Source code for openquake.baselib.parallel

# -*- coding: utf-8 -*-
# vim: tabstop=4 shiftwidth=4 softtabstop=4
#
# Copyright (C) 2010-2023 GEM Foundation
#
# OpenQuake is free software: you can redistribute it and/or modify it
# under the terms of the GNU Affero General Public License as published
# by the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# OpenQuake is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU Affero General Public License for more details.
#
# You should have received a copy of the GNU Affero General Public License
# along with OpenQuake. If not, see <http://www.gnu.org/licenses/>.
"""\
The Starmap API
====================================

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

Typically one wants to apply a callable to a list of arguments in
parallel, 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, by using the following `count`
function:

.. code-block:: python

  def count(word):
      return collections.Counter(word)

The `collections.Counter` class works sequentially, and can
solve the problem in parallel by using
:class:`openquake.baselib.parallel.Starmap`:

>>> arglist = [('hello',), ('world',)]  # list of arguments
>>> smap = Starmap(count, arglist)  # Starmap instance, nothing started yet
>>> sorted(smap.reduce().items())  # build the counts per letter
[('d', 1), ('e', 1), ('h', 1), ('l', 3), ('o', 2), ('r', 1), ('w', 1)]

A `Starmap` object is an iterable: when iterating over it produces
task results. It also has a `reduce` method similar to `functools.reduce`
with 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
   :class:`openquake.baselib.AccumDict`) working as a `Counter`, so there
   is no need to specify it.

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

>>> res = Starmap(count, arglist).reduce(acc=collections.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 "processpool":
  use multiprocessing
`OQ_DISTRIBUTE` set to "no":
  disable the parallelization, useful for debugging
`OQ_DISTRIBUTE` set tp "zmq"
   use the zmq 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.

If you are using a pool, is always a good idea to cleanup resources at the end
with

>>> Starmap.shutdown()

`Starmap.shutdown` is always defined. It does nothing if there is
no pool, but it is still better to call it: in the future, you may change
idea and use another parallelization strategy requiring cleanup. In this
way your code is future-proof.

Monitoring
=============================

A major feature of the Starmap API is the ability to monitor the time spent
in each task and the memory allocated. Such information is written into an
HDF5 file that can be provided by the user or autogenerated. To autogenerate
the file you can use :func:`openquake.commonlib.datastore.hdf5new` which
will create a file named ``calc_XXX.hdf5`` in your $OQ_DATA directory
(if the environment variable is not set, the engine will use $HOME/oqdata).
Here is an example of usage:

>>> from openquake.commonlib.datastore import hdf5new
>>> h5 = hdf5new()
>>> smap = Starmap(count, [['hello'], ['world']], h5=h5)
>>> print(sorted(smap.reduce().items()))
[('d', 1), ('e', 1), ('h', 1), ('l', 3), ('o', 2), ('r', 1), ('w', 1)]

After the calculation, or even while the calculation is running, you can
open the calculation file for reading and extract the performance information
for it. The engine provides a command to do that, `oq show performance`,
but you can also get it manually, with a call to
`openquake.baselib.performance.performance_view(h5)` which will return
the performance information as a numpy array:

>>> from openquake.baselib.performance import performance_view
>>> performance_view(h5).dtype.names
('operation', 'time_sec', 'memory_mb', 'counts')
>>> h5.close()

The four columns are as follows:

operation:
  the name of the function running in parallel (in this case 'count')
time_sec:
  the cumulative time in second spent running the function
memory_mb:
  the maximum allocated memory per core
counts:
  the number of times the function was called (in this case 2)

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(count, (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 :class:`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.

"""
import os
import re
import ast
import sys
import stat
import time
import socket
import signal
import pickle
import inspect
import logging
import operator
import traceback
import subprocess
import collections
from unittest import mock
import multiprocessing.dummy
from multiprocessing.connection import wait
import multiprocessing.shared_memory as shmem
import psutil
import numpy

from openquake.baselib import config, hdf5
from openquake.baselib.python3compat import decode
from openquake.baselib.zeromq import zmq, Socket
from openquake.baselib.performance import (
    Monitor, memory_rss, init_performance)
from openquake.baselib.general import (
    split_in_blocks, block_splitter, AccumDict, humansize, CallableDict,
    gettemp, engine_version, shortlist, compress, decompress, mp as mp_context)

sys.setrecursionlimit(2000)  # raised to make pickle happier
# see https://github.com/gem/oq-engine/issues/5230
submit = CallableDict()
MB = 1024 ** 2
GB = 1024 ** 3
host_cores = config.zworkers.host_cores.split(',')

# see https://scicomp.aalto.fi/triton/tut/array
SLURM_BATCH = '''\
#!/bin/bash
#SBATCH --job-name={mon.operation}
#SBATCH --array=1-{mon.task_no}
#SBATCH --time=10:00:00
#SBATCH --mem-per-cpu=1G
#SBATCH --output={mon.calc_dir}/%a.out
#SBATCH --error={mon.calc_dir}/%a.err
srun {python} -m openquake.baselib.slurm {mon.calc_dir} $SLURM_ARRAY_TASK_ID
'''

[docs]def sbatch(mon): """ Start a SLURM script via sbatch """ path = os.path.join(mon.calc_dir, 'slurm.sh') with open(path, 'w') as f: python = config.distribution.python f.write(SLURM_BATCH.format(python=python, mon=mon)) os.chmod(path, os.stat(path).st_mode | stat.S_IEXEC) sbatch = subprocess.run(['which', 'sbatch'], capture_output=True).stdout if sbatch: subprocess.run(['sbatch', path], capture_output=True) return # TODO: return the SLURM job ID # if SLURM is not installed, fake it logging.info(f'Faking SLURM for {mon.operation}') pool = mp_context.Pool() for task_id in range(1, mon.task_no + 1): pool.apply_async(slurm_task, (mon.calc_dir, str(task_id))) pool.close() pool.join()
[docs]@submit.add('no') def no_submit(self, func, args, monitor): safely_call(func, args, self.task_no, monitor)
[docs]@submit.add('processpool') def processpool_submit(self, func, args, monitor): self.pool.apply_async(safely_call, (func, args, self.task_no, monitor))
[docs]@submit.add('threadpool') def threadpool_submit(self, func, args, monitor): self.pool.apply_async(safely_call, (func, args, self.task_no, monitor))
[docs]@submit.add('zmq') def zmq_submit(self, func, args, monitor): idx = self.task_no % len(host_cores) host = host_cores[idx].split()[0] port = int(config.zworkers.ctrl_port) dest = 'tcp://%s:%d' % (host, port) with Socket(dest, zmq.REQ, 'connect', timeout=300) as sock: sub = sock.send((func, args, self.task_no, monitor)) assert sub == 'submitted', sub
[docs]@submit.add('ipp') def ipp_submit(self, func, args, monitor): self.executor.submit(safely_call, func, args, self.task_no, monitor)
[docs]@submit.add('slurm') def slurm_submit(self, func, args, monitor): calc_dir = monitor.calc_dir # $HOME/oqdata/calc_XXX if not os.path.exists(calc_dir): os.mkdir(calc_dir) inpname = str(self.task_no + 1) + '.inp' with open(os.path.join(calc_dir, inpname), 'wb') as f: pickle.dump((func, args, monitor), f, pickle.HIGHEST_PROTOCOL) logging.debug('saved %s', os.path.join(calc_dir, inpname))
[docs]def oq_distribute(task=None): """ :returns: the value of OQ_DISTRIBUTE or config.distribution.oq_distribute """ dist = os.environ.get('OQ_DISTRIBUTE', config.distribution.oq_distribute) if dist not in ('no', 'processpool', 'threadpool', 'zmq', 'ipp', 'slurm'): raise ValueError('Invalid oq_distribute=%s' % dist) return dist
[docs]def init_workers(): """Used to initialize the process pool""" try: from setproctitle import setproctitle except ImportError: pass else: setproctitle('oq-worker')
[docs]class Pickled(object): """ An utility to manually pickling/unpickling objects. Pickled instances have a nice string representation and length giving the size of the pickled bytestring. :param obj: the object to pickle """ compressed = False def __init__(self, obj): self.clsname = obj.__class__.__name__ self.calc_id = str(getattr(obj, 'calc_id', '')) # for monitors try: self.pik = pickle.dumps(obj, pickle.HIGHEST_PROTOCOL) except TypeError as exc: # can't pickle, show the obj in the message raise TypeError('%s: %s' % (exc, obj)) self.compressed = len(self.pik) > MB and config.distribution.compress if self.compressed: self.pik = compress(self.pik) def __repr__(self): """String representation of the pickled object""" return '<Pickled %s #%s %s>' % ( self.clsname, self.calc_id, humansize(len(self))) def __len__(self): """Length of the pickled bytestring""" return len(self.pik)
[docs] def unpickle(self): """Unpickle the underlying object""" pik = decompress(self.pik) if self.compressed else self.pik return pickle.loads(pik)
[docs]def get_pickled_sizes(obj): """ 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. """ sizes = [] attrs = getattr(obj, '__dict__', {}) for name, value in attrs.items(): sizes.append((name, len(Pickled(value)))) return len(Pickled(obj)), sorted( sizes, key=lambda pair: pair[1], reverse=True)
[docs]def pickle_sequence(objects): """ 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. :param objects: a sequence of objects to pickle """ cache = {} out = [] for obj in objects: obj_id = id(obj) if obj_id not in cache: if isinstance(obj, Pickled): # already pickled cache[obj_id] = obj else: # pickle the object cache[obj_id] = Pickled(obj) out.append(cache[obj_id]) return out
[docs]class FakePickle: def __init__(self, sentbytes): self.sentbytes = sentbytes
[docs] def unpickle(self): pass
def __len__(self): return self.sentbytes
[docs]class Result(object): """ :param val: value to return or exception instance :param mon: Monitor instance :param tb_str: traceback string (empty if there was no exception) :param msg: message string (default empty) """ func = None def __init__(self, val, mon, tb_str='', msg=''): if isinstance(val, dict): self.pik = Pickled(val) self.nbytes = {k: len(Pickled(v)) for k, v in val.items()} elif isinstance(val, tuple) and callable(val[0]): self.func = val[0] self.pik = pickle_sequence(val[1:]) self.nbytes = {'args': sum(len(p) for p in self.pik)} elif msg == 'TASK_ENDED': self.pik = Pickled(None) self.nbytes = {} else: self.pik = Pickled(val) self.nbytes = {'tot': len(self.pik)} self.mon = mon self.tb_str = tb_str self.msg = msg self.workerid = (socket.gethostname(), os.getpid())
[docs] def get(self): """ Returns the underlying value or raise the underlying exception """ t0 = time.time() val = self.pik.unpickle() self.dt = time.time() - t0 if self.tb_str: etype = val.__class__ msg = '\n%s%s: %s' % (self.tb_str, etype.__name__, val) if issubclass(etype, KeyError): raise RuntimeError(msg) # nicer message else: raise etype(msg) return val
def __repr__(self): nbytes = ['%s: %s' % (k, humansize(v)) for k, v in self.nbytes.items()] return '<%s %s>' % (self.__class__.__name__, ' '.join(nbytes))
[docs] @classmethod def new(cls, func, args, mon, sentbytes=0): """ :returns: a new Result instance """ try: if mon.version and mon.version != engine_version(): raise RuntimeError( 'The master is at version %s while the worker %s is at ' 'version %s' % (mon.version, socket.gethostname(), engine_version())) if mon.config.dbserver.host != config.dbserver.host: raise RuntimeError( 'The master has dbserver.host=%s while the worker has %s' % (mon.config.dbserver.host, config.dbserver.host)) with mon: val = func(*args) except StopIteration: mon.counts -= 1 # StopIteration does not count res = Result(None, mon, msg='TASK_ENDED') res.pik = FakePickle(sentbytes) except Exception: _etype, exc, tb = sys.exc_info() res = Result(exc, mon, ''.join(traceback.format_tb(tb))) else: res = Result(val, mon) return res
[docs]def check_mem_usage(soft_percent=None, hard_percent=None): """ Display a warning if we are running out of memory """ soft_percent = soft_percent or config.memory.soft_mem_limit hard_percent = hard_percent or config.memory.hard_mem_limit used_mem_percent = psutil.virtual_memory().percent if used_mem_percent > hard_percent: raise MemoryError('Using more memory than allowed by configuration ' '(Used: %d%% / Allowed: %d%%)! Shutting down.' % (used_mem_percent, hard_percent)) elif used_mem_percent > soft_percent: msg = 'Using over %d%% of the memory in %s!' return msg % (used_mem_percent, socket.gethostname())
dummy_mon = Monitor() dummy_mon.config = config dummy_mon.backurl = None DEBUG = False
[docs]def sendback(res, zsocket): """ Send back to the master node the result by using the zsocket. :returns: the accumulated number of bytes sent """ calc_id = res.mon.calc_id task_no = res.mon.task_no nbytes = len(res.pik) try: zsocket.send(res) if DEBUG: from openquake.commonlib.logs import dblog if calc_id: # None when building the png maps msg = 'sent back %s' % humansize(nbytes) dblog('DEBUG', calc_id, task_no, msg) except Exception: # like OverflowError _etype, exc, tb = sys.exc_info() tb_str = ''.join(traceback.format_tb(tb)) if DEBUG and calc_id: dblog('ERROR', calc_id, task_no, tb_str) res = Result(exc, res.mon, tb_str) zsocket.send(res) return nbytes
[docs]def safely_call(func, args, task_no=0, mon=dummy_mon): """ 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. :param func: the function to call :param args: the arguments :param task_no: the task number :param mon: a monitor """ isgenfunc = inspect.isgeneratorfunction(func) if hasattr(args[0], 'unpickle'): # args is a list of Pickled objects args = [a.unpickle() for a in args] if mon is dummy_mon: # in the DbServer assert not isgenfunc, func return Result.new(func, args, mon) # debug(f'{mon.backurl=}, {task_no=}') if mon.operation.endswith('_'): name = mon.operation[:-1] elif func is split_task: name = args[1].__name__ else: name = func.__name__ mon = mon.new(operation='total ' + name, measuremem=True) mon.weight = getattr(args[0], 'weight', 1.) # used in task_info mon.task_no = task_no if mon.inject: args += (mon,) sentbytes = 0 if isgenfunc: with Socket(mon.backurl, zmq.PUSH, 'connect') as zsocket: it = func(*args) while True: res = Result.new(next, (it,), mon, sentbytes) # StopIteration -> TASK_ENDED if res.msg == 'TASK_ENDED': zsocket.send(res) break sentbytes += sendback(res, zsocket) else: res = Result.new(func, args, mon) # send back a single result and a TASK_ENDED with Socket(mon.backurl, zmq.PUSH, 'connect') as zsocket: sentbytes += sendback(res, zsocket) end = Result(None, mon, msg='TASK_ENDED') end.pik = FakePickle(sentbytes) zsocket.send(end)
if oq_distribute() == 'ipp': from ipyparallel import Cluster
[docs]class IterResult(object): """ :param iresults: an iterator over Result objects :param taskname: the name of the task :param done_total: a function returning the number of done tasks and the total :param sent: a nested dictionary name -> {argname: number of bytes sent} :param progress: a logging function for the progress report :param hdf5path: a path where to store persistently the performance info """ def __init__(self, iresults, taskname, argnames, sent, h5): self.iresults = iresults self.name = taskname self.argnames = ' '.join(argnames) self.sent = sent self.h5 = h5 def _iter(self): first_time = True self.counts = 0 self.dt = 0 for result in self.iresults: msg = check_mem_usage() # log a warning if too much memory is used if msg and first_time: logging.warning(msg) first_time = False # warn only once self.nbytes += result.nbytes self.counts += 1 out = result.get() self.dt += result.dt yield out def __iter__(self): if self.iresults == (): return () t0 = time.time() self.nbytes = AccumDict() try: yield from self._iter() finally: items = sorted(self.nbytes.items(), key=operator.itemgetter(1)) nb = {k: humansize(v) for k, v in list(reversed(items))[:3]} recv = sum(self.nbytes.values()) mean = recv / (self.counts or 1) pu = '[unpik=%.2fs]' % self.dt logging.info('Received %d * %s in %d seconds %s from %s\n%s', self.counts, humansize(mean), time.time() - t0, pu, self.name, nb)
[docs] def reduce(self, agg=operator.add, acc=None): if acc is None: acc = AccumDict() for result in self: acc = agg(acc, result) return acc
[docs] @classmethod def sum(cls, iresults): """ Sum the data transfer information of a set of results """ res = object.__new__(cls) res.sent = 0 for iresult in iresults: res.sent += iresult.sent name = iresult.name.split('#', 1)[0] if hasattr(res, 'name'): assert res.name.split('#', 1)[0] == name, (res.name, name) else: res.name = iresult.name.split('#')[0] return res
[docs]def getargnames(task_func): # a task can be a function, a method, a class or a callable instance if inspect.isfunction(task_func): return inspect.getfullargspec(task_func).args elif inspect.ismethod(task_func): return inspect.getfullargspec(task_func).args[1:] elif inspect.isclass(task_func): return inspect.getfullargspec(task_func.__init__).args[1:] else: # instance with a __call__ method return inspect.getfullargspec(task_func.__call__).args[1:]
[docs]def get_return_ip(receiver_host): if receiver_host: return socket.gethostbyname(receiver_host) hostname = socket.gethostname() return socket.gethostbyname(hostname)
# ########################### SharedMemory ############################## #
[docs]class SharedArray(object): """ Wrapper over a SharedMemory array to be used as a context manager. """
[docs] @classmethod def new(cls, array): return cls(array.shape, array.dtype, array)
def __init__(self, shape, dtype, value): nbytes = numpy.zeros(1, dtype).nbytes * numpy.prod(shape) # NOTE: on Windows size wants an int an not a numpy.int self.sm = shmem.SharedMemory(create=True, size=int(nbytes)) self.shape = shape self.dtype = dtype # fill the SharedMemory buffer with the value arr = numpy.ndarray(shape, dtype, buffer=self.sm.buf) arr[:] = value def __enter__(self): # this is called in the workers self._sm = shmem.SharedMemory(name=self.sm.name) return numpy.ndarray(self.shape, self.dtype, buffer=self._sm.buf) def __exit__(self, etype, exc, tb): # this is called in the workers self._sm.close()
[docs]class Starmap(object): pids = () running_tasks = [] # currently running tasks maxtasksperchild = None # with 1 it hangs on the EUR calculation! num_cores = int(config.distribution.get('num_cores', '0')) if not num_cores: # use only the "visible" cores, not the total system cores # if the underlying OS supports it (macOS does not) try: num_cores = len(psutil.Process().cpu_affinity()) except AttributeError: num_cores = psutil.cpu_count() CT = num_cores * 2
[docs] @classmethod def init(cls, distribute=None): cls.distribute = distribute or oq_distribute() if cls.distribute == 'processpool' and not hasattr(cls, 'pool'): # unregister custom handlers before starting the processpool term_handler = signal.signal(signal.SIGTERM, signal.SIG_DFL) int_handler = signal.signal(signal.SIGINT, signal.SIG_IGN) # we use spawn here to avoid deadlocks with logging, see # https://github.com/gem/oq-engine/pull/3923 and # https://codewithoutrules.com/2018/09/04/python-multiprocessing/ cls.pool = mp_context.Pool( cls.num_cores, init_workers, maxtasksperchild=cls.maxtasksperchild) cls.pids = [proc.pid for proc in cls.pool._pool] # after spawning the processes restore the original handlers # i.e. the ones defined in openquake.engine.engine signal.signal(signal.SIGTERM, term_handler) signal.signal(signal.SIGINT, int_handler) elif cls.distribute == 'threadpool' and not hasattr(cls, 'pool'): cls.pool = multiprocessing.dummy.Pool(cls.num_cores) elif cls.distribute == 'ipp' and not hasattr(cls, 'executor'): rc = Cluster(n=cls.num_cores).start_and_connect_sync() cls.executor = rc.executor()
[docs] @classmethod def shutdown(cls): # shutting down the pool during the runtime causes mysterious # race conditions with errors inside atexit._run_exitfuncs if hasattr(cls, 'pool'): cls.pool.close() cls.pool.terminate() cls.pool.join() del cls.pool cls.pids = [] elif hasattr(cls, 'executor'): cls.executor.shutdown()
[docs] @classmethod def apply(cls, task, allargs, concurrent_tasks=None, maxweight=None, weight=lambda item: 1, key=lambda item: 'Unspecified', distribute=None, progress=logging.info, h5=None): r""" 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`). :param task: a task to run in parallel :param args: the arguments to be passed to the task function :param concurrent_tasks: hint about how many tasks to generate :param maxweight: if not None, used to split the tasks :param weight: function to extract the weight of an item in arg0 :param key: function to extract the kind of an item in arg0 :param distribute: if not given, inferred from OQ_DISTRIBUTE :param progress: logging function to use (default logging.info) :param h5: an open hdf5.File where to store the performance info :returns: an :class:`IterResult` object """ arg0, *args = allargs if maxweight: # block_splitter is lazy taskargs = ([blk] + args for blk in block_splitter( arg0, maxweight, weight, key)) else: # split_in_blocks is eager if concurrent_tasks is None: concurrent_tasks = cls.CT taskargs = [[blk] + args for blk in split_in_blocks( arg0, concurrent_tasks or 1, weight, key)] return cls(task, taskargs, distribute, progress, h5)
[docs] def apply_split(cls, task, allargs, concurrent_tasks=None, maxweight=None, weight=lambda item: 1, key=lambda item: 'Unspecified', distribute=None, progress=logging.info, h5=None, duration=300, outs_per_task=5): """ Same as Starmap.apply, but possibly produces subtasks """ args = (allargs[0], task, allargs[1:], duration, outs_per_task) return cls.apply(split_task, args, concurrent_tasks or 2*cls.num_cores, maxweight, weight, key, distribute, progress, h5)
def __init__(self, task_func, task_args=(), distribute=None, progress=logging.info, h5=None): self.__class__.init(distribute=distribute) self.task_func = task_func if h5: match = re.search(r'(\d+)', os.path.basename(h5.filename)) self.calc_id = int(match.group(1)) else: # TODO: see if we can forbid this case self.calc_id = None h5 = hdf5.File(gettemp(suffix='.hdf5'), 'w') init_performance(h5) if task_func is split_task: self.name = task_args[0][1].__name__ else: self.name = task_func.__name__ self.monitor = Monitor(self.name) self.monitor.filename = h5.filename self.monitor.calc_id = self.calc_id self.task_args = task_args self.progress = progress self.h5 = h5 self.task_queue = [] try: self.num_tasks = len(self.task_args) except TypeError: # generators have no len self.num_tasks = None self.argnames = getargnames(task_func) self.sent = AccumDict(accum=AccumDict()) # fname -> argname -> nbytes self.monitor.inject = (self.argnames[-1].startswith('mon') or self.argnames[-1].endswith('mon')) self.receiver = 'tcp://0.0.0.0:%s' % config.dbserver.receiver_ports if self.distribute in ('no', 'processpool') or sys.platform != 'linux': self.return_ip = '127.0.0.1' # zmq returns data to localhost else: # zmq returns data to the receiver_host self.return_ip = get_return_ip(config.dbserver.receiver_host) logging.debug(f'{self.return_ip=}') self.monitor.backurl = None # overridden later self.tasks = [] # populated by .submit self.task_no = 0 self._shared = {}
[docs] def log_percent(self): """ Log the progress of the computation in percentage """ queued = len(self.task_queue) total = self.task_no done = total - len(self.tasks) percent = int(float(done) / total * 100) if not hasattr(self, 'prev_percent'): # first time self.prev_percent = 0 elif percent > self.prev_percent: self.progress('%s %3d%% [%d submitted, %d queued]', self.name, percent, self.task_no, queued) self.prev_percent = percent return done
[docs] def submit(self, args, func=None): """ Submit the given arguments to the underlying task """ func = func or self.task_func if not hasattr(self, 'socket'): # setup the PULL socket the first time self.__class__.running_tasks = self.tasks self.socket = Socket(self.receiver, zmq.PULL, 'bind').__enter__() self.monitor.shared = self._shared self.monitor.backurl = 'tcp://%s:%s' % ( self.return_ip, self.socket.port) self.monitor.config = config OQ_TASK_NO = os.environ.get('OQ_TASK_NO') if OQ_TASK_NO is not None and self.task_no != int(OQ_TASK_NO): self.task_no += 1 return dist = 'no' if self.num_tasks == 1 or OQ_TASK_NO else self.distribute if dist != 'no': pickled = isinstance(args[0], Pickled) if not pickled: assert not isinstance(args[-1], Monitor) # sanity check args = pickle_sequence(args) if func is None: fname = self.task_func.__name__ argnames = self.argnames[:-1] else: fname = func.__name__ argnames = getargnames(func)[:-1] self.sent[fname] += {a: len(p) for a, p in zip(argnames, args)} submit[dist](self, func, args, self.monitor) self.tasks.append(self.task_no) self.task_no += 1
[docs] def submit_split(self, args, duration, outs_per_task): """ Submit the given arguments to the underlying task """ self.monitor.operation = self.task_func.__name__ + '_' self.submit( (args[0], self.task_func, args[1:], duration, outs_per_task), split_task)
[docs] def submit_all(self): """ :returns: an IterResult object """ if self.num_tasks is None: # loop on the iterator for args in self.task_args: self.submit(args) else: # build a task queue in advance self.task_queue = [(self.task_func, args) for args in self.task_args] dist = 'no' if self.num_tasks == 1 else self.distribute if dist == 'slurm': for func, args in self.task_queue: self.submit(args, func=func) self.task_queue.clear() return self.get_results()
[docs] def get_results(self): """ :returns: an :class:`IterResult` instance """ return IterResult(self._loop(), self.name, self.argnames, self.sent, self.h5)
[docs] def reduce(self, agg=operator.add, acc=None): """ Submit all tasks and reduce the results """ return self.submit_all().reduce(agg, acc)
def __iter__(self): return iter(self.submit_all()) def _submit_many(self, howmany): for _ in range(howmany): if self.task_queue: # remove in LIFO order func, args = self.task_queue[0] del self.task_queue[0] self.submit(args, func=func) # NB: the shared dictionary will be attached to the monitor # and used in the workers; to see an example of usage, look at # the event_based calculator
[docs] def share(self, **dictarray): """ Apply SharedArray.new to a dictionary of arrays """ self._shared = {k: SharedArray.new(a) for k, a in dictarray.items()}
def _loop(self): self.busytime = AccumDict(accum=[]) # pid -> time dist = 'no' if self.num_tasks == 1 else self.distribute if dist == 'slurm': self.monitor.task_no = self.task_no # total number of tasks sbatch(self.monitor) elif self.task_queue: first_args = self.task_queue[:self.CT] self.task_queue[:] = self.task_queue[self.CT:] for func, args in first_args: self.submit(args, func=func) if not hasattr(self, 'socket'): # no submit was ever made return () nbytes = sum(self.sent[self.task_func.__name__].values()) logging.warning('Sent %d %s tasks, %s', len(self.tasks), self.name, humansize(nbytes)) isocket = iter(self.socket) # read from the PULL socket finished = set() while self.tasks: self.log_percent() res = next(isocket) if self.calc_id != res.mon.calc_id: logging.warning('Discarding a result from job %s, since this ' 'is job %s', res.mon.calc_id, self.calc_id) elif res.msg == 'TASK_ENDED': finished.add(res.mon.task_no) self.busytime += {res.workerid: res.mon.duration} self.tasks.remove(res.mon.task_no) self._submit_many(1) todo = set(range(self.task_no)) - finished logging.debug('%d tasks todo %s', len(todo), shortlist(sorted(todo))) task_sent = ast.literal_eval(decode(self.h5['task_sent'][()])) task_sent.update(self.sent) del self.h5['task_sent'] self.h5['task_sent'] = str(task_sent) name = res.mon.operation[6:] # strip 'total ' n = self.name + ':' + name if name == 'split_task' else name if sys.platform != 'darwin': # it normally works on macOS, but not in notebooks calling # notebooks, which is the case relevant for Marco Pagani mem_gb = (memory_rss(os.getpid()) + sum( memory_rss(pid) for pid in Starmap.pids)) / GB else: # measure only the memory used by the main process mem_gb = memory_rss(os.getpid()) / GB res.mon.save_task_info(self.h5, res, n, mem_gb) res.mon.flush(self.h5) elif res.func: # add subtask self.task_queue.append((res.func, res.pik)) self._submit_many(1) else: yield res self.log_percent() self.socket.__exit__(None, None, None) self.tasks.clear() self.unlink() if dist == 'slurm': for fname in os.listdir(self.monitor.calc_dir): os.remove(os.path.join(self.monitor.calc_dir, fname)) if len(self.busytime) > 1: times = numpy.array(list(self.busytime.values())) logging.info( 'Mean time per core=%ds, std=%.1fs, min=%ds, max=%ds', times.mean(), times.std(), times.min(), times.max())
[docs]def sequential_apply(task, args, concurrent_tasks=Starmap.CT, maxweight=None, weight=lambda item: 1, key=lambda item: 'Unspecified', progress=logging.info): """ Apply sequentially task to args by splitting args[0] in blocks """ with mock.patch.dict('os.environ', {'OQ_DISTRIBUTE': 'no'}): return Starmap.apply(task, args, concurrent_tasks, maxweight, weight, key, progress=progress)
[docs]def count(word): """ Used as example in the documentation """ return collections.Counter(word)
[docs]class List(list): weight = 0
[docs]def split_task(elements, func, args, duration, outs_per_task, monitor): """ :param func: a task function with a monitor as last argument :param args: arguments of the task function, with args[0] being a sequence :param duration: split the task if it exceeds the duration :param outs_per_task: number of splits to try (ex. 5) :yields: a partial result, 0 or more task objects """ n = len(elements) if outs_per_task > n: # too many splits outs_per_task = n elements = numpy.array(elements) # from WeightedSequence to array idxs = numpy.arange(n) split_elems = [elements[idxs % outs_per_task == i] for i in range(outs_per_task)] # see how long it takes to run the first slice t0 = time.time() for i, elems in enumerate(split_elems): monitor.out_no = monitor.task_no + i * 65536 res = func(elems, *args, monitor=monitor) dt = time.time() - t0 yield res if dt > duration: # spawn subtasks for the rest and exit, used in classical/case_14 for els in split_elems[i + 1:]: ls = List(els) ls.weight = sum(getattr(el, 'weight', 1.) for el in els) yield (func, ls) + args break
[docs]def logfinish(n, tot): logging.info('Finished %d of %d jobs', n, tot) return n + 1
[docs]def multispawn(func, allargs, chunksize=Starmap.num_cores): """ Spawn processes with the given arguments """ tot = len(allargs) logging.info('Running %d jobs', tot) allargs = allargs[::-1] # so that the first argument is submitted first procs = {} # sentinel -> process n = 1 while allargs: args = allargs.pop() proc = mp_context.Process(target=func, args=args) proc.start() procs[proc.sentinel] = proc while len(procs) >= chunksize: # wait for something to finish for finished in wait(procs): procs[finished].join() del procs[finished] n = logfinish(n, tot) while procs: for finished in wait(procs): procs[finished].join() del procs[finished] n = logfinish(n, tot)
[docs]def slurm_task(calc_dir: str, task_id: str): """ Task in a SLURM job array """ with open(calc_dir + '/' + task_id + '.inp', 'rb') as f: func, args, mon = pickle.load(f) safely_call(func, args, int(task_id) - 1, mon)