1 # Copyright (C) 2010 W. Trevor King <wking@drexel.edu>
3 # This program is free software: you can redistribute it and/or modify
4 # it under the terms of the GNU General Public License as published by
5 # the Free Software Foundation, either version 3 of the License, or
6 # (at your option) any later version.
8 # This program is distributed in the hope that it will be useful,
9 # but WITHOUT ANY WARRANTY; without even the implied warranty of
10 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11 # GNU General Public License for more details.
13 # You should have received a copy of the GNU General Public License
14 # along with this program. If not, see <http://www.gnu.org/licenses/>.
16 # The author may be contacted at <wking@drexel.edu> on the Internet, or
17 # write to Trevor King, Drudge's University, Physics Dept., 3141 Chestnut St.,
18 # Philadelphia PA 19104, USA.
20 """Functions for running external commands on other hosts.
22 mpi4py_ is a Python wrapper around MPI.
24 .. _mpi4py: http://mpi4py.scipy.org/
26 The MPIManager data flow is a bit complicated, so I've made a
29 [original intracommunicator] [intercom] [spawned intracommunicator]
31 ,->--(spawn-thread)-->--+-->------------->-(worker-0)-.
32 (main-thread) | `--->-(worker-1)-|
33 `-<-(receive-thread)-<--+--<-----. `->-(worker-2)-|
34 | `-------<-----------'
38 ============== ============== ===============================
39 Source Target Connection
40 ============== ============== ===============================
41 main-thread spawn-thread spawn_queue
42 spawn-thread worker-* SpawnThread.comm(SPAWN_TAG)
43 worker-* receive-thread ReceiveThread.comm(RECEIVE_TAG)
44 receive-thread main-thread receive_queue
45 ============== ============== ===============================
47 There is also a `free_queue` running from `receive-thread` to
48 `spawn-thread` to mark job completion so `spawn-thread` knows which
49 nodes are free (and therefore ready to receive new jobs).
52 from Queue import Queue, Empty
54 from threading import Thread
57 from mpi4py import MPI
59 _DISABLING_ERROR = None
60 if MPI.COMM_WORLD.Get_rank() == 0:
63 _SKIP = ' # doctest: +SKIP'
64 except ImportError, _DISABLING_ERROR:
66 _SKIP = ' # doctest: +SKIP'
70 from .thread import CLOSE_MESSAGE, ThreadManager
73 CLOSE_MESSAGE = "close"
78 def MPI_worker_death():
81 if MPI.COMM_WORLD.Get_rank() != 0:
86 raise _DISABLING_ERROR
87 rank = MPI.COMM_WORLD.Get_rank()
89 'process %d should have been killed by an MPI_worker_death() call'
93 class WorkerProcess (object):
95 self.comm = MPI.Comm.Get_parent() # intercommunicator
96 self.rank = self.comm.Get_rank() # *intracom* rank?
98 self.name = 'worker-%d' % self.rank
99 log().debug('%s started' % self.name)
103 # only one worker needs to disconnect from the intercommunicator.
104 self.comm.Disconnect()
109 msg = self.comm.recv(source=self.manager, tag=SPAWN_TAG, status=s)
110 if msg == CLOSE_MESSAGE:
111 log().debug('%s closing' % self.name)
113 assert isinstance(msg, Job), msg
114 log().debug('%s running job %s' % (self.name, msg))
116 self.comm.send(msg, dest=self.manager, tag=RECEIVE_TAG)
118 # forward close message to receive-thread
119 self.comm.send(CLOSE_MESSAGE, dest=self.manager, tag=RECEIVE_TAG)
122 class ManagerThread (Thread):
123 def __init__(self, job_queue, free_queue, comm, rank, size,
125 super(ManagerThread, self).__init__(*args, **kwargs)
126 self.job_queue = job_queue
127 self.free_queue = free_queue
131 self.name = self.getName() # work around Pythons < 2.6
132 log().debug('%s starting' % self.name)
135 class SpawnThread (ManagerThread):
137 for i in range(self.size):
139 self.comm.send(CLOSE_MESSAGE, dest=i, tag=SPAWN_TAG)
141 while len(free) < self.size:
142 free.append(self.free_queue.get())
143 # close recieve-thread via worker-0
144 self.comm.send(CLOSE_MESSAGE, dest=0, tag=SPAWN_TAG)
148 msg = self.job_queue.get()
149 if msg == CLOSE_MESSAGE:
150 log().debug('%s closing' % self.name)
153 assert isinstance(msg, Job), msg
154 rank = self.free_queue.get()
155 log().debug('%s sending job %s to %d' % (self.name, msg, rank))
156 self.comm.send(msg, dest=rank, tag=SPAWN_TAG)
159 class ReceiveThread (ManagerThread):
163 msg = self.comm.recv(
164 source=MPI.ANY_SOURCE, tag=RECEIVE_TAG, status=s)
165 if msg == CLOSE_MESSAGE:
166 log().debug('%s closing' % self.name)
167 self.comm.Disconnect()
169 rank = s.Get_source()
170 self.free_queue.put(rank)
171 log().debug('%s got job %s from %d' % (self.name, msg, rank))
172 assert isinstance(msg, Job), msg
173 self.job_queue.put(msg)
176 class MPIManager (ThreadManager):
177 __doc__ = """Manage asynchronous `Job` execution via :mod:`pbs`.
179 >>> from math import sqrt
180 >>> m = MPIManager()%(skip)s
182 >>> for i in range(10):
183 ... group_A.append(m.async_invoke(Job(target=sqrt, args=[i])))%(skip)s
185 >>> for i in range(10):
186 ... group_B.append(m.async_invoke(Job(target=sqrt, args=[i],
187 ... blocks_on=[j.id for j in group_A])))%(skip)s
188 >>> jobs = m.wait(ids=[j.id for j in group_A[5:8]])%(skip)s
189 >>> print sorted(jobs.values(), key=lambda j: j.id)%(skip)s
190 [<Job 5>, <Job 6>, <Job 7>]
191 >>> jobs = m.wait()%(skip)s
192 >>> print sorted(jobs.values(), key=lambda j: j.id)%(skip)s
193 ... # doctest: +NORMALIZE_WHITESPACE
194 [<Job 0>, <Job 1>, <Job 2>, <Job 3>, <Job 4>, <Job 8>, <Job 9>, <Job 10>,
195 <Job 11>, <Job 12>, <Job 13>, <Job 14>, <Job 15>, <Job 16>, <Job 17>,
197 >>> m.teardown()%(skip)s
198 """ % {'skip': _SKIP}
200 def __init__(self, worker_pool=None):
202 super(MPIManager, self).__init__(worker_pool)
204 def _spawn_workers(self, worker_pool):
205 spawn_script = ';'.join([
206 'from %s import WorkerProcess' % __name__,
207 'w = WorkerProcess()',
211 if worker_pool == None:
212 worker_pool = MPI.COMM_WORLD.Get_size()
213 comm = MPI.COMM_SELF.Spawn( # locks with mpich2 if no mpd running
214 sys.executable, args=['-c', spawn_script], maxprocs=worker_pool)
215 rank = comm.Get_rank()
216 assert rank == 0, rank
217 # `comm` connects `COMM_WORLD` with the spawned group
218 # (intercommunicator), so `comm.Get_size() == 1` regardless of
219 # `worker_pool`. We want to know the size of the worker pool,
223 for worker_rank in range(size):
224 free_queue.put(worker_rank)
227 for worker in [SpawnThread(self._spawn_queue, free_queue,
229 name='spawn-thread'),
230 ReceiveThread(self._receive_queue, free_queue,
232 name='receive-thread'),
234 log().debug('start %s' % worker.name)
236 self._workers.append(worker)