTree.


General notes

First thing to say: there is an overhead involved in having a DOM-like C library mimic the ElementTree API. As opposed to ElementTree, lxml has to generate Python representations of tree nodes on the fly when asked for them, and the internal tree structure of libxml2 results in a higher maintenance overhead than the simpler top-down structure of ElementTree. What this means is: the more of your code runs in Python, the less you can benefit from the speed of lxml and libxml2. Note, however, that this is true for most performance critical Python applications. No one would implement fourier transformations in pure Python when you can use NumPy.

The up side then is that lxml provides powerful tools like tree iterators, XPath and XSLT, that can handle complex operations at the speed of C. Their pythonic API in lxml makes them so flexible that most applications can easily benefit from them.

How to read the timings

The statements made here are backed by the (micro-)benchmark scripts, and that come with the lxml source distribution. They are distributed under the same BSD license as lxml itself, and the lxml project would like to promote them as a general benchmarking suite for all ElementTree implementations. New benchmarks are very easy to add as tiny test methods, so if you write a performance test for a specific part of the API yourself, please consider sending it to the lxml mailing list.

The timings cited below compare lxml 2.3 (with libxml2 2.7.6) to the latest developer versions of ElementTree (1.3beta2) and cElementTree (1.0.6a3). They were run single-threaded on a 2.5GHz 64bit Intel Core Duo machine under Ubuntu Linux 9.10 (Karmic). The C libraries were compiled with the same platform specific optimisation flags. The Python interpreter (2.6.4) was also manually compiled for the platform. Note that many of the following ElementTree timings are therefore better then what a normal Python installation with the standard library (c)ElementTree modules would yield.

The scripts run a number of simple tests on the different libraries, using different XML tree configurations: different tree sizes (T1-4), with or without attributes (-/A), with or without ASCII string or unicode text (-/S/U), and either against a tree or its serialised XML form (T/X). In the result extracts cited below, T1 refers to a 3-level tree with many children at the third level, T2 is swapped around to have many children below the root element, T3 is a deep tree with few children at each level and T4 is a small tree, slightly broader than deep. If repetition is involved, this usually means running the benchmark in a loop over all children of the tree root, otherwise, the operation is run on the root node (C/R).

As an example, the character code (SATR T1) states that the benchmark was running for tree T1, with plain string text (S) and attributes (A). It was run against the root element (R) in the tree structure of the data (T).

Note that very small operations are repeated in integer loops to make them measurable. It is therefore not always possible to compare the absolute timings of, say, a single access benchmark (which usually loops) and a 'get all in one step' benchmark, which already takes enough time to be measurable and is therefore measured as is. An example is the index access to a single child, which cannot be compared to the timings for getchildren(). Take a look at the concrete benchmarks in the scripts to understand how the numbers compare.

Parsing and Serialising

Serialisation is an area where lxml excels. The reason is that it executes entirely at the C level, without any interaction with Python code. The results are rather impressive, especially for UTF-8, which is native to libxml2. While 20 to 40 times faster than (c)ElementTree 1.2 (which is part of the standard library since Python 2.5), lxml is still more than 7 times as fast as the much improved ElementTree 1.3:

lxe: tostring_utf16  (S-TR T1)    9.8219 msec/pass
cET: tostring_utf16  (S-TR T1)   88.7740 msec/pass
ET : tostring_utf16  (S-TR T1)   99.6690 msec/pass

lxe: tostring_utf16  (UATR T1)   10.3750 msec/pass
cET: tostring_utf16  (UATR T1)   90.7581 msec/pass
ET : tostring_utf16  (UATR T1)  102.3569 msec/pass

lxe: tostring_utf16  (S-TR T2)   10.2711 msec/pass
cET: tostring_utf16  (S-TR T2)   93.5340 msec/pass
ET : tostring_utf16  (S-TR T2)  105.8500 msec/pass

lxe: tostring_utf8   (S-TR T2)    7.1261 msec/pass
cET: tostring_utf8   (S-TR T2)   93.4091 msec/pass
ET : tostring_utf8   (S-TR T2)  105.5419 msec/pass

lxe: tostring_utf8   (U-TR T3)    1.4591 msec/pass
cET: tostring_utf8   (U-TR T3)   29.6180 msec/pass
ET : tostring_utf8   (U-TR T3)   31.9080 msec/pass

The same applies to plain text serialisation. Note that the cElementTree version in the standard library does not currently support this, as it is a new feature in ET 1.3 and lxml.etree 2.0:

lxe: tostring_text_ascii     (S-TR T1)    1.9400 msec/pass
cET: tostring_text_ascii     (S-TR T1)   41.6231 msec/pass
ET : tostring_text_ascii     (S-TR T1)   52.7501 msec/pass

lxe: tostring_text_ascii     (S-TR T3)    0.5331 msec/pass
cET: tostring_text_ascii     (S-TR T3)   12.9712 msec/pass
ET : tostring_text_ascii     (S-TR T3)   15.3620 msec/pass

lxe: tostring_text_utf16     (S-TR T1)    3.2430 msec/pass
cET: tostring_text_utf16     (S-TR T1)   41.9259 msec/pass
ET : tostring_text_utf16     (S-TR T1)   53.4091 msec/pass

lxe: tostring_text_utf16     (U-TR T1)    3.6838 msec/pass
cET: tostring_text_utf16     (U-TR T1)   38.7859 msec/pass
ET : tostring_text_utf16     (U-TR T1)   50.8440 msec/pass

Unlike ElementTree, the tostring() function in lxml also supports serialisation to a Python unicode string object:

lxe: tostring_text_unicode   (S-TR T1)    2.4869 msec/pass
lxe: tostring_text_unicode   (U-TR T1)    3.0370 msec/pass
lxe: tostring_text_unicode   (S-TR T3)    0.6518 msec/pass
lxe: tostring_text_unicode   (U-TR T3)    0.7300 msec/pass

For parsing, lxml.etree and cElementTree compete for the medal. Depending on the input, either of the two can be faster. The (c)ET libraries use a very thin layer on top of the expat parser, which is known to be very fast. Here are some timings from the benchmarking suite:

lxe: parse_stringIO  (SAXR T1)   19.9990 msec/pass
cET: parse_stringIO  (SAXR T1)    8.4970 msec/pass
ET : parse_stringIO  (SAXR T1)  183.9781 msec/pass

lxe: parse_stringIO  (S-XR T3)    2.0790 msec/pass
cET: parse_stringIO  (S-XR T3)    2.7430 msec/pass
ET : parse_stringIO  (S-XR T3)   47.4229 msec/pass

lxe: parse_stringIO  (UAXR T3)   11.1630 msec/pass
cET: parse_stringIO  (UAXR T3)   15.0940 msec/pass
ET : parse_stringIO  (UAXR T3)   92.6890 msec/pass

And another couple of timings from a benchmark that Fredrik Lundh used to promote cElementTree, comparing a number of different parsers. First, parsing a 274KB XML file containing Shakespeare's Hamlet:

lxml.etree.parse done in 0.005 seconds
cElementTree.parse done in 0.012 seconds
elementtree.ElementTree.parse done in 0.136 seconds
elementtree.XMLTreeBuilder: 6636 nodes read in 0.243 seconds
elementtree.SimpleXMLTreeBuilder: 6636 nodes read in 0.314 seconds
elementtree.SgmlopXMLTreeBuilder: 6636 nodes read in 0.104 seconds
minidom tree read in 0.137 seconds

And a 3.4MB XML file containing the Old Testament:

lxml.etree.parse done in 0.031 seconds
cElementTree.parse done in 0.039 seconds
elementtree.ElementTree.parse done in 0.537 seconds
elementtree.XMLTreeBuilder: 25317 nodes read in 0.577 seconds
elementtree.SimpleXMLTreeBuilder: 25317 nodes read in 1.265 seconds
elementtree.SgmlopXMLTreeBuilder: 25317 nodes read in 0.331 seconds
minidom tree read in 0.643 seconds

Here are the same benchmarks run under an early alpha of CPython 3.3, but on a different machine, which makes the absolute numbers uncomparable. This time, however, we include the memory usage of the process in KB before and after parsing (using os.fork() to make sure we start from a clean state each time). For the 274KB hamlet.xml file:

Memory usage at start: 7288
xml.etree.ElementTree.parse done in 0.104 seconds
Memory usage: 14252 (+6964)
xml.etree.cElementTree.parse done in 0.016 seconds
Memory usage: 9748 (+2460)
lxml.etree.parse done in 0.017 seconds
Memory usage: 11040 (+3752)
lxml.etree[remove_blank_space].parse done in 0.015 seconds
Memory usage: 10088 (+2800)
minidom tree read in 0.152 seconds
Memory usage: 30376 (+23088)

And for the 3.4MB Old Testament XML file:

Memory usage at start: 20456
xml.etree.ElementTree.parse done in 0.419 seconds
Memory usage: 46112 (+25656)
xml.etree.cElementTree.parse done in 0.054 seconds
Memory usage: 32644 (+12188)
lxml.etree.parse done in 0.041 seconds
Memory usage: 37508 (+17052)
lxml.etree[remove_blank_space].parse done in 0.037 seconds
Memory usage: 34356 (+13900)
minidom tree read in 0.671 seconds
Memory usage: 110448 (+89992)

As can be seen from the sizes, both lxml.etree and cElementTree are rather memory friendly compared to the pure Python libraries ElementTree and (especially) minidom. And the timings speak for themselves anyway.

For plain parser performance, lxml.etree and cElementTree tend to stay rather close to each other, usually within a factor of two, with winners well distributed over both sides. Similar timings can be observed for the iterparse() function:

lxe: iterparse_stringIO  (SAXR T1)   24.8621 msec/pass
cET: iterparse_stringIO  (SAXR T1)   17.3280 msec/pass
ET : iterparse_stringIO  (SAXR T1)  199.1270 msec/pass

lxe: iterparse_stringIO  (UAXR T3)   12.3630 msec/pass
cET: iterparse_stringIO  (UAXR T3)   17.5190 msec/pass
ET : iterparse_stringIO  (UAXR T3)   95.8610 msec/pass

However, if you benchmark the complete round-trip of a serialise-parse cycle, the numbers will look similar to these:

lxe: write_utf8_parse_stringIO  (S-TR T1)   27.5791 msec/pass
cET: write_utf8_parse_stringIO  (S-TR T1)  158.9060 msec/pass
ET : write_utf8_parse_stringIO  (S-TR T1)  347.8320 msec/pass

lxe: write_utf8_parse_stringIO  (UATR T2)   34.4141 msec/pass
cET: write_utf8_parse_stringIO  (UATR T2)  187.7041 msec/pass
ET : write_utf8_parse_stringIO  (UATR T2)  388.9449 msec/pass

lxe: write_utf8_parse_stringIO  (S-TR T3)    3.7861 msec/pass
cET: write_utf8_parse_stringIO  (S-TR T3)   52.4600 msec/pass
ET : write_utf8_parse_stringIO  (S-TR T3)  101.4550 msec/pass

lxe: write_utf8_parse_stringIO  (SATR T4)    0.5522 msec/pass
cET: write_utf8_parse_stringIO  (SATR T4)    3.8941 msec/pass
ET : write_utf8_parse_stringIO  (SATR T4)    5.9431 msec/pass

For applications that require a high parser throughput of large files, and that do little to no serialization, both cET and lxml.etree are a good choice. The cET library is particularly fast for iterparse applications that extract small amounts of data or aggregate information from large XML data sets that do not fit into memory. If it comes to round-trip performance, however, lxml is multiple times faster in total. So, whenever the input documents are not considerably larger than the output, lxml is the clear winner.

Regarding HTML parsing, Ian Bicking has done some benchmarking on lxml's HTML parser, comparing it to a number of other famous HTML parser tools for Python. lxml wins this contest by quite a length. To give an idea, the numbers suggest that lxml.html can run a couple of parse-serialise cycles in the time that other tools need for parsing alone. The comparison even shows some very favourable results regarding memory consumption.

Liza Daly has written an article that presents a couple of tweaks to get the most out of lxml's parser for very large XML documents. She quite favourably positions lxml.etree as a tool for high-performance XML parsing.

Finally, has a couple of publications about XML parser performance. Farwick and Hafner have written two interesting articles that compare the parser of libxml2 to some major Java based XML parsers. One deals with event-driven parser performance, the other one presents benchmark results comparing DOM parsers. Both comparisons suggest that libxml2's parser performance is largely superiour to all commonly used Java parsers in almost all cases. Note that the C parser benchmark results are based on xmlbench, which uses a simpler setup for libxml2 than lxml does.

The ElementTree API

Since all three libraries implement the same API, their performance is easy to compare in this area. A major disadvantage for lxml's performance is the different tree model that underlies libxml2. It allows lxml to provide parent pointers for elements and full XPath support, but also increases the overhead of tree building and restructuring. This can be seen from the tree setup times of the benchmark (given in seconds):

lxe:       --     S-     U-     -A     SA     UA
     T1: 0.0407 0.0470 0.0506 0.0396 0.0464 0.0504
     T2: 0.0480 0.0557 0.0584 0.0520 0.0608 0.0627
     T3: 0.0118 0.0132 0.0136 0.0319 0.0322 0.0319
     T4: 0.0002 0.0002 0.0002 0.0006 0.0006 0.0006
cET:       --     S-     U-     -A     SA     UA
     T1: 0.0045 0.0043 0.0043 0.0045 0.0043 0.0043
     T2: 0.0068 0.0069 0.0066 0.0078 0.0070 0.0069
     T3: 0.0040 0.0040 0.0040 0.0050 0.0052 0.0067
     T4: 0.0000 0.0000 0.0000 0.0001 0.0001 0.0001
ET :       --     S-     U-     -A     SA     UA
     T1: 0.0479 0.1051 0.1279 0.0487 0.1597 0.0484
     T2: 0.1995 0.0553 0.2297 0.2550 0.0550 0.2881
     T3: 0.0177 0.0169 0.0174 0.0185 0.2895 0.0189
     T4: 0.0003 0.0002 0.0003 0.0003 0.0014 0.0003

While lxml is still a lot faster than ET in most cases, cET can be several times faster than lxml here. One of the reasons is that lxml must encode incoming string data and tag names into UTF-8, and additionally discard the created Python elements after their use, when they are no longer referenced. ET and cET represent the tree itself through these objects, which reduces the overhead in creating them.

Child access

The same reason makes operations like collecting children as in list(element) more costly in lxml. Where ET and cET can quickly create a shallow copy of their list of children, lxml has to create a Python object for each child and collect them in a list:

lxe: root_list_children        (--TR T1)    0.0079 msec/pass
cET: root_list_children        (--TR T1)    0.0029 msec/pass
ET : root_list_children        (--TR T1)    0.0100 msec/pass

lxe: root_list_children        (--TR T2)    0.0849 msec/pass
cET: root_list_children        (--TR T2)    0.0110 msec/pass
ET : root_list_children        (--TR T2)    0.1481 msec/pass

This handicap is also visible when accessing single children:

lxe: first_child               (--TR T2)    0.0699 msec/pass
cET: first_child               (--TR T2)    0.0608 msec/pass
ET : first_child               (--TR T2)    0.3419 msec/pass

lxe: last_child                (--TR T1)    0.0710 msec/pass
cET: last_child                (--TR T1)    0.0648 msec/pass
ET : last_child                (--TR T1)    0.3309 msec/pass

... unless you also add the time to find a child index in a bigger list. ET and cET use Python lists here, which are based on arrays. The data structure used by libxml2 is a linked tree, and thus, a linked list of children:

lxe: middle_child              (--TR T1)    0.0989 msec/pass
cET: middle_child              (--TR T1)    0.0598 msec/pass
ET : middle_child              (--TR T1)    0.3390 msec/pass

lxe: middle_child              (--TR T2)    2.7599 msec/pass
cET: middle_child              (--TR T2)    0.0620 msec/pass
ET : middle_child              (--TR T2)    0.3610 msec/pass

Element creation

As opposed to ET, libxml2 has a notion of documents that each element must be in. This results in a major performance difference for creating independent Elements that end up in independently created documents:

lxe: create_elements           (--TC T2)    1.1640 msec/pass
cET: create_elements           (--TC T2)    0.0808 msec/pass
ET : create_elements           (--TC T2)    0.5801 msec/pass

Therefore, it is always preferable to create Elements for the document they are supposed to end up in, either as SubElements of an Element or using the explicit Element.makeelement() call:

lxe: makeelement               (--TC T2)    1.2751 msec/pass
cET: makeelement               (--TC T2)    0.1469 msec/pass
ET : makeelement               (--TC T2)    0.7451 msec/pass

lxe: create_subelements        (--TC T2)    1.1470 msec/pass
cET: create_subelements        (--TC T2)    0.1080 msec/pass
ET : create_subelements        (--TC T2)    1.4369 msec/pass

So, if the main performance bottleneck of an application is creating large XML trees in memory through calls to Element and SubElement, cET is the best choice. Note, however, that the serialisation performance may even out this advantage, especially for smaller trees and trees with many attributes.

Merging different sources

A critical action for lxml is moving elements between document contexts. It requires lxml to do recursive adaptations throughout the moved tree structure.

The following benchmark appends all root children of the second tree to the root of the first tree:

lxe: append_from_document      (--TR T1,T2)    2.0740 msec/pass
cET: append_from_document      (--TR T1,T2)    0.1271 msec/pass
ET : append_from_document      (--TR T1,T2)    0.4020 msec/pass

lxe: append_from_document      (--TR T3,T4)    0.0229 msec/pass
cET: append_from_document      (--TR T3,T4)    0.0088 msec/pass
ET : append_from_document      (--TR T3,T4)    0.0291 msec/pass

Although these are fairly small numbers compared to parsing, this easily shows the different performance classes for lxml and (c)ET. Where the latter do not have to care about parent pointers and tree structures, lxml has to deep traverse the appended tree. The performance difference therefore increases with the size of the tree that is moved.

This difference is not always as visible, but applies to most parts of the API, like inserting newly created elements:

lxe: insert_from_document       (--TR T1,T2)    7.2598 msec/pass
cET: insert_from_document       (--TR T1,T2)    0.1578 msec/pass
ET : insert_from_document       (--TR T1,T2)    0.5150 msec/pass

or replacing the child slice by a newly created element:

lxe: replace_children_element   (--TC T1)    0.1149 msec/pass
cET: replace_children_element   (--TC T1)    0.0110 msec/pass
ET : replace_children_element   (--TC T1)    0.0558 msec/pass

as opposed to replacing the slice with an existing element from the same document:

lxe: replace_children           (--TC T1)    0.0091 msec/pass
cET: replace_children           (--TC T1)    0.0060 msec/pass
ET : replace_children           (--TC T1)    0.0188 msec/pass

While these numbers are too small to provide a major performance impact in practice, you should keep this difference in mind when you merge very large trees.


Deep copying a tree is fast in lxml:

lxe: deepcopy_all              (--TR T1)    5.0900 msec/pass
cET: deepcopy_all              (--TR T1)   57.9181 msec/pass
ET : deepcopy_all              (--TR T1)  499.1000 msec/pass

lxe: deepcopy_all              (-ATR T2)    6.3980 msec/pass
cET: deepcopy_all              (-ATR T2)   65.6390 msec/pass
ET : deepcopy_all              (-ATR T2)  526.5379 msec/pass

lxe: deepcopy_all              (S-TR T3)    1.4491 msec/pass
cET: deepcopy_all              (S-TR T3)   14.7018 msec/pass
ET : deepcopy_all              (S-TR T3)  123.5120 msec/pass

So, for example, if you have a database-like scenario where you parse in a large tree and then search and copy independent subtrees from it for further processing, lxml is by far the best choice here.

Tree traversal

Another area where lxml is very fast is iteration for tree traversal. If your algorithms can benefit from step-by-step traversal of the XML tree and especially if few elements are of interest or the target element tag name is known, lxml is a good choice:

lxe: getiterator_all      (--TR T1)    1.6890 msec/pass
cET: getiterator_all      (--TR T1)   23.8621 msec/pass
ET : getiterator_all      (--TR T1)   11.1070 msec/pass

lxe: getiterator_islice   (--TR T2)    0.0188 msec/pass
cET: getiterator_islice   (--TR T2)    0.1841 msec/pass
ET : getiterator_islice   (--TR T2)   11.7059 msec/pass

lxe: getiterator_tag      (--TR T2)    0.0119 msec/pass
cET: getiterator_tag      (--TR T2)    0.3560 msec/pass
ET : getiterator_tag      (--TR T2)   10.6668 msec/pass

lxe: getiterator_tag_all  (--TR T2)    0.2429 msec/pass
cET: getiterator_tag_all  (--TR T2)   20.3710 msec/pass
ET : getiterator_tag_all  (--TR T2)   10.6280 msec/pass

This translates directly into similar timings for Element.findall():

lxe: findall              (--TR T2)    2.4588 msec/pass
cET: findall              (--TR T2)   24.1358 msec/pass
ET : findall              (--TR T2)   13.0949 msec/pass

lxe: findall              (--TR T3)    0.5939 msec/pass
cET: findall              (--TR T3)    6.9802 msec/pass
ET : findall              (--TR T3)    3.8991 msec/pass

lxe: findall_tag          (--TR T2)    0.2789 msec/pass
cET: findall_tag          (--TR T2)   20.5719 msec/pass
ET : findall_tag          (--TR T2)   10.8678 msec/pass

lxe: findall_tag          (--TR T3)    0.1638 msec/pass
cET: findall_tag          (--TR T3)    5.0790 msec/pass
ET : findall_tag          (--TR T3)    2.5120 msec/pass

Note that all three libraries currently use the same Python implementation for .findall(), except for their native tree iterator (element.iter()).


The following timings are based on the benchmark script

This part of lxml does not have an equivalent in ElementTree. However, lxml provides more than one way of accessing it and you should take care which part of the lxml API you use. The most straight forward way is to call the xpath() method on an Element or ElementTree:

lxe: xpath_method         (--TC T1)    0.7598 msec/pass
lxe: xpath_method         (--TC T2)   12.6798 msec/pass
lxe: xpath_method         (--TC T3)    0.0758 msec/pass
lxe: xpath_method         (--TC T4)    0.6182 msec/pass

This is well suited for testing and when the XPath expressions are as diverse as the trees they are called on. However, if you have a single XPath expression that you want to apply to a larger number of different elements, the XPath class is the most efficient way to do it:

lxe: xpath_class          (--TC T1)    0.2189 msec/pass
lxe: xpath_class          (--TC T2)    1.4110 msec/pass
lxe: xpath_class          (--TC T3)    0.0319 msec/pass
lxe: xpath_class          (--TC T4)    0.0880 msec/pass

Note that this still allows you to use variables in the expression, so you can parse it once and then adapt it through variables at call time. In other cases, where you have a fixed Element or ElementTree and want to run different expressions on it, you should consider the XPathEvaluator:

lxe: xpath_element        (--TR T1)    0.1669 msec/pass
lxe: xpath_element        (--TR T2)    6.9060 msec/pass
lxe: xpath_element        (--TR T3)    0.0451 msec/pass
lxe: xpath_element        (--TR T4)    0.1681 msec/pass

While it looks slightly slower, creating an XPath object for each of the expressions generates a much higher overhead here:

lxe: xpath_class_repeat   (--TC T1)    0.7451 msec/pass
lxe: xpath_class_repeat   (--TC T2)   12.2290 msec/pass
lxe: xpath_class_repeat   (--TC T3)    0.0730 msec/pass
lxe: xpath_class_repeat   (--TC T4)    0.5970 msec/pass

A longer example

... based on lxml 1.3.

A while ago, Uche Ogbuji posted a benchmark proposal that would read in a 3MB XML version of the Old Testament of the Bible and look for the word begat in all verses. Apparently, it is contained in 120 out of almost 24000 verses. This is easy to implement in ElementTree using findall(). However, the fastest and most memory friendly way to do this is obviously iterparse(), as most of the data is not of any interest.

Now, Uche's original proposal was more or less the following:

def bench_ET():
    tree = ElementTree.parse("ot.xml")
    result = []
    for v in tree.findall("//v"):
        text = v.text
        if 'begat' in text:
    return len(result)

which takes about one second on my machine today. The faster iterparse() variant looks like this:

def bench_ET_iterparse():
    result = []
    for event, v in ElementTree.iterparse("ot.xml"):
        if v.tag == 'v':
            text = v.text
            if 'begat' in text:
    return len(result)

The improvement is about 10%. At the time I first tried (early 2006), lxml didn't have iterparse() support, but the findall() variant was already faster than ElementTree. This changes immediately when you switch to cElementTree. The latter only needs 0.17 seconds to do the trick today and only some impressive 0.10 seconds when running the iterparse version. And even back then, it was quite a bit faster than what lxml could achieve.

Since then, lxml has matured a lot and has gotten much faster. The iterparse variant now runs in 0.14 seconds, and if you remove the v.clear(), it is even a little faster (which isn't the case for cElementTree).

One of the many great tools in lxml is XPath, a swiss army knife for finding things in XML documents. It is possible to move the whole thing to a pure XPath implementation, which looks like this:

def bench_lxml_xpath_all():
    tree = etree.parse("ot.xml")
    result = tree.xpath("//v[contains(., 'begat')]/text()")
    return len(result)

This runs in about 0.13 seconds and is about the shortest possible implementation (in lines of Python code) that I could come up with. Now, this is already a rather complex XPath expression compared to the simple "//v" ElementPath expression we started with. Since this is also valid XPath, let's try this instead:

def bench_lxml_xpath():
    tree = etree.parse("ot.xml")
    result = []
    for v in tree.xpath("//v"):
        text = v.text
        if 'begat' in text:
    return len(result)

This gets us down to 0.12 seconds, thus showing that a generic XPath evaluation engine cannot always compete with a simpler, tailored solution. However, since this is not much different from the original findall variant, we can remove the complexity of the XPath call completely and just go with what we had in the beginning. Under lxml, this runs in the same 0.12 seconds.

But there is one thing left to try. We can replace the simple ElementPath expression with a native tree iterator:

def bench_lxml_getiterator():
    tree = etree.parse("ot.xml")
    result = []
    for v in tree.getiterator("v"):
        text = v.text
        if 'begat' in text:
    return len(result)

This implements the same thing, just without the overhead of parsing and evaluating a path expression. And this makes it another bit faster, down to 0.11 seconds. For comparison, cElementTree runs this version in 0.17 seconds.

So, what have we learned?


The following timings are based on the benchmark script

Objectify is a data-binding API for XML based on lxml.etree, that was added in version 1.1. It uses standard Python attribute access to traverse the XML tree. It also features ObjectPath, a fast path language based on the same meme.

Just like lxml.etree, lxml.objectify creates Python representations of elements on the fly. To save memory, the normal Python garbage collection mechanisms will discard them when their last reference is gone. In cases where deeply nested elements are frequently accessed through the objectify API, the create-discard cycles can become a bottleneck, as elements have to be instantiated over and over again.


ObjectPath can be used to speed up the access to elements that are deep in the tree. It avoids step-by-step Python element instantiations along the path, which can substantially improve the access time:

lxe: attribute                  (--TR T1)    4.8928 msec/pass
lxe: attribute                  (--TR T2)   25.5480 msec/pass
lxe: attribute                  (--TR T4)    4.6349 msec/pass

lxe: objectpath                 (--TR T1)    1.4842 msec/pass
lxe: objectpath                 (--TR T2)   21.1990 msec/pass
lxe: objectpath                 (--TR T4)    1.4892 msec/pass

lxe: attributes_deep            (--TR T1)   11.9710 msec/pass
lxe: attributes_deep            (--TR T2)   32.4290 msec/pass
lxe: attributes_deep            (--TR T4)   11.4839 msec/pass

lxe: objectpath_deep            (--TR T1)    4.8139 msec/pass
lxe: objectpath_deep            (--TR T2)   24.6511 msec/pass
lxe: objectpath_deep            (--TR T4)    4.7588 msec/pass

Note, however, that parsing ObjectPath expressions is not for free either, so this is most effective for frequently accessing the same element.

Caching Elements

A way to improve the normal attribute access time is static instantiation of the Python objects, thus trading memory for speed. Just create a cache dictionary and run:

cache[root] = list(root.iter())

after parsing and:

del cache[root]

when you are done with the tree. This will keep the Python element representations of all elements alive and thus avoid the overhead of repeated Python object creation. You can also consider using filters or generator expressions to be more selective. By choosing the right trees (or even subtrees and elements) to cache, you can trade memory usage against access speed:

lxe: attribute_cached           (--TR T1)    3.8228 msec/pass
lxe: attribute_cached           (--TR T2)   23.7138 msec/pass
lxe: attribute_cached           (--TR T4)    3.5269 msec/pass

lxe: attributes_deep_cached     (--TR T1)    4.6771 msec/pass
lxe: attributes_deep_cached     (--TR T2)   24.8699 msec/pass
lxe: attributes_deep_cached     (--TR T4)    4.3321 msec/pass

lxe: objectpath_deep_cached     (--TR T1)    1.1430 msec/pass
lxe: objectpath_deep_cached     (--TR T2)   19.7470 msec/pass
lxe: objectpath_deep_cached     (--TR T4)    1.1740 msec/pass

Things to note: you cannot currently use weakref.WeakKeyDictionary objects for this as lxml's element objects do not support weak references (which are costly in terms of memory). Also note that new element objects that you add to these trees will not turn up in the cache automatically and will therefore still be garbage collected when all their Python references are gone, so this is most effective for largely immutable trees. You should consider using a set instead of a list in this case and add new elements by hand.

Further optimisations

Here are some more things to try if optimisation is required:

  • A lot of time is usually spent in tree traversal to find the addressed elements in the tree. If you often work in subtrees, do what you would also do with deep Python objects: assign the parent of the subtree to a variable or pass it into functions instead of starting at the root. This allows accessing its descendents more directly.
  • Try assigning data values directly to attributes instead of passing them through DataElement.
  • If you use custom data types that are costly to parse, try running objectify.annotate() over read-only trees to speed up the attribute type inference on read access.

Note that none of these measures is guaranteed to speed up your application. As usual, you should prefer readable code over premature optimisations and profile your expected use cases before bothering to apply optimisations at random.