westpa 2022.13__cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl

westpa/core/sim_manager.py

@@ -0,0 +1,839 @@
+import logging
+import math
+import operator
+import time
+from datetime import timedelta
+from pickle import PickleError
+from itertools import zip_longest
+from collections import Counter
+
+import numpy as np
+from numpy.random import Generator, MT19937
+
+import westpa
+from .data_manager import weight_dtype
+from .segment import Segment
+from .states import InitialState
+from . import extloader
+from . import wm_ops
+
+
+log = logging.getLogger(__name__)
+
+EPS = np.finfo(weight_dtype).eps
+
+
+def grouper(n, iterable, fillvalue=None):
+    "Collect data into fixed-length chunks or blocks"
+    # grouper(3, 'ABCDEFG', 'x') --> ABC DEF Gxx
+    args = [iter(iterable)] * n
+    return zip_longest(fillvalue=fillvalue, *args)
+
+
+class PropagationError(RuntimeError):
+    pass
+
+
+class WESimManager:
+    def process_config(self):
+        config = self.rc.config
+        for entry, type_ in [('gen_istates', bool), ('block_size', int), ('save_transition_matrices', bool)]:
+            config.require_type_if_present(['west', 'propagation', entry], type_)
+
+        self.do_gen_istates = config.get(['west', 'propagation', 'gen_istates'], False)
+        self.propagator_block_size = config.get(['west', 'propagation', 'block_size'], 1)
+        self.save_transition_matrices = config.get(['west', 'propagation', 'save_transition_matrices'], False)
+        self.max_run_walltime = config.get(['west', 'propagation', 'max_run_wallclock'], default=None)
+        self.max_total_iterations = config.get(['west', 'propagation', 'max_total_iterations'], default=None)
+
+    def __init__(self, rc=None):
+        self.rc = rc or westpa.rc
+        self.work_manager = self.rc.get_work_manager()
+        self.data_manager = self.rc.get_data_manager()
+        self.we_driver = self.rc.get_we_driver()
+        self.system = self.rc.get_system_driver()
+
+        # A table of function -> list of (priority, name, callback) tuples
+        self._callback_table = {}
+        self._valid_callbacks = {
+            self.prepare_run,
+            self.finalize_run,
+            self.prepare_iteration,
+            self.finalize_iteration,
+            self.pre_propagation,
+            self.post_propagation,
+            self.pre_we,
+            self.post_we,
+            self.prepare_new_iteration,
+        }
+        self._callbacks_by_name = {fn.__name__: fn for fn in self._valid_callbacks}
+        self.n_propagated = 0
+
+        # config items
+        self.do_gen_istates = False
+        self.propagator_block_size = 1
+        self.save_transition_matrices = False
+        self.max_run_walltime = None
+        self.max_total_iterations = None
+        self.process_config()
+
+        # Per-iteration variables
+        self.n_iter = None  # current iteration
+
+        # Basis and initial states for this iteration, in case the propagator needs them
+        self.current_iter_bstates = None  # BasisStates valid at this iteration
+        self.current_iter_istates = None  # InitialStates used in this iteration
+
+        # Basis states for next iteration
+        self.next_iter_bstates = None  # BasisStates valid for the next iteration
+        self.next_iter_bstate_cprobs = None  # Cumulative probabilities for basis states, used for selection
+
+        # Tracking of this iteration's segments
+        self.segments = None  # Mapping of seg_id to segment for all segments in this iteration
+        self.completed_segments = None  # Mapping of seg_id to segment for all completed segments in this iteration
+        self.incomplete_segments = None  # Mapping of seg_id to segment for all incomplete segments in this iteration
+
+        # Tracking of binning
+        self.bin_mapper_hash = None  # Hash of bin mapper from most recently-run WE, for use by post-WE analysis plugins
+
+        # Pseudo Random Number Generator
+        self.rng = Generator(MT19937())
+
+    def register_callback(self, hook, function, priority=0):
+        '''Registers a callback to execute during the given ``hook`` into the simulation loop. The optional
+        priority is used to order when the function is called relative to other registered callbacks.'''
+
+        if hook not in self._valid_callbacks:
+            try:
+                hook = self._callbacks_by_name[hook]
+            except KeyError:
+                raise KeyError('invalid hook {!r}'.format(hook))
+
+        # It's possible to register a callback that's a duplicate function, but at a different place in memory.
+        #   For example, if you launch a run without clearing state from a previous run.
+        #   More details on this are available in https://github.com/westpa/westpa/issues/182 but the below code
+        #   handles specifically the problem that causes in plugin loading.
+        try:
+            # Before checking for set membership of (priority, function.__name__, function), just check
+            #   function hash for collisions in this hook.
+            hook_function_hash = [hash(callback[2]) for callback in self._callback_table[hook]]
+        except KeyError:
+            # If there's no entry in self._callback_table for this hook, then there definitely aren't any collisions
+            #   because no plugins are registered to it yet in the first place.
+            pass
+        else:
+            # If there are plugins registered to this hook, check for duplicate hash, which will definitely have the same name, module, function.
+            try:
+                if hash(function) in hook_function_hash:
+                    log.info('{!r} has already been loaded, skipping'.format(function))
+                    return
+            except KeyError:
+                pass
+
+        try:
+            self._callback_table[hook].add((priority, function.__name__, function))
+        except KeyError:
+            self._callback_table[hook] = set([(priority, function.__name__, function)])
+
+        # Raise warning if there are multiple callback with same priority.
+        for priority, count in Counter([callback[0] for callback in self._callback_table[hook]]).items():
+            if count > 1:
+                log.warning(
+                    f'{count} callbacks in {hook} have identical priority {priority}. The order of callback execution is not guaranteed.'
+                )
+                log.warning(f'{hook}: {self._callback_table[hook]}')
+
+        log.debug('registered callback {!r} for hook {!r}'.format(function, hook))
+
+    def invoke_callbacks(self, hook, *args, **kwargs):
+        callbacks = self._callback_table.get(hook, [])
+        # Sort by priority, function name, then module name
+        sorted_callbacks = sorted(callbacks, key=lambda x: (x[0], x[1], x[2].__module__))
+        for priority, name, fn in sorted_callbacks:
+            log.debug('invoking callback {!r} for hook {!r}'.format(fn, hook))
+            fn(*args, **kwargs)
+
+    def load_plugins(self, plugins=None):
+        if plugins is None:
+            plugins = []
+
+        try:
+            plugins_config = westpa.rc.config['west', 'plugins']
+        except KeyError:
+            plugins_config = []
+
+        if plugins_config is None:
+            plugins_config = []
+
+        plugins += plugins_config
+
+        for plugin_config in plugins:
+            plugin_name = plugin_config['plugin']
+            if plugin_config.get('enabled', True):
+                log.info('loading plugin {!r}'.format(plugin_name))
+                plugin = extloader.get_object(plugin_name)(self, plugin_config)
+                log.debug('loaded plugin {!r}'.format(plugin))
+
+    def report_bin_statistics(self, bins, target_states, save_summary=False):
+        segments = list(self.segments.values())
+        bin_counts = np.fromiter(map(len, bins), dtype=np.int_, count=len(bins))
+        target_counts = self.we_driver.bin_target_counts
+
+        # Do not include bins with target count zero (e.g. sinks, never-filled bins) in the (non)empty bins statistics
+        n_active_bins = len(target_counts[target_counts != 0])
+
+        if target_states:
+            n_active_bins -= len(target_states)
+
+        seg_probs = np.fromiter(map(operator.attrgetter('weight'), segments), dtype=weight_dtype, count=len(segments))
+        bin_probs = np.fromiter(map(operator.attrgetter('weight'), bins), dtype=weight_dtype, count=len(bins))
+        norm = seg_probs.sum()
+
+        assert abs(1 - norm) < EPS * (len(segments) + n_active_bins)
+
+        min_seg_prob = seg_probs[seg_probs != 0].min()
+        max_seg_prob = seg_probs.max()
+        seg_drange = math.log(max_seg_prob / min_seg_prob)
+        min_bin_prob = bin_probs[bin_probs != 0].min()
+        max_bin_prob = bin_probs.max()
+        bin_drange = math.log(max_bin_prob / min_bin_prob)
+        n_pop = len(bin_counts[bin_counts != 0])
+
+        self.rc.pstatus('{:d} of {:d} ({:%}) active bins are populated'.format(n_pop, n_active_bins, n_pop / n_active_bins))
+        self.rc.pstatus('per-bin minimum non-zero probability:       {:g}'.format(min_bin_prob))
+        self.rc.pstatus('per-bin maximum probability:                {:g}'.format(max_bin_prob))
+        self.rc.pstatus('per-bin probability dynamic range (kT):     {:g}'.format(bin_drange))
+        self.rc.pstatus('per-segment minimum non-zero probability:   {:g}'.format(min_seg_prob))
+        self.rc.pstatus('per-segment maximum non-zero probability:   {:g}'.format(max_seg_prob))
+        self.rc.pstatus('per-segment probability dynamic range (kT): {:g}'.format(seg_drange))
+        self.rc.pstatus('norm = {:g}, error in norm = {:g} ({:.2g}*epsilon)'.format(norm, (norm - 1), (norm - 1) / EPS))
+        self.rc.pflush()
+
+        if min_seg_prob < 1e-100:
+            log.warning(
+                '\nMinimum segment weight is < 1e-100 and might not be physically relevant. Please reconsider your progress coordinate or binning scheme.'
+            )
+
+        if save_summary:
+            iter_summary = self.data_manager.get_iter_summary()
+            iter_summary['n_particles'] = len(segments)
+            iter_summary['norm'] = norm
+            iter_summary['min_bin_prob'] = min_bin_prob
+            iter_summary['max_bin_prob'] = max_bin_prob
+            iter_summary['min_seg_prob'] = min_seg_prob
+            iter_summary['max_seg_prob'] = max_seg_prob
+            if np.isnan(iter_summary['cputime']):
+                iter_summary['cputime'] = 0.0
+            if np.isnan(iter_summary['walltime']):
+                iter_summary['walltime'] = 0.0
+            self.data_manager.update_iter_summary(iter_summary)
+
+    def get_bstate_pcoords(self, basis_states, label='basis'):
+        '''For each of the given ``basis_states``, calculate progress coordinate values
+        as necessary.  The HDF5 file is not updated. The BasisState objects are explicitly
+        copied from the futures in order to retain auxdata/restart files (under BasisState.data)
+        from certain work managers (e.g., the ``processes`` work manager.)'''
+
+        self.rc.pstatus('Calculating progress coordinate values for {} states.'.format(label))
+        futures = [self.work_manager.submit(wm_ops.get_pcoord, args=(basis_state,)) for basis_state in basis_states]
+        fmap = {future: i for (i, future) in enumerate(futures)}
+        for future in self.work_manager.as_completed(futures):
+            basis_states[fmap[future]] = future.get_result()
+
+    def report_basis_states(self, basis_states, label='basis'):
+        pstatus = self.rc.pstatus
+        pstatus('{:d} {} state(s) present'.format(len(basis_states), label), end='')
+        if self.rc.verbose_mode:
+            pstatus(':')
+            pstatus(
+                '{:6s}    {:12s}    {:20s}    {:20s}    {}'.format(
+                    'ID', 'Label', 'Probability', 'Aux Reference', 'Progress Coordinate'
+                )
+            )
+            for basis_state in basis_states:
+                pstatus(
+                    '{:<6d}    {:12s}    {:<20.14g}    {:20s}    {}'.format(
+                        basis_state.state_id,
+                        basis_state.label,
+                        basis_state.probability,
+                        basis_state.auxref or '',
+                        ', '.join(map(str, basis_state.pcoord)),
+                    )
+                )
+        pstatus()
+        self.rc.pflush()
+
+    def report_target_states(self, target_states):
+        pstatus = self.rc.pstatus
+        pstatus('{:d} target state(s) present'.format(len(target_states)), end='')
+        if self.rc.verbose_mode and target_states:
+            pstatus(':')
+            pstatus('{:6s}    {:12s}    {}'.format('ID', 'Label', 'Progress Coordinate'))
+            for target_state in target_states:
+                pstatus(
+                    '{:<6d}    {:12s}    {}'.format(
+                        target_state.state_id, target_state.label, ','.join(map(str, target_state.pcoord))
+                    )
+                )
+        pstatus()
+        self.rc.pflush()
+
+    def initialize_simulation(self, basis_states, target_states, start_states, segs_per_state=1, suppress_we=False):
+        '''Initialize a new weighted ensemble simulation, taking ``segs_per_state`` initial
+        states from each of the given ``basis_states``.
+
+        ``w_init`` is the forward-facing version of this function'''
+
+        data_manager = self.data_manager
+        work_manager = self.work_manager
+        pstatus = self.rc.pstatus
+        system = self.system
+
+        pstatus('Creating HDF5 file {!r}'.format(self.data_manager.we_h5filename))
+        data_manager.prepare_backing()
+
+        # Process target states
+        data_manager.save_target_states(target_states)
+        self.report_target_states(target_states)
+
+        # Process basis states
+        self.get_bstate_pcoords(basis_states)
+        self.data_manager.create_ibstate_group(basis_states)
+        self.data_manager.create_ibstate_iter_h5file(basis_states)
+        self.report_basis_states(basis_states)
+
+        # Process start states
+        # Unlike the above, does not create an ibstate group.
+        # Should it? I don't think so, if needed it can be traced back through basis_auxref
+
+        # Here, we are trying to assign a state_id to the start state to be initialized, without actually
+        # saving it to the ibstates records in any of the h5 files. It might actually be a problem
+        # when tracing trajectories with westpa.analysis (especially with HDF5 framework) since it would
+        # try to look for a basis state id > len(basis_state).
+        # Since start states are only used while initializing and iteration 1, it's ok to not save it to save space. If necessary,
+        # the structure can be traced directly to the parent file using the standard basis state logic referencing
+        # west[iterations/iter_00000001/ibstates/istate_index/basis_auxref] of that istate.
+
+        if len(start_states) > 0 and start_states[0].state_id is None:
+            last_id = basis_states[-1].state_id
+            for start_state in start_states:
+                start_state.state_id = last_id + 1
+                last_id += 1
+
+        self.get_bstate_pcoords(start_states, label='start')
+        self.report_basis_states(start_states, label='start')
+
+        pstatus('Preparing initial states')
+        initial_states = []
+        weights = []
+        if self.do_gen_istates:
+            istate_type = InitialState.ISTATE_TYPE_GENERATED
+        else:
+            istate_type = InitialState.ISTATE_TYPE_BASIS
+
+        for basis_state in basis_states:
+            for _iseg in range(segs_per_state):
+                initial_state = data_manager.create_initial_states(1, 1)[0]
+                initial_state.basis_state_id = basis_state.state_id
+                initial_state.basis_state = basis_state
+                initial_state.istate_type = istate_type
+                weights.append(basis_state.probability / segs_per_state)
+                initial_states.append(initial_state)
+
+        for start_state in start_states:
+            for _iseg in range(segs_per_state):
+                initial_state = data_manager.create_initial_states(1, 1)[0]
+                initial_state.basis_state_id = start_state.state_id
+                initial_state.basis_state = start_state
+                initial_state.basis_auxref = start_state.auxref
+
+                # Start states are assigned their own type, so they can be identified later
+                initial_state.istate_type = InitialState.ISTATE_TYPE_START
+                weights.append(start_state.probability / segs_per_state)
+                initial_state.iter_used = 1
+                initial_states.append(initial_state)
+
+        if self.do_gen_istates:
+            futures = [
+                work_manager.submit(wm_ops.gen_istate, args=(initial_state.basis_state, initial_state))
+                for initial_state in initial_states
+            ]
+            for future in work_manager.as_completed(futures):
+                rbstate, ristate = future.get_result()
+                initial_states[ristate.state_id].pcoord = ristate.pcoord
+        else:
+            for initial_state in initial_states:
+                basis_state = initial_state.basis_state
+                initial_state.pcoord = basis_state.pcoord
+                initial_state.istate_status = InitialState.ISTATE_STATUS_PREPARED
+
+        for initial_state in initial_states:
+            log.debug('initial state created: {!r}'.format(initial_state))
+
+        # save list of initial states just generated
+        # some of these may not be used, depending on how WE shakes out
+        data_manager.update_initial_states(initial_states, n_iter=1)
+
+        if not suppress_we:
+            self.we_driver.populate_initial(initial_states, weights, system)
+            segments = list(self.we_driver.next_iter_segments)
+            binning = self.we_driver.next_iter_binning
+        else:
+            segments = list(self.we_driver.current_iter_segments)
+            binning = self.we_driver.final_binning
+
+        bin_occupancies = np.fromiter(map(len, binning), dtype=np.uint, count=self.we_driver.bin_mapper.nbins)
+        target_occupancies = np.require(self.we_driver.bin_target_counts, dtype=np.uint)
+
+        # total_bins/replicas defined here to remove target state bin from "active" bins
+        total_bins = len(bin_occupancies) - len(target_states)
+        total_replicas = int(sum(target_occupancies)) - int(self.we_driver.bin_target_counts[-1]) * len(target_states)
+
+        # Make sure we have
+        for segment in segments:
+            segment.n_iter = 1
+            segment.status = Segment.SEG_STATUS_PREPARED
+            assert segment.parent_id < 0
+            assert initial_states[segment.initial_state_id].iter_used == 1
+
+        data_manager.prepare_iteration(1, segments)
+        data_manager.update_initial_states(initial_states, n_iter=1)
+
+        if self.rc.verbose_mode:
+            pstatus('\nSegments generated:')
+            for segment in segments:
+                pstatus('{!r}'.format(segment))
+
+        pstatus(
+            '''
+        Total bins:            {total_bins:d}
+        Initial replicas:      {init_replicas:d} in {occ_bins:d} bins, total weight = {weight:g}
+        Total target replicas: {total_replicas:d}
+        '''.format(
+                total_bins=total_bins,
+                init_replicas=int(sum(bin_occupancies)),
+                occ_bins=len(bin_occupancies[bin_occupancies > 0]),
+                weight=float(sum(segment.weight for segment in segments)),
+                total_replicas=total_replicas,
+            )
+        )
+
+        total_prob = float(sum(segment.weight for segment in segments))
+        pstatus(f'1-prob: {1 - total_prob:.4e}')
+
+        target_counts = self.we_driver.bin_target_counts
+        # Do not include bins with target count zero (e.g. sinks, never-filled bins) in the (non)empty bins statistics
+        n_active_bins = len(target_counts[target_counts != 0])
+        seg_probs = np.fromiter(map(operator.attrgetter('weight'), segments), dtype=weight_dtype, count=len(segments))
+        norm = seg_probs.sum()
+
+        if not abs(1 - norm) < EPS * (len(segments) + n_active_bins):
+            pstatus("Normalization check failed at w_init, explicitly renormalizing")
+            for segment in segments:
+                segment.weight /= norm
+
+        # Send the segments over to the data manager to commit to disk
+        data_manager.current_iteration = 1
+
+        # Save BinMapper stuff for iteration 1
+        try:
+            pickled, hashed = self.we_driver.bin_mapper.pickle_and_hash()
+        except PickleError:
+            pickled = hashed = ''
+        data_manager.save_iter_binning(data_manager.current_iteration, hashed, pickled, target_counts)
+
+        # Report statistics
+        pstatus('Simulation prepared.')
+        self.segments = {segment.seg_id: segment for segment in segments}
+        self.report_bin_statistics(binning, target_states, save_summary=True)
+        data_manager.flush_backing()
+        data_manager.close_backing()
+
+    def prepare_iteration(self):
+        log.debug('beginning iteration {:d}'.format(self.n_iter))
+
+        # the WE driver needs a list of all target states for this iteration
+        # along with information about any new weights introduced (e.g. by recycling)
+        target_states = self.data_manager.get_target_states(self.n_iter)
+        new_weights = self.data_manager.get_new_weight_data(self.n_iter)
+
+        self.we_driver.new_iteration(target_states=target_states, new_weights=new_weights)
+
+        # Get basis states used in this iteration
+        self.current_iter_bstates = self.data_manager.get_basis_states(self.n_iter)
+
+        # Get the segments for this iteration and separate into complete and incomplete
+        if self.segments is None:
+            segments = self.segments = {segment.seg_id: segment for segment in self.data_manager.get_segments()}
+            log.debug('loaded {:d} segments'.format(len(segments)))
+        else:
+            segments = self.segments
+            log.debug('using {:d} pre-existing segments'.format(len(segments)))
+
+        completed_segments = self.completed_segments = {}
+        incomplete_segments = self.incomplete_segments = {}
+        for segment in segments.values():
+            if segment.status == Segment.SEG_STATUS_COMPLETE:
+                completed_segments[segment.seg_id] = segment
+            else:
+                incomplete_segments[segment.seg_id] = segment
+        log.debug('{:d} segments are complete; {:d} are incomplete'.format(len(completed_segments), len(incomplete_segments)))
+
+        if len(incomplete_segments) == len(segments):
+            # Starting a new iteration
+            self.rc.pstatus('Beginning iteration {:d}'.format(self.n_iter))
+        elif incomplete_segments:
+            self.rc.pstatus('Continuing iteration {:d}'.format(self.n_iter))
+        self.rc.pstatus(
+            '{:d} segments remain in iteration {:d} ({:d} total)'.format(len(incomplete_segments), self.n_iter, len(segments))
+        )
+
+        # Get the initial states active for this iteration (so that the propagator has them if necessary)
+        self.current_iter_istates = {
+            state.state_id: state for state in self.data_manager.get_segment_initial_states(list(segments.values()))
+        }
+        log.debug('This iteration uses {:d} initial states'.format(len(self.current_iter_istates)))
+
+        # Assign this iteration's segments' initial points to bins and report on bin population
+        initial_pcoords = self.system.new_pcoord_array(len(segments))
+        initial_binning = self.system.bin_mapper.construct_bins()
+        for iseg, segment in enumerate(segments.values()):
+            initial_pcoords[iseg] = segment.pcoord[0]
+        initial_assignments = self.system.bin_mapper.assign(initial_pcoords)
+        for segment, assignment in zip(iter(segments.values()), initial_assignments):
+            initial_binning[assignment].add(segment)
+        self.report_bin_statistics(initial_binning, [], save_summary=True)
+        del initial_pcoords, initial_binning
+
+        self.rc.pstatus('Waiting for segments to complete...')
+
+        # Let the WE driver assign completed segments
+        if completed_segments:
+            self.we_driver.assign(list(completed_segments.values()))
+
+        # load restart data
+        self.data_manager.prepare_segment_restarts(
+            incomplete_segments.values(), self.current_iter_bstates, self.current_iter_istates
+        )
+
+        # Get the basis states and initial states for the next iteration, necessary for doing on-the-fly recycling
+        self.next_iter_bstates = self.data_manager.get_basis_states(self.n_iter + 1)
+        self.next_iter_bstate_cprobs = np.add.accumulate([bstate.probability for bstate in self.next_iter_bstates])
+
+        self.we_driver.avail_initial_states = {
+            istate.state_id: istate for istate in self.data_manager.get_unused_initial_states(n_iter=self.n_iter + 1)
+        }
+        log.debug('{:d} unused initial states found'.format(len(self.we_driver.avail_initial_states)))
+
+        # Invoke callbacks
+        self.invoke_callbacks(self.prepare_iteration)
+
+        # dispatch and immediately wait on result for prep_iter
+        log.debug('dispatching propagator prep_iter to work manager')
+        self.work_manager.submit(wm_ops.prep_iter, args=(self.n_iter, segments)).get_result()
+
+    def finalize_iteration(self):
+        '''Clean up after an iteration and prepare for the next.'''
+        log.debug('finalizing iteration {:d}'.format(self.n_iter))
+
+        self.invoke_callbacks(self.finalize_iteration)
+
+        # dispatch and immediately wait on result for post_iter
+        log.debug('dispatching propagator post_iter to work manager')
+        self.work_manager.submit(wm_ops.post_iter, args=(self.n_iter, list(self.segments.values()))).get_result()
+
+        # Move existing segments into place as new segments
+        del self.segments
+        self.segments = {segment.seg_id: segment for segment in self.we_driver.next_iter_segments}
+
+        self.rc.pstatus("Iteration completed successfully")
+
+    def get_istate_futures(self):
+        '''Add ``n_states`` initial states to the internal list of initial states assigned to
+        recycled particles.  Spare states are used if available, otherwise new states are created.
+        If created new initial states requires generation, then a set of futures is returned
+        representing work manager tasks corresponding to the necessary generation work.'''
+
+        n_recycled = self.we_driver.n_recycled_segs
+        n_istates_needed = self.we_driver.n_istates_needed
+
+        log.debug('{:d} unused initial states available'.format(len(self.we_driver.avail_initial_states)))
+        log.debug('{:d} new initial states required for recycling {:d} walkers'.format(n_istates_needed, n_recycled))
+
+        futures = set()
+        updated_states = []
+        for _i in range(n_istates_needed):
+            # Select a basis state according to its weight
+            ibstate = np.digitize([self.rng.random()], self.next_iter_bstate_cprobs)
+            basis_state = self.next_iter_bstates[ibstate[0]]
+            initial_state = self.data_manager.create_initial_states(1, n_iter=self.n_iter + 1)[0]
+            initial_state.iter_created = self.n_iter
+            initial_state.basis_state_id = basis_state.state_id
+            initial_state.istate_status = InitialState.ISTATE_STATUS_PENDING
+
+            if self.do_gen_istates:
+                log.debug('generating new initial state from basis state {!r}'.format(basis_state))
+                initial_state.istate_type = InitialState.ISTATE_TYPE_GENERATED
+                futures.add(self.work_manager.submit(wm_ops.gen_istate, args=(basis_state, initial_state)))
+            else:
+                log.debug('using basis state {!r} directly'.format(basis_state))
+                initial_state.istate_type = InitialState.ISTATE_TYPE_BASIS
+                initial_state.pcoord = basis_state.pcoord.copy()
+                initial_state.istate_status = InitialState.ISTATE_STATUS_PREPARED
+                self.we_driver.avail_initial_states[initial_state.state_id] = initial_state
+            updated_states.append(initial_state)
+        self.data_manager.update_initial_states(updated_states, n_iter=self.n_iter + 1)
+        return futures
+
+    def propagate(self):
+        segments = list(self.incomplete_segments.values())
+        log.debug('iteration {:d}: propagating {:d} segments'.format(self.n_iter, len(segments)))
+
+        # all futures dispatched for this iteration
+        futures = set()
+        segment_futures = set()
+
+        # Immediately dispatch any necessary initial state generation
+        istate_gen_futures = self.get_istate_futures()
+        futures.update(istate_gen_futures)
+
+        # Dispatch propagation tasks using work manager
+        for segment_block in grouper(self.propagator_block_size, segments):
+            segment_block = [_f for _f in segment_block if _f]
+            pbstates, pistates = westpa.core.states.pare_basis_initial_states(
+                self.current_iter_bstates, list(self.current_iter_istates.values()), segment_block
+            )
+            future = self.work_manager.submit(wm_ops.propagate, args=(pbstates, pistates, segment_block))
+            futures.add(future)
+            segment_futures.add(future)
+
+        while futures and self.work_manager.running:
+            future = self.work_manager.wait_any(futures)
+            futures.remove(future)
+
+            if future in segment_futures:
+                segment_futures.remove(future)
+                incoming = future.get_result()
+                self.n_propagated += 1
+
+                self.segments.update({segment.seg_id: segment for segment in incoming})
+                self.completed_segments.update({segment.seg_id: segment for segment in incoming})
+
+                self.we_driver.assign(incoming)
+                new_istate_futures = self.get_istate_futures()
+                istate_gen_futures.update(new_istate_futures)
+                futures.update(new_istate_futures)
+
+                with self.data_manager.expiring_flushing_lock():
+                    self.data_manager.update_segments(self.n_iter, incoming)
+
+            elif future in istate_gen_futures:
+                istate_gen_futures.remove(future)
+                _basis_state, initial_state = future.get_result()
+                log.debug('received newly-prepared initial state {!r}'.format(initial_state))
+                initial_state.istate_status = InitialState.ISTATE_STATUS_PREPARED
+                with self.data_manager.expiring_flushing_lock():
+                    self.data_manager.update_initial_states([initial_state], n_iter=self.n_iter + 1)
+                self.we_driver.avail_initial_states[initial_state.state_id] = initial_state
+            else:
+                log.error('unknown future {!r} received from work manager'.format(future))
+                raise AssertionError('untracked future {!r}'.format(future))
+
+        log.debug('done with propagation')
+        self.save_bin_data()
+        self.data_manager.flush_backing()
+
+    def save_bin_data(self):
+        '''Calculate and write flux and transition count matrices to HDF5. Population and rate matrices
+        are likely useless at the single-tau level and are no longer written.'''
+        # save_bin_data(self, populations, n_trans, fluxes, rates, n_iter=None)
+
+        if self.save_transition_matrices:
+            with self.data_manager.expiring_flushing_lock():
+                iter_group = self.data_manager.get_iter_group(self.n_iter)
+                for key in ['bin_ntrans', 'bin_fluxes']:
+                    try:
+                        del iter_group[key]
+                    except KeyError:
+                        pass
+                iter_group['bin_ntrans'] = self.we_driver.transition_matrix
+                iter_group['bin_fluxes'] = self.we_driver.flux_matrix
+
+    def check_propagation(self):
+        '''Check for failures in propagation or initial state generation, and raise an exception
+        if any are found.'''
+
+        failed_segments = [segment for segment in self.segments.values() if segment.status != Segment.SEG_STATUS_COMPLETE]
+
+        if failed_segments:
+            failed_ids = '  \n'.join(str(segment.seg_id) for segment in failed_segments)
+            log.error('propagation failed for {:d} segment(s):\n{}'.format(len(failed_segments), failed_ids))
+            raise PropagationError('propagation failed for {:d} segments'.format(len(failed_segments)))
+        else:
+            log.debug('propagation complete for iteration {:d}'.format(self.n_iter))
+
+        failed_istates = [
+            istate
+            for istate in self.we_driver.used_initial_states.values()
+            if istate.istate_status != InitialState.ISTATE_STATUS_PREPARED
+        ]
+        log.debug('{!r}'.format(failed_istates))
+        if failed_istates:
+            failed_ids = '  \n'.join(str(istate.state_id) for istate in failed_istates)
+            log.error('initial state generation failed for {:d} states:\n{}'.format(len(failed_istates), failed_ids))
+            raise PropagationError('initial state generation failed for {:d} states'.format(len(failed_istates)))
+        else:
+            log.debug('initial state generation complete for iteration {:d}'.format(self.n_iter))
+
+    def run_we(self):
+        '''Run the weighted ensemble algorithm based on the binning in self.final_bins and
+        the recycled particles in self.to_recycle, creating and committing the next iteration's
+        segments to storage as well.'''
+
+        # The WE driver now does almost everything; we just have to record the
+        # mapper used for binning this iteration, and update initial states
+        # that have been used
+
+        try:
+            pickled, hashed = self.we_driver.bin_mapper.pickle_and_hash()
+        except PickleError:
+            pickled = hashed = ''
+
+        self.bin_mapper_hash = hashed
+        self.we_driver.construct_next()
+
+        if self.we_driver.used_initial_states:
+            for initial_state in self.we_driver.used_initial_states.values():
+                initial_state.iter_used = self.n_iter + 1
+            self.data_manager.update_initial_states(list(self.we_driver.used_initial_states.values()))
+
+        self.data_manager.update_segments(self.n_iter, list(self.segments.values()))
+
+        self.data_manager.require_iter_group(self.n_iter + 1)
+        self.data_manager.save_iter_binning(self.n_iter + 1, hashed, pickled, self.we_driver.bin_target_counts)
+
+        # Report on recycling
+        recycling_events = {}
+        for nw in self.we_driver.new_weights:
+            try:
+                recycling_events[nw.target_state_id].append(nw.weight)
+            except KeyError:
+                recycling_events[nw.target_state_id] = list([nw.weight])
+
+        tstates_by_id = {state.state_id: state for state in self.we_driver.target_states.values()}
+
+        for tstate_id, weights in recycling_events.items():
+            tstate = tstates_by_id[tstate_id]
+            self.rc.pstatus(
+                'Recycled {:g} probability ({:d} walkers) from target state {!r}'.format(sum(weights), len(weights), tstate.label)
+            )
+
+    def prepare_new_iteration(self):
+        '''Commit data for the coming iteration to the HDF5 file.'''
+        self.invoke_callbacks(self.prepare_new_iteration)
+
+        if self.rc.debug_mode:
+            self.rc.pstatus('\nSegments generated:')
+            for segment in self.we_driver.next_iter_segments:
+                self.rc.pstatus('{!r} pcoord[0]={!r}'.format(segment, segment.pcoord[0]))
+
+        self.data_manager.prepare_iteration(self.n_iter + 1, list(self.we_driver.next_iter_segments))
+        self.data_manager.save_new_weight_data(self.n_iter + 1, self.we_driver.new_weights)
+
+    def run(self):
+        run_starttime = time.time()
+        max_walltime = self.max_run_walltime
+        if max_walltime:
+            run_killtime = run_starttime + max_walltime
+            self.rc.pstatus('Maximum wallclock time: %s' % timedelta(seconds=max_walltime or 0))
+        else:
+            run_killtime = None
+
+        self.n_iter = self.data_manager.current_iteration
+        max_iter = self.max_total_iterations or self.n_iter + 1
+
+        iter_elapsed = 0
+        while self.n_iter <= max_iter:
+            if max_walltime and time.time() + 1.1 * iter_elapsed >= run_killtime:
+                self.rc.pstatus('Iteration {:d} would require more than the allotted time. Ending run.'.format(self.n_iter))
+                return
+
+            try:
+                iter_start_time = time.time()
+
+                self.rc.pstatus('\n%s' % time.asctime())
+                self.rc.pstatus('Iteration %d (%d requested)' % (self.n_iter, max_iter))
+
+                self.prepare_iteration()
+                self.rc.pflush()
+
+                self.pre_propagation()
+                self.propagate()
+                self.rc.pflush()
+                self.check_propagation()
+                self.rc.pflush()
+                self.post_propagation()
+
+                cputime = sum(segment.cputime for segment in self.segments.values())
+
+                self.rc.pflush()
+                self.pre_we()
+                self.run_we()
+                self.post_we()
+                self.rc.pflush()
+
+                self.prepare_new_iteration()
+
+                self.finalize_iteration()
+
+                iter_elapsed = time.time() - iter_start_time
+                iter_summary = self.data_manager.get_iter_summary()
+                iter_summary['walltime'] += iter_elapsed
+                iter_summary['cputime'] = cputime
+                self.data_manager.update_iter_summary(iter_summary)
+
+                self.n_iter += 1
+                self.data_manager.current_iteration += 1
+
+                try:
+                    # This may give NaN if starting a truncated simulation
+                    walltime = timedelta(seconds=float(iter_summary['walltime']))
+                except ValueError:
+                    walltime = 0.0
+
+                try:
+                    cputime = timedelta(seconds=float(iter_summary['cputime']))
+                except ValueError:
+                    cputime = 0.0
+
+                self.rc.pstatus('Iteration wallclock: {0!s}, cputime: {1!s}\n'.format(walltime, cputime))
+                self.rc.pflush()
+            finally:
+                self.data_manager.flush_backing()
+
+        self.rc.pstatus('\n%s' % time.asctime())
+        self.rc.pstatus('WEST run complete.')
+
+    def prepare_run(self):
+        '''Prepare a new run.'''
+        self.data_manager.prepare_run()
+        self.system.prepare_run()
+        self.invoke_callbacks(self.prepare_run)
+
+    def finalize_run(self):
+        '''Perform cleanup at the normal end of a run'''
+        self.invoke_callbacks(self.finalize_run)
+        self.system.finalize_run()
+        self.data_manager.finalize_run()
+
+    def pre_propagation(self):
+        self.invoke_callbacks(self.pre_propagation)
+
+    def post_propagation(self):
+        self.invoke_callbacks(self.post_propagation)
+
+    def pre_we(self):
+        self.invoke_callbacks(self.pre_we)
+
+    def post_we(self):
+        self.invoke_callbacks(self.post_we)