openquake.baselib package


Utility functions of general interest.

class openquake.baselib.general.AccumDict(dic=None, accum=None, **kw)[source]

Bases: dict

An accumulating dictionary, useful to accumulate variables:

>> acc = AccumDict()
>> acc += {'a': 1}
>> acc += {'a': 1, 'b': 1}
>> acc
{'a': 2, 'b': 1}
>> {'a': 1} + acc
{'a': 3, 'b': 1}
>> acc + 1
{'a': 3, 'b': 2}
>> 1 - acc
{'a': -1, 'b': 0}
>> acc - 1
{'a': 1, 'b': 0}

Also the multiplication has been defined:

>> prob1 = AccumDict(a=0.4, b=0.5)
>> prob2 = AccumDict(b=0.5)
>> prob1 * prob2
{'a': 0.4, 'b': 0.25}
>> prob1 * 1.2
{'a': 0.48, 'b': 0.6}
>> 1.2 * prob1
{'a': 0.48, 'b': 0.6}

It is very common to use an AccumDict of accumulators; here is an example using the empty list as accumulator:

>>> acc = AccumDict(accum=[])
>>> acc['a'] += [1]
>>> acc['b'] += [2]
>>> sorted(acc.items())
[('a', [1]), ('b', [2])]

The implementation is smart enough to make (deep) copies of the accumulator, therefore each key has a different accumulator, which initially is the empty list (in this case).

apply(func, *extras)[source]

>> a = AccumDict({‘a’: 1, ‘b’: 2}) >> a.apply(lambda x, y: 2 * x + y, 1) {‘a’: 3, ‘b’: 5}

class openquake.baselib.general.CallableDict(keyfunc=<function <lambda>>, keymissing=None)[source]

Bases: collections.OrderedDict

A callable object built on top of a dictionary of functions, used as a smart registry or as a poor man generic function dispatching on the first argument. It is typically used to implement converters. Here is an example:

>>> format_attrs = CallableDict()  # dict of functions (fmt, obj) -> str
>>> @format_attrs.add('csv')  # implementation for csv
... def format_attrs_csv(fmt, obj):
...     items = sorted(vars(obj).items())
...     return '\n'.join('%s,%s' % item for item in items)
>>> @format_attrs.add('json')  # implementation for json
... def format_attrs_json(fmt, obj):
...     return json.dumps(vars(obj))

format_attrs(fmt, obj) calls the correct underlying function depending on the fmt key. If the format is unknown a KeyError is raised. It is also possible to set a keymissing function to specify what to return if the key is missing.

For a more practical example see the implementation of the exporters in openquake.calculators.export


Return a decorator registering a new implementation for the CallableDict for the given keys.

exception openquake.baselib.general.CodeDependencyError[source]

Bases: exceptions.Exception

exception openquake.baselib.general.DeprecationWarning[source]

Bases: exceptions.UserWarning

Raised the first time a deprecated function is called

class openquake.baselib.general.DictArray(imtls)[source]

Bases: _abcoll.Mapping

A small wrapper over a dictionary of arrays serializable to HDF5:

>>> d = DictArray({'PGA': [0.01, 0.02, 0.04], 'PGV': [0.1, 0.2]})
>>> from openquake.baselib import hdf5
>>> with hdf5.File('/tmp/x.h5', 'w') as f:
...      f['d'] = d
...      f['d']
PGA: [ 0.01  0.02  0.04]
PGV: [ 0.1  0.2]>

The DictArray maintains the lexicographic order of the keys.


Convert an array of compatible length into a DictArray:

>>> d = DictArray({'PGA': [0.01, 0.02, 0.04], 'PGV': [0.1, 0.2]})
>>>, 5, 1))  # array of lenght 5 = 3 + 2
PGA: [0 1 2]
PGV: [3 4]>
class openquake.baselib.general.WeightedSequence(seq=())[source]

Bases: _abcoll.MutableSequence

A wrapper over a sequence of weighted items with a total weight attribute. Adding items automatically increases the weight.

insert(i, item_weight)[source]

Insert an item with the given weight in the sequence

classmethod merge(ws_list)[source]

Merge a set of WeightedSequence objects.

Parameters:ws_list – a sequence of :class: openquake.baselib.general.WeightedSequence instances
Returns:a openquake.baselib.general.WeightedSequence instance
openquake.baselib.general.assert_close(a, b, rtol=1e-07, atol=0, context=None)[source]

Compare for equality up to a given precision two composite objects which may contain floats. NB: if the objects are or contain generators, they are exhausted.

  • a – an object
  • b – another object
  • rtol – relative tolerance
  • atol – absolute tolerance
openquake.baselib.general.assert_independent(package, *packages)[source]
  • package – Python name of a module/package
  • packages – Python names of modules/packages

Make sure the package does not depend from the packages.

openquake.baselib.general.block_splitter(items, max_weight, weight=<function <lambda>>, kind=<function nokey>)[source]
  • items – an iterator over items
  • max_weight – the max weight to split on
  • weight – a function returning the weigth of a given item
  • kind – a function returning the kind of a given item

Group together items of the same kind until the total weight exceeds the max_weight and yield WeightedSequence instances. Items with weight zero are ignored.

For instance

>>> items = 'ABCDE'
>>> list(block_splitter(items, 3))
[<WeightedSequence ['A', 'B', 'C'], weight=3>, <WeightedSequence ['D', 'E'], weight=2>]

The default weight is 1 for all items.

openquake.baselib.general.ceil(a, b)[source]

Divide a / b and return the biggest integer close to the quotient.

  • a – a number
  • b – a positive number

the biggest integer close to the quotient


Return a decorator to make deprecated functions.

Parameters:message – the message to print the first time the deprecated function is used.

Here is an example of usage:

>>> @deprecated('Use new_function instead')
... def old_function():
...     'Do something'

Notice that if the function is called several time, the deprecation warning will be displayed only the first time.


Return the distinct keys in order.

openquake.baselib.general.get_array(array, **kw)[source]

Extract a subarray by filtering on the given keyword arguments

Returns:<short git hash> if Git repository found
openquake.baselib.general.group_array(array, *kfields)[source]

Convert an array into an OrderedDict kfields -> array

openquake.baselib.general.groupby(objects, key, reducegroup=<type 'list'>)[source]
  • objects – a sequence of objects with a key value
  • key – the key function to extract the key value
  • reducegroup – the function to apply to each group

an OrderedDict {key value: map(reducegroup, group)}

>>> groupby(['A1', 'A2', 'B1', 'B2', 'B3'], lambda x: x[0],
...         lambda group: ''.join(x[1] for x in group))
OrderedDict([('A', '12'), ('B', '123')])
openquake.baselib.general.groupby2(records, kfield, vfield)[source]
  • records – a sequence of records with positional or named fields
  • kfield – the index/name/tuple specifying the field to use as a key
  • vfield – the index/name/tuple specifying the field to use as a value

an list of pairs of the form (key, [value, ...]).

>>> groupby2(['A1', 'A2', 'B1', 'B2', 'B3'], 0, 1)
[('A', ['1', '2']), ('B', ['1', '2', '3'])]

Here is an example where the keyfield is a tuple of integers:

>>> groupby2(['A11', 'A12', 'B11', 'B21'], (0, 1), 2)
[(('A', '1'), ['1', '2']), (('B', '1'), ['1']), (('B', '2'), ['1'])]
openquake.baselib.general.humansize(nbytes, suffixes=('B', 'KB', 'MB', 'GB', 'TB', 'PB'))[source]

Return file size in a human-friendly format


If module_or_package is a module, just import it; if it is a package, recursively imports all the modules it contains. Returns the names of the modules that were imported as a set. The set can be empty if the modules were already in sys.modules.


Dummy function to apply to items without a key

openquake.baselib.general.run_in_process(code, *args)[source]

Run in an external process the given Python code and return the output as a Python object. If there are arguments, then code is taken as a template and traditional string interpolation is performed.

  • code – string or template describing Python code
  • args – arguments to be used for interpolation

the output of the process, as a Python object

openquake.baselib.general.safeprint(*args, **kwargs)[source]

Convert and print characters using the proper encoding

openquake.baselib.general.search_module(module, syspath=['/usr/share/sphinx/scripts/python2', '/home/ubuntu/oq-engine', '/opt/openquake/lib/python2.7/site-packages', '/usr/lib/python2.7', '/usr/lib/python2.7/plat-x86_64-linux-gnu', '/usr/lib/python2.7/lib-tk', '/usr/lib/python2.7/lib-old', '/usr/lib/python2.7/lib-dynload', '/usr/local/lib/python2.7/dist-packages', '/usr/lib/python2.7/dist-packages'])[source]

Given a module name (possibly with dots) returns the corresponding filepath, or None, if the module cannot be found.

  • module – (dotted) name of the Python module to look for
  • syspath – a list of directories to search (default sys.path)
openquake.baselib.general.split_in_blocks(sequence, hint, weight=<function <lambda>>, key=<function nokey>)[source]

Split the sequence in a number of WeightedSequences close to hint.

  • sequence – a finite sequence of items
  • hint – an integer suggesting the number of subsequences to generate
  • weight – a function returning the weigth of a given item
  • key – a function returning the key of a given item

The WeightedSequences are of homogeneous key and they try to be balanced in weight. For instance

>>> items = 'ABCDE'
>>> list(split_in_blocks(items, 3))
[<WeightedSequence ['A', 'B'], weight=2>, <WeightedSequence ['C', 'D'], weight=2>, <WeightedSequence ['E'], weight=1>]
openquake.baselib.general.split_in_slices(number, num_slices)[source]
  • number – a positive number to split in slices
  • num_slices – the number of slices to return (at most)

a list of slices

>>> split_in_slices(4, 2)
[slice(0, 2, None), slice(2, 4, None)]
>>> split_in_slices(5, 1)
[slice(0, 5, None)]
>>> split_in_slices(5, 2)
[slice(0, 3, None), slice(3, 5, None)]
>>> split_in_slices(2, 4)
[slice(0, 1, None), slice(1, 2, None)]
openquake.baselib.general.writetmp(content=None, dir=None, prefix='tmp', suffix='tmp')[source]

Create temporary file with the given content.

Please note: the temporary file must be deleted by the caller.

  • content (string) – the content to write to the temporary file.
  • dir (string) – directory where the file should be created
  • prefix (string) – file name prefix
  • suffix (string) – file name suffix

a string with the path to the temporary file


class openquake.baselib.hdf5.ByteCounter(nbytes=0)[source]

Bases: object

A visitor used to measure the dimensions of a HDF5 dataset or group. Use it as ByteCounter.get_nbytes(dset_or_group).

classmethod get_nbytes(dset)[source]
class openquake.baselib.hdf5.File(name, mode=None, driver=None, libver=None, userblock_size=None, swmr=False, **kwds)[source]

Bases: h5py._hl.files.File

Subclass of h5py.File able to store and retrieve objects conforming to the HDF5 protocol used by the OpenQuake software. It works recursively also for dictionaries of the form name->obj.

>>> f = File('/tmp/x.h5', 'w')
>>> f['dic'] = dict(a=dict(x=1, y=2), b=3)
>>> dic = f['dic']
>>> dic['a']['x'].value
>>> dic['b'].value
>>> f.close()
save(nodedict, root='')[source]

Save a node dictionary in the .hdf5 file, starting from the root dataset. A common application is to convert XML files into .hdf5 files, see the usage in openquake.commands.to_hdf5.

Parameters:nodedict – a dictionary with keys ‘tag’, ‘attrib’, ‘text’, ‘nodes’
set_nbytes(key, nbytes=None)[source]

Set the nbytes attribute on the HDF5 object identified by key.

classmethod temporary()[source]

Returns a temporary hdf5 file, open for writing. The temporary name is stored in the .path attribute. It is the user responsability to remove the file when closed.

class openquake.baselib.hdf5.LiteralAttrs[source]

Bases: object

A class to serialize a set of parameters in HDF5 format. The goal is to store simple parameters as an HDF5 table in a readable way. Each parameter can be retrieved as an attribute, given its name. The implementation treats specially dictionary attributes, by storing them as attrname.keyname strings, see the example below:

>>> class Ser(LiteralAttrs):
...     def __init__(self, a, b):
...         self.a = a
...         self.b = b
>>> ser = Ser(1, dict(x='xxx', y='yyy'))
>>> arr, attrs = ser.__toh5__()
>>> for k, v in arr:
...     print('%s=%s' % (k, v))
>>> s = object.__new__(Ser)
>>> s.__fromh5__(arr, attrs)
>>> s.a
>>> s.b['x']

The implementation is not recursive, i.e. there will be at most one dot in the serialized names (in the example here a, b.x, b.y).

class openquake.baselib.hdf5.PickleableSequence(objects)[source]

Bases: _abcoll.Sequence

An immutable sequence of pickleable objects that can be serialized in HDF5 format. Here is an example, using the LiteralAttrs class defined in this module, but any pickleable class would do:

>>> seq = PickleableSequence([LiteralAttrs(), LiteralAttrs()])
>>> with File('/tmp/x.h5', 'w') as f:
...     f['data'] = seq
>>> with File('/tmp/x.h5') as f:
...     f['data']
(<LiteralAttrs >, <LiteralAttrs >)
Parameters:lst – a list of strings or bytes
Returns:an array of variable length ASCII strings

The full Python name (i.e. pkg.subpkg.mod.cls) of a class

openquake.baselib.hdf5.create(hdf5, name, dtype, shape=(None, ), compression=None, fillvalue=0, attrs=None)[source]
  • hdf5 – a h5py.File object
  • name – an hdf5 key string
  • dtype – dtype of the dataset (usually composite)
  • shape – shape of the dataset (can be extendable)
  • compression – None or ‘gzip’ are recommended
  • attrs – dictionary of attributes of the dataset

a HDF5 dataset


The class associated to the given dotname (i.e. pkg.subpkg.mod.cls)

openquake.baselib.hdf5.extend(dset, array)[source]

Extend an extensible dataset with an array of a compatible dtype.

  • dset – an h5py dataset
  • array – an array of length L

the total length of the dataset (i.e. initial length + L)

openquake.baselib.hdf5.extend3(hdf5path, key, array, **attrs)[source]

Extend an HDF5 file dataset with the given array


If the dataset has an attribute ‘nbytes’, return it. Otherwise get the size of the underlying array. Returns None if the dataset is actually a group., dic, **extra)[source]
  • path – an .hdf5 pathname
  • dic – a dictionary of array-like objects
  • extra – extra attributes to be saved in the file


This module defines a Node class, together with a few conversion functions which are able to convert NRML files into hierarchical objects (DOM). That makes it easier to read and write XML from Python and viceversa. Such features are used in the command-line conversion tools. The Node class is kept intentionally similar to an Element class, however it overcomes the limitation of ElementTree: in particular a node can manage a lazy iterable of subnodes, whereas ElementTree wants to keep everything in memory. Moreover the Node class provides a convenient dot notation to access subnodes.

The Node class is instantiated with four arguments:

  1. the node tag (a mandatory string)
  2. the node attributes (a dictionary)
  3. the node value (a string or None)
  4. the subnodes (an iterable over nodes)

If a node has subnodes, its value should be None.

For instance, here is an example of instantiating a root node with two subnodes a and b:

>>> from openquake.baselib.node import Node
>>> a = Node('a', {}, 'A1')
>>> b = Node('b', {'attrb': 'B'}, 'B1')
>>> root = Node('root', nodes=[a, b])
>>> root
<root {} None ...>

Node objects can be converted into nicely indented strings:

>>> print(root.to_str())
  a 'A1'
  b{attrb='B'} 'B1'

The subnodes can be retrieved with the dot notation:

>>> root.a
<a {} A1 >

The value of a node can be extracted with the ~ operator:

>>> ~root.a

If there are multiple subnodes with the same name

>>> root.append(Node('a', {}, 'A2'))  # add another 'a' node

the dot notation will retrieve the first node.

It is possible to retrieve the other nodes from the ordinal index:

>>> root[0], root[1], root[2]
(<a {} A1 >, <b {'attrb': 'B'} B1 >, <a {} A2 >)

The list of all subnodes with a given name can be retrieved as follows:

>>> list(root.getnodes('a'))
[<a {} A1 >, <a {} A2 >]

It is also possible to delete a node given its index:

>>> del root[2]

A node is an iterable object yielding its subnodes:

>>> list(root)
[<a {} A1 >, <b {'attrb': 'B'} B1 >]

The attributes of a node can be retrieved with the square bracket notation:

>>> root.b['attrb']

It is possible to add and remove attributes freely:

>>> root.b['attr'] = 'new attr'
>>> del root.b['attr']

Node objects can be easily converted into ElementTree objects:

>>> node_to_elem(root)  
<Element 'root' at ...>

Then is trivial to generate the XML representation of a node:

>>> from xml.etree import ElementTree
>>> print(ElementTree.tostring(node_to_elem(root)).decode('utf-8'))
<root><a>A1</a><b attrb="B">B1</b></root>

Generating XML files larger than the available memory requires some care. The trick is to use a node generator, such that it is not necessary to keep the entire tree in memory. Here is an example:

>>> def gen_many_nodes(N):
...     for i in xrange(N):
...         yield Node('a', {}, 'Text for node %d' % i)
>>> lazytree = Node('lazytree', {}, nodes=gen_many_nodes(10))

The lazytree object defined here consumes no memory, because the nodes are not created a instantiation time. They are created as soon as you start iterating on the lazytree. In particular list(lazytree) will generated all of them. If your goal is to store the tree on the filesystem in XML format you should use a writing routine converting a subnode at the time, without requiring the full list of them. The routines provided by ElementTree are no good, however commonlib.writers provide an StreamingXMLWriter just for that purpose.

Lazy trees should not be used unless it is absolutely necessary in order to save memory; the problem is that if you use a lazy tree the slice notation will not work (the underlying generator will not accept it); moreover it will not be possible to iterate twice on the subnodes, since the generator will be exhausted. Notice that even accessing a subnode with the dot notation will avance the generator. Finally, nodes containing lazy nodes will not be pickleable.

class openquake.baselib.node.Node(fulltag, attrib=None, text=None, nodes=None, lineno=None)[source]

Bases: object

A class to make it easy to edit hierarchical structures with attributes, such as XML files. Node objects must be pickleable and must consume as little memory as possible. Moreover they must be easily converted from and to ElementTree objects. The advantage over ElementTree objects is that subnodes can be lazily generated and that they can be accessed with the dot notation.


Append a new subnode


Return the direct subnodes with name ‘name’

to_str(expandattrs=True, expandvals=True)[source]

Convert the node into a string, intended for testing/debugging purposes

  • expandattrs – print the values of the attributes if True, else print only the names
  • expandvals – print the values if True, else print only the tag names
class openquake.baselib.node.SourceLineParser(html=0, target=None, encoding=None)[source]

Bases: xml.etree.ElementTree.XMLParser

A custom parser managing line numbers: works for Python <= 3.3

class openquake.baselib.node.StreamingXMLWriter(bytestream, indent=4, encoding='utf-8', nsmap=None)[source]

Bases: object

A bynary stream XML writer. The typical usage is something like this:

with StreamingXMLWriter(output_file) as writer:
    for node in nodegenerator():
emptyElement(name, attrs)[source]

Add an empty element (may have attributes)


Close an XML tag


Serialize a node object (typically an ElementTree object)


Get the short representation of a fully qualified tag

Parameters:tag (str) – a (fully qualified or not) XML tag
start_tag(name, attrs=None)[source]

Open an XML tag

class openquake.baselib.node.ValidatingXmlParser(validators, stop=None)[source]

Bases: object

Validating XML Parser based on Expat. It has two methods .parse_file and .parse_bytes returning a validated Node object.

  • validators – a dictionary of validation functions
  • stop – the tag where to stop the parsing (if any)
exception Exit[source]

Bases: exceptions.Exception

Raised when the parsing is stopped before the end on purpose

ValidatingXmlParser.parse_bytes(bytestr, isfinal=True)[source]

Parse a byte string. If the string is very large, split it in chuncks and parse each chunk with isfinal=False, then parse an empty chunk with isfinal=True.


Parse a file or a filename

openquake.baselib.node.context(*args, **kwds)[source]

Context manager managing exceptions and adding line number of the current node and name of the current file to the error message.

  • fname – the current file being processed
  • node – the current node being processed
openquake.baselib.node.floatformat(*args, **kwds)[source]

Context manager to change the default format string for the function openquake.commonlib.writers.scientificformat().

Parameters:fmt_string – the format to use; for instance ‘%13.9E’

Parse an XML string and return a tree

openquake.baselib.node.iterparse(source, events=('end', ), remove_comments=True, **kw)[source]

Thin wrapper around ElementTree.iterparse

openquake.baselib.node.node_copy(node, nodefactory=<class 'openquake.baselib.node.Node'>)[source]

Make a deep copy of the node

openquake.baselib.node.node_display(root, expandattrs=False, expandvals=False, output=<open file '<stdout>', mode 'w'>)[source]

Write an indented representation of the Node object on the output; this is intended for testing/debugging purposes.

  • root – a Node object
  • expandattrs (bool) – if True, the values of the attributes are also printed, not only the names
  • expandvals (bool) – if True, the values of the tags are also printed, not only the names.
  • output – stream where to write the string representation of the node
openquake.baselib.node.node_from_dict(dic, nodefactory=<class 'openquake.baselib.node.Node'>)[source]

Convert a (nested) dictionary with attributes tag, attrib, text, nodes into a Node object.

openquake.baselib.node.node_from_elem(elem, nodefactory=<class 'openquake.baselib.node.Node'>, lazy=())[source]

Convert (recursively) an ElementTree object into a Node object.

openquake.baselib.node.node_from_ini(ini_file, nodefactory=<class 'openquake.baselib.node.Node'>, root_name='ini')[source]

Convert a .ini file into a Node object.

Parameters:ini_file – a filename or a file like object in read mode
openquake.baselib.node.node_from_xml(xmlfile, nodefactory=<class 'openquake.baselib.node.Node'>)[source]

Convert a .xml file into a Node object.

Parameters:xmlfile – a file name or file object open for reading

Convert a Node object into a (nested) dictionary with attributes tag, attrib, text, nodes.

Parameters:node – a Node-compatible object

Convert (recursively) a Node object into an ElementTree object.

openquake.baselib.node.node_to_ini(node, output=<open file '<stdout>', mode 'w'>)[source]

Convert a Node object with the right structure into a .ini file.

Params node:a Node object
Params output:a file-like object opened in write mode
openquake.baselib.node.node_to_xml(node, output=<open file '<stdout>', mode 'w'>, nsmap=None)[source]

Convert a Node object into a pretty .xml file without keeping everything in memory. If you just want the string representation use tostring(node).

  • node – a Node-compatible object (ElementTree nodes are fine)
  • nsmap – if given, shorten the tags with aliases
openquake.baselib.node.parse(source, remove_comments=True, **kw)[source]

Thin wrapper around ElementTree.parse

openquake.baselib.node.pprint(self, stream=None, indent=1, width=80, depth=None)[source]

Pretty print the underlying literal Python object

openquake.baselib.node.read_nodes(fname, filter_elem, nodefactory=<class 'openquake.baselib.node.Node'>, remove_comments=True)[source]

Convert an XML file into a lazy iterator over Node objects satifying the given specification, i.e. a function element -> boolean.

  • fname – file name of file object
  • filter_elem – element specification

In case of errors, add the file name to the error message.

openquake.baselib.node.scientificformat(value, fmt='%13.9E', sep=' ', sep2=':')[source]
  • value – the value to convert into a string
  • fmt – the formatting string to use for float values
  • sep – separator to use for vector-like values
  • sep2 – second separator to use for matrix-like values

Convert a float or an array into a string by using the scientific notation and a fixed precision (by default 10 decimal digits). For instance:

>>> scientificformat(-0E0)
>>> scientificformat(-0.004)
>>> scientificformat([0.004])
>>> scientificformat([0.01, 0.02], '%10.6E')
'1.000000E-02 2.000000E-02'
>>> scientificformat([[0.1, 0.2], [0.3, 0.4]], '%4.1E')
'1.0E-01:2.0E-01 3.0E-01:4.0E-01'

Get the short representation of a fully qualified tag

Parameters:tag (str) – a (fully qualified or not) XML tag

Convert the node into a literal Python object

openquake.baselib.node.tostring(node, indent=4, nsmap=None)[source]

Convert a node into an XML string by using the StreamingXMLWriter. This is useful for testing purposes.

  • node – a node object (typically an ElementTree object)
  • indent – the indentation to use in the XML (default 4 spaces)


The Starmap API

There are several good libraries to manage parallel programming, both in the standard library and in third party packages. Since we are not interested in reinventing the wheel, OpenQuake does not offer any new parallel library; however, it does offer some glue code so that you can use your library of choice. Currently multiprocessing, concurrent.futures, celery and ipython-parallel are supported. Moreover, openquake.baselib.parallel offers some additional facilities that make it easier to parallelize scientific computations, i.e. embarrassing parallel problems.

Typically one wants to apply a callable to a list of arguments in parallel rather then sequentially, and then combine together the results. This is known as a MapReduce problem. As a simple example, we will consider the problem of counting the letters in a text. Here is how you can solve the problem sequentially:

>>> from itertools import starmap  # map a function with multiple arguments
>>> from functools import reduce  # reduce an iterable with a binary operator
>>> from operator import add  # addition function
>>> from collections import Counter  # callable doing the counting
>>> arglist = [('hello',), ('world',)]  # list of arguments
>>> results = starmap(Counter, arglist)  # iterator over the results
>>> res = reduce(add, results, Counter())  # aggregated counts
>>> sorted(res.items())  # counts per letter
[('d', 1), ('e', 1), ('h', 1), ('l', 3), ('o', 2), ('r', 1), ('w', 1)]

Here is how you can solve the problem in parallel by using openquake.baselib.parallel.Starmap:

>>> res2 = Starmap(Counter, arglist).reduce()
>>> assert res2 == res  # the same as before

As you see there are some notational advantages with respect to use itertools.starmap. First of all, Starmap has a reduce method, so there is no need to import functools.reduce; secondly, the reduce method has sensible defaults:

  1. the default aggregation function is add, so there is no need to specify it
  2. the default accumulator is an empty accumulation dictionary (see openquake.baselib.AccumDict) working as a Counter, so there is no need to specify it.

You can of course ovverride the defaults, so if you really want to return a Counter you can do

>>> res3 = Starmap(Counter, arglist).reduce(acc=Counter())

In the engine we use nearly always callables that return dictionaries and we aggregate nearly always with the addition operator, so such defaults are very convenient. You are encouraged to do the same, since we found that approach to be very flexible. Typically in a scientific application you will return a dictionary of numpy arrays.

The parallelization algorithm used by Starmap will depend on the environment variable OQ_DISTRIBUTE. Here are the possibilities available at the moment:

OQ_DISTRIBUTE not set or set to “futures”:
use multiprocessing via the concurrent.futures interface
OQ_DISTRIBUTE set to “no”:
disable the parallelization, useful for debugging
OQ_DISTRIBUTE set to “celery”:
use celery, useful if you have multiple machines in a cluster
OQ_DISTRIBUTE set tp “ipython”
use the ipyparallel concurrency mechanism (experimental)

There is also an `OQ_DISTRIBUTE`=”threadpool”; however the performance of using threads instead of processes is normally bad for the kind of applications we are interested in (CPU-dominated, which large tasks such that the time to spawn a new process is negligible with respect to the time to perform the task), so it is not recommended.

The Starmap.apply API

The Starmap class has a very convenient classmethod Starmap.apply which is used in several places in the engine. Starmap.apply is useful when you have a sequence of objects that you want to split in homogenous chunks and then apply a callable to each chunk (in parallel). For instance, in the letter counting example discussed before, Starmap.apply could be used as follows:

>>> text = 'helloworld'  # sequence of characters
>>> res3 = Starmap.apply(Counter, (text,)).reduce()
>>> assert res3 == res

The API of Starmap.apply is designed to extend the one of apply, a builtin of Python 2; the second argument is the tuple of arguments passed to the first argument. The difference with apply is that Starmap.apply returns a Starmap object so that nothing is actually done until you iterate on it (reduce is doing that).

How many chunks will be produced? That depends on the parameter concurrent_tasks; it it is not passed, it has a default of 5 times the number of cores in your machine - as returned by os.cpu_count() - and Starmap.apply will try to produce a number of chunks close to that number. The nice thing is that it is also possible to pass a weight function. Suppose for instance that instead of a list of letters you have a list of seismic sources: some sources requires a long computation time (such as ComplexFaultSources), some requires a short computation time (such as PointSources). By giving an heuristic weight to the different sources it is possible to produce chunks with nearly homogeneous weight; in particular PointSource tasks will contain a lot more sources than tasks with ComplexFaultSources.

It is essential in large computations to have a homogeneous task distribution, otherwise you will end up having a big task dominating the computation time (i.e. you may have 1000 cores of which 999 are free, having finished all the short tasks, but you have to wait for days for the single core processing the slow task). The OpenQuake engine does a great deal of work trying to split slow sources in more manageable fast sources.

class openquake.baselib.parallel.BaseStarmap(func, iterargs, poolsize=None, progress=<function info>)[source]

Bases: object

add_task_no(iterargs, pickle=True)

Add .task_no and .weight to the monitor and yield back the arguments by pickling them if pickle is True.

classmethod apply(func, args, concurrent_tasks=80, weight=<function <lambda>>, key=<function <lambda>>)[source]

The number of tasks, if known, or the empty string otherwise.

static poolfactory(size)
reduce(agg=<built-in function add>, acc=None)[source]
Returns:an IterResult instance
class openquake.baselib.parallel.IterResult(futures, taskname, num_tasks, progress=<function info>, sent=0)[source]

Bases: object

  • futures – an iterator over futures
  • taskname – the name of the task
  • num_tasks – the total number of expected futures
  • progress – a logging function for the progress report
  • sent – the number of bytes sent (0 if OQ_DISTRIBUTE=no)
reduce(agg=<built-in function add>, acc=None)[source]
classmethod sum(iresults)[source]

Sum the data transfer information of a set of results

task_data_dt = dtype([('taskno', '<u4'), ('weight', '<f4'), ('duration', '<f4')])
class openquake.baselib.parallel.Pickled(obj)[source]

Bases: object

An utility to manually pickling/unpickling objects. The reason is that celery does not use the HIGHEST_PROTOCOL, so relying on celery is slower. Moreover Pickled instances have a nice string representation and length giving the size of the pickled bytestring.

Parameters:obj – the object to pickle

Unpickle the underlying object

class openquake.baselib.parallel.Processmap(func, iterargs, poolsize=None, progress=<function info>)[source]

Bases: openquake.baselib.parallel.BaseStarmap

MapReduce implementation based on processes. For instance

>>> from collections import Counter
>>> c = Processmap(Counter, [('hello',), ('world',)], poolsize=4).reduce()
>>> sorted(c.items())
[('d', 1), ('e', 1), ('h', 1), ('l', 3), ('o', 2), ('r', 1), ('w', 1)]
class openquake.baselib.parallel.Sequential(func, iterargs, poolsize=None, progress=<function info>)[source]

Bases: openquake.baselib.parallel.BaseStarmap

A sequential Starmap, useful for debugging purpose.

class openquake.baselib.parallel.Starmap(oqtask, task_args, name=None)[source]

Bases: object

A manager to submit several tasks of the same type. The usage is:

tm = Starmap(do_something,
tm.send(arg1, arg2)
tm.send(arg3, arg4)

Progress report is built-in.

add_task_no(iterargs, pickle=True)[source]

Add .task_no and .weight to the monitor and yield back the arguments by pickling them if pickle is True.

classmethod apply(task, task_args, concurrent_tasks=80, maxweight=None, weight=<function <lambda>>, key=<function <lambda>>, name=None)[source]

Apply a task to a tuple of the form (sequence, *other_args) by first splitting the sequence in chunks, according to the weight of the elements and possibly to a key (see :func: openquake.baselib.general.split_in_blocks).

  • task – a task to run in parallel
  • task_args – the arguments to be passed to the task function
  • agg – the aggregation function
  • acc – initial value of the accumulator (default empty AccumDict)
  • concurrent_tasks – hint about how many tasks to generate
  • maxweight – if not None, used to split the tasks
  • weight – function to extract the weight of an item in arg0
  • key – function to extract the kind of an item in arg0
executor = <concurrent.futures.process.ProcessPoolExecutor object>

The number of tasks, if known, or the empty string otherwise.


Log in INFO mode regular tasks and in DEBUG private tasks

reduce(agg=<built-in function add>, acc=None)[source]

Loop on a set of results and update the accumulator by using the aggregation function.

  • agg – the aggregation function, (acc, val) -> new acc
  • acc – the initial value of the accumulator

the final value of the accumulator

classmethod restart()[source]

Submit a function with the given arguments to the process pool and add a Future to the list .results. If the attribute distribute is set, the function is run in process and the result is returned.

Returns:an IterResult object
task_ids = []

Wait until all the task terminate. Discard the results.

Returns:the total number of tasks that were spawned
class openquake.baselib.parallel.Threadmap(func, iterargs, poolsize=None, progress=<function info>)[source]

Bases: openquake.baselib.parallel.BaseStarmap

MapReduce implementation based on threads. For instance

>>> from collections import Counter
>>> c = Threadmap(Counter, [('hello',), ('world',)], poolsize=4).reduce()
>>> sorted(c.items())
[('d', 1), ('e', 1), ('h', 1), ('l', 3), ('o', 2), ('r', 1), ('w', 1)]
static poolfactory(size)
openquake.baselib.parallel.check_mem_usage(monitor=<Monitor dummy>, soft_percent=90, hard_percent=100)[source]

Display a warning if we are running out of memory

Parameters:mem_percent (int) – the memory limit as a percentage
openquake.baselib.parallel.do_not_aggregate(acc, value)[source]

Do nothing aggregation function.

  • acc – the accumulator
  • value – the value to accumulate

the accumulator unchanged


Return the pickled sizes of an object and its direct attributes, ordered by decreasing size. Here is an example:

>> total_size, partial_sizes = get_pickled_sizes(Monitor(‘’)) >> total_size 345 >> partial_sizes [(‘_procs’, 214), (‘exc’, 4), (‘mem’, 4), (‘start_time’, 4), (‘_start_time’, 4), (‘duration’, 4)]

Notice that the sizes depend on the operating system and the machine.


If the task has an attribute shared_dir_on which is false, return ‘futures’ even if OQ_DISTRIBUTE is celery, otherwise return the current value of the variable OQ_DISTRIBUTE; if undefined, return ‘futures’.


Convert an iterable of objects into a list of pickled objects. If the iterable contains copies, the pickling will be done only once. If the iterable contains objects already pickled, they will not be pickled again.

Parameters:objects – a sequence of objects to pickle
openquake.baselib.parallel.qsub(func, allargs, authkey=None)[source]

Map functions to arguments by means of the Grid Engine.

  • func – a pickleable callable object
  • allargs – a list of tuples of arguments
  • authkey – authentication token used to send back the results

an iterable over results of the form (res, etype, mon)

openquake.baselib.parallel.safely_call(func, args)[source]

Call the given function with the given arguments safely, i.e. by trapping the exceptions. Return a pair (result, exc_type) where exc_type is None if no exceptions occur, otherwise it is the exception class and the result is a string containing error message and traceback.

  • func – the function to call
  • args – the arguments

This is used at startup, only when the ProcessPoolExecutor is used, to fork the processes before loading any big data structure. It is called once once, and adds the list of PIDs spawned to the executor.


class openquake.baselib.performance.Monitor(operation='dummy', hdf5path=None, autoflush=False, measuremem=False)[source]

Bases: object

Measure the resident memory occupied by a list of processes during the execution of a block of code. Should be used as a context manager, as follows:

with Monitor('do_something') as mon:
print mon.mem

At the end of the block the Monitor object will have the following 5 public attributes:

.start_time: when the monitor started (a datetime object) .duration: time elapsed between start and stop (in seconds) .exc: usually None; otherwise the exception happened in the with block .mem: the memory delta in bytes

The behaviour of the Monitor can be customized by subclassing it and by overriding the method on_exit(), called at end and used to display or store the results of the analysis.

NB: if the .address attribute is set, it is possible for the monitor to send commands to that address, assuming there is a multiprocessing.connection.Listener listening.

address = None
authkey = None
calc_id = None

Last time interval measured


Save the measurements on the performance file (or on stdout)

Returns:an array of dtype perf_dt, with the information of the monitor (operation, time_sec, memory_mb, counts); the lenght of the array can be 0 (for counts=0) or 1 (otherwise).

A memory measurement (in bytes)

new(operation='no operation', **kw)[source]

Return a copy of the monitor usable for a different operation.


To be overridden in subclasses


Save (name, value) information in the associated hdf5path


Datetime instance recording when the monitoring started



Compatibility layer for Python 2 and 3. Mostly copied from six and future, but reduced to the subset of utilities needed by GEM. This is done to avoid an external dependency.


Recursively check all modules in the given package for compatibility with Python 3 syntax. No imports are performed.

Parameters:pkg – a Python package

Decode an object assuming the encoding is UTF-8.

Param:a unicode or bytes object
Returns:a unicode object

Encode a string assuming the encoding is UTF-8.

Param:a unicode or bytes object
openquake.baselib.python3compat.exec_(_code_, _globs_=None, _locs_=None)[source]

Execute code in a namespace.

openquake.baselib.python3compat.raise_(tp, value=None, tb=None)[source]
openquake.baselib.python3compat.with_metaclass(meta, *bases)[source]

Returns an instance of meta inheriting from the given bases. To be used to replace the __metaclass__ syntax., *args)[source]


class openquake.baselib.runtests.TestLoader[source]

Bases: object

loadTestsFromNames(suitename, module=None)[source]
class openquake.baselib.runtests.TestResult(stream, descriptions, verbosity)[source]

Bases: unittest.runner.TextTestResult

timedict = {}
openquake.baselib.runtests.addTest(self, test)[source]


Here is a minimal example of usage:

>>> from openquake.baselib import sap
>>> def fun(input, inplace, output=None, out='/tmp'):
...     'Example'
...     for item in sorted(locals().items()):
...         print('%s = %s' % item)

>>> p = sap.Script(fun)
>>> p.arg('input', 'input file or archive')
>>> p.flg('inplace', 'convert inplace')
>>> p.arg('output', 'output archive')
>>> p.opt('out', 'optional output file')

>>> p.callfunc(['a'])
inplace = False
input = a
out = /tmp
output = None

>>> p.callfunc(['a', 'b', '-i', '-o', 'OUT'])
inplace = True
input = a
out = OUT
output = b

Parsers can be composed too.

class, name=None, parentparser=None, help=True, registry=True)[source]

Bases: object

A simple way to define command processors based on argparse. Each parser is associated to a function and parsers can be composed together, by dispatching on a given name (if not given, the function name is used).

arg(name, help, type=None, choices=None, metavar=None, nargs=None)[source]

Describe a positional argument


Parse the argv list and extract a dictionary of arguments which is then passed to the function underlying the Script.


Make sure all arguments have a specification

flg(name, help, abbrev=None)[source]

Describe a flag


Added a new group of arguments with the given description


Return the help message as a string

opt(name, help, abbrev=None, type=None, choices=None, metavar=None, nargs=None)[source]

Describe an option

registry = {}, name='main', description=None, prog=None, version=None)[source]

Collects together different Scripts and builds a single Script dispatching to the subparsers depending on the first argument, i.e. the name of the subparser to invoke.

  • scripts – a list of Script instances
  • name – the name of the composed parser
  • description – description of the composed parser
  • prog – name of the script printed in the usage message
  • version – version of the script printed with –version, description=None, help=True)[source]
  • parserargparse.ArgumentParser instance or None
  • description – string used to build a new parser if parser is None
  • help – flag used to build a new parser if parser is None

if parser is None the new parser; otherwise the .parentparser attribute (if set) or the parser itself (if not set)[source]

Returns {choice1, ..., choiceN} or the empty string



Decorator for a class with _slots_. It automatically defines the methods __eq__, __ne__, assert_equal.