CraySystem.py 66.8 KB
Newer Older
1
2
"""Resource management for Cray ALPS based systems"""

3
4
5
6
import logging
import threading
import thread
import time
7
import sys
8
import xmlrpclib
9
import json
10

11
12
import Cobalt.Util
import Cobalt.Components.system.AlpsBridge as ALPSBridge
13

14
from Cobalt.Components.base import Component, exposed, automatic, query, locking
15
16
from Cobalt.Components.system.base_system import BaseSystem
from Cobalt.Components.system.CrayNode import CrayNode
17
from Cobalt.Components.system.base_pg_manager import ProcessGroupManager
18
from Cobalt.Components.system.ALPSProcessGroup import ALPSProcessGroup
19
from Cobalt.Exceptions import ComponentLookupError
Paul Rich's avatar
Paul Rich committed
20
from Cobalt.Exceptions import JobNotInteractive
21
from Cobalt.Components.system.ALPSProcessGroup import ALPSProcessGroup
22
from Cobalt.Exceptions import JobValidationError
23
from Cobalt.DataTypes.ProcessGroup import ProcessGroup
24
from Cobalt.Util import compact_num_list, expand_num_list
25
from Cobalt.Util import init_cobalt_config, get_config_option
26

27
28
_logger = logging.getLogger(__name__)

29
init_cobalt_config()
30

31
SYSTEM_SIZE = int(get_config_option('system', 'size'))
32
33
34
35
36
UPDATE_THREAD_TIMEOUT = int(get_config_option('alpssystem',
    'update_thread_timeout', 10))
TEMP_RESERVATION_TIME = int(get_config_option('alpssystem',
    'temp_reservation_time', 300))
SAVE_ME_INTERVAL = float(get_config_option('alpsssytem', 'save_me_interval', 10.0))
Paul Rich's avatar
Paul Rich committed
37
38
PENDING_STARTUP_TIMEOUT = float(get_config_option('alpssystem',
    'pending_startup_timeout', 1200)) #default 20 minutes to account for boot.
Paul Rich's avatar
Paul Rich committed
39
APKILL_CMD = get_config_option('alps', 'apkill', '/opt/cray/alps/default/bin/apkill')
40
DRAIN_MODE = get_config_option('system', 'drain_mode', 'first-fit')
41
42
43
#cleanup time in seconds
CLEANUP_DRAIN_WINDOW = get_config_option('system', 'cleanup_drain_window', 300)

44
45
#Epsilon for backfilling.  This system does not do this on a per-node basis.
BACKFILL_EPSILON = int(get_config_option('system', 'backfill_epsilon', 120))
46
47
48
ELOGIN_HOSTS = [host for host in get_config_option('alpssystem', 'elogin_hosts', '').split(':')]
if ELOGIN_HOSTS == ['']:
    ELOGIN_HOSTS = []
Paul Rich's avatar
Paul Rich committed
49
DRAIN_MODES = ['first-fit', 'backfill']
50
CLEANING_ID = -1
51

52
53


54
55
56
57
58
59
60
61
62
63
def chain_loc_list(loc_list):
    '''Take a list of compact Cray locations,
    expand and concatenate them.

    '''
    retlist = []
    for locs in loc_list:
        retlist.extend(expand_num_list(locs))
    return retlist

64

65
class CraySystem(BaseSystem):
66
67
68
    '''Cray/ALPS-specific system component.  Behaviors should go here.  Direct
    ALPS interaction through BASIL/other APIs should go through the ALPSBridge
    (or other bridge) module as appropriate.
69

70
    '''
71
72
    name = "system"
    implementation = "alps_system"
73
    logger = _logger
74

75
    def __init__(self, *args, **kwargs):
76
77
78
79
        '''Initialize system.  Read initial states from bridge.
        Get current state

        '''
80
        start_time = time.time()
81
        super(CraySystem, self).__init__(*args, **kwargs)
82
        _logger.info('BASE SYSTEM INITIALIZED')
83
84
85
86
87
88
89
90
91
92
        self._common_init_restart()
        _logger.info('ALPS SYSTEM COMPONENT READY TO RUN')
        _logger.info('Initilaization complete in %s sec.', (time.time() -
                start_time))

    def _common_init_restart(self, spec=None):
        '''Common routine for cold and restart intialization of the system
        component.

        '''
93
94
95
96
97
        if DRAIN_MODE not in DRAIN_MODES:
            #abort startup, we have a completely invalid config.
            _logger.critical('ALPS SYSTEM: ABORT STARTUP: %s is not a valid drain mode.  Must be one of %s.',
                DRAIN_MODE, ", ".join(DRAIN_MODES))
            sys.exit(1)
98
        #initilaize bridge.
99
100
101
102
        bridge_pending = True
        while bridge_pending:
            # purge stale children from prior run.  Also ensure the
            # system_script_forker is currently up.
Paul Rich's avatar
Paul Rich committed
103
104
            # These attempts may fail due to system_script_forker not being up.
            # We don't want to trash the statefile in this case.
105
106
107
108
            try:
                ALPSBridge.init_bridge()
            except ALPSBridge.BridgeError:
                _logger.error('Bridge Initialization failed.  Retrying.')
109
110
111
112
                Cobalt.Util.sleep(10)
            except ComponentLookupError:
                _logger.warning('Error reaching forker.  Retrying.')
                Cobalt.Util.sleep(10)
113
114
115
            else:
                bridge_pending = False
                _logger.info('BRIDGE INITIALIZED')
116
        #process manager setup
117
        if spec is None:
Paul Rich's avatar
Paul Rich committed
118
            self.process_manager = ProcessGroupManager(pgroup_type=ALPSProcessGroup)
119
        else:
Paul Rich's avatar
Paul Rich committed
120
            self.process_manager = ProcessGroupManager(pgroup_type=ALPSProcessGroup).__setstate__(spec['process_manager'])
Paul Rich's avatar
Paul Rich committed
121
            self.logger.debug('pg type %s', self.process_manager.process_groups.item_cls)
122
123
        #self.process_manager.forkers.append('alps_script_forker')
        self.process_manager.update_launchers()
Paul Rich's avatar
Paul Rich committed
124
        self.pending_start_timeout = PENDING_STARTUP_TIMEOUT
125
126
        _logger.info('PROCESS MANAGER INTIALIZED')
        #resource management setup
127
        self.nodes = {} #cray node_id: CrayNode
128
129
130
131
        self.node_name_to_id = {} #cray node name to node_id map
        self.alps_reservations = {} #cobalt jobid : AlpsReservation object
        if spec is not None:
            self.alps_reservations = spec['alps_reservations']
132
        self._init_nodes_and_reservations()
133
134
135
        if spec is not None:
            node_info = spec.get('node_info', {})
            for nid, node in node_info.items():
136
137
138
139
                try:
                    self.nodes[nid].reset_info(node)
                except: #check the exception types later.  Carry on otherwise.
                    self.logger.warning("Node nid: %s not found in restart information.  Bringing up node in a clean configuration.", nid)
Paul Rich's avatar
Paul Rich committed
140
141
142
        #storage for pending job starts.  Allows us to handle slow starts vs
        #user deletes
        self.pending_starts = {} #jobid: time this should be cleared.
143
        self.nodes_by_queue = {} #queue:[node_ids]
144
        #populate initial state
145
        #state update thread and lock
146
        self._node_lock = threading.RLock()
147
        self._gen_node_to_queue()
Paul Rich's avatar
Paul Rich committed
148
        self.node_update_thread = thread.start_new_thread(self._run_update_state, tuple())
149
        _logger.info('UPDATE THREAD STARTED')
150
        self.current_equivalence_classes = []
Paul Rich's avatar
Paul Rich committed
151
        self.killing_jobs = {}
152
153
154
155
        #hold on to the initial spec in case nodes appear out of nowhere.
        self.init_spec = None
        if spec is not None:
            self.init_spec = spec
156

157
158
159
160
161
162
163
164
    def __getstate__(self):
        '''Save process, alps-reservation information, along with base
        information'''
        state = {}
        state.update(super(CraySystem, self).__getstate__())
        state['alps_system_statefile_version'] = 1
        state['process_manager'] = self.process_manager.__getstate__()
        state['alps_reservations'] = self.alps_reservations
165
        state['node_info'] = self.nodes
166
167
168
169
170
171
172
        return state

    def __setstate__(self, state):
        start_time = time.time()
        super(CraySystem, self).__setstate__(state)
        _logger.info('BASE SYSTEM INITIALIZED')
        self._common_init_restart(state)
173
        _logger.info('ALPS SYSTEM COMPONENT READY TO RUN')
174
        _logger.info('Reinitilaization complete in %s sec.', (time.time() -
175
                start_time))
176

177
178
179
180
181
    def save_me(self):
        '''Automatically save a copy of the state of the system component.'''
        #should we be holding the block lock as well?
        Component.save(self)
    save_me = automatic(save_me, SAVE_ME_INTERVAL)
182

183
184
185
186
    def _init_nodes_and_reservations(self):
        '''Initialize nodes from ALPS bridge data'''

        retnodes = {}
187
188
189
        pending = True
        while pending:
            try:
190
191
192
193
194
                # None of these queries has strictly degenerate data.  Inventory
                # is needed for raw reservation data.  System gets memory and a
                # much more compact representation of data.  Reservednodes gives
                # which notes are reserved.
                inventory = ALPSBridge.fetch_inventory()
Paul Rich's avatar
Paul Rich committed
195
                _logger.info('INVENTORY FETCHED')
196
                system = ALPSBridge.extract_system_node_data(ALPSBridge.system())
Paul Rich's avatar
Paul Rich committed
197
                _logger.info('SYSTEM DATA FETCHED')
198
                reservations = ALPSBridge.fetch_reservations()
Paul Rich's avatar
Paul Rich committed
199
                _logger.info('ALPS RESERVATION DATA FETCHED')
200
                # reserved_nodes = ALPSBridge.reserved_nodes()
201
202
203
            except Exception:
                #don't crash out here.  That may trash a statefile.
                _logger.error('Possible transient encountered during initialization. Retrying.',
204
                        exc_info=True)
205
206
207
208
                Cobalt.Util.sleep(10)
            else:
                pending = False

209
        self._assemble_nodes(inventory, system)
210
211
        #Reversing the node name to id lookup is going to save a lot of cycles.
        for node in self.nodes.values():
212
            self.node_name_to_id[node.name] = node.node_id
213
214
        _logger.info('NODE INFORMATION INITIALIZED')
        _logger.info('ALPS REPORTS %s NODES', len(self.nodes))
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
        # self._assemble_reservations(reservations, reserved_nodes)
        return

    def _assemble_nodes(self, inventory, system):
        '''merge together the INVENTORY and SYSTEM query data to form as
        complete a picture of a node as we can.

        '''
        nodes = {}
        for nodespec in inventory['nodes']:
            node = CrayNode(nodespec)
            node.managed = True
            nodes[node.node_id] = node
        for node_id, nodespec in system.iteritems():
            nodes[node_id].attributes.update(nodespec['attrs'])
230
231
232
233
            # Should this be a different status?
            nodes[node_id].role = nodespec['role'].upper()
            if nodes[node_id].role.upper() not in ['BATCH']:
                nodes[node_id].status = 'down'
234
235
236
237
238
239
            nodes[node_id].status = nodespec['state']
        self.nodes = nodes

    def _assemble_reservations(self, reservations, reserved_nodes):
        # FIXME: we can recover reservations now.  Implement this.
        pass
240

241
    def _gen_node_to_queue(self):
242
        '''(Re)Generate a mapping for fast lookup of node-id's to queues.'''
243
        with self._node_lock:
244
            self.nodes_by_queue = {}
245
246
247
            for node in self.nodes.values():
                for queue in node.queues:
                    if queue in self.nodes_by_queue.keys():
248
                        self.nodes_by_queue[queue].add(node.node_id)
249
                    else:
250
                        self.nodes_by_queue[queue] = set([node.node_id])
251

252
    @exposed
253
    def get_nodes(self, as_dict=False, node_ids=None, params=None, as_json=False):
254
255
        '''fetch the node dictionary.

256
257
258
259
260
261
262
            as_dict  - Return node information as a dictionary keyed to string
                        node_id value.
            node_ids - A list of node names to return, if None, return all nodes
                       (default None).
            params   - If requesting a dict, only request this list of
                       parameters of the node.
            json     - Encode to json before sending.  Useful on large systems.
263
264
265
266
267

            returns the node dictionary.  Can reutrn underlying node data as
            dictionary for XMLRPC purposes

        '''
268
269
270
271
272
273
274
275
276
        def node_filter(node):
            if node_ids is not None:
                return (str(node[0]) in [str(nid) for nid in node_ids])
            return True

        node_info = None
        if as_dict:
            retdict = {k:v.to_dict(True, params) for k, v in self.nodes.items()}
            node_info = dict(filter(node_filter, retdict.items()))
277
        else:
278
279
280
281
282
            node_info = dict(filter(node_filter, self.nodes.items()))
        if as_json:
            return json.dumps(node_info)
        return node_info

283
284
285
    def _run_update_state(self):
        '''automated node update functions on the update timer go here.'''
        while True:
286
287
288
289
290
291
292
293
294
            try:
                self.process_manager.update_launchers()
                self.update_node_state()
                self._get_exit_status()
            except Exception:
                # prevent the update thread from dying
                _logger.critical('Error in _run_update_state', exc_info=True)
            finally:
                Cobalt.Util.sleep(UPDATE_THREAD_TIMEOUT)
295

296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
    def _reconstruct_node(self, inven_node, inventory):
        '''Reconstruct a node from statefile information.  Needed whenever we
        find a new node.  If no statefile information from the orignal cobalt
        invocation exists, bring up a node in default states and mark node
        administratively down.

        This node was disabled and invisible to ALPS at the time Cobalt was
        initialized and so we have no current record of that node.

        '''
        nid = inven_node['node_id']
        new_node = None
        #construct basic node from inventory
        for node_info in inventory['nodes']:
            if int(node_info['node_id']) == int(nid):
                new_node = CrayNode(node_info)
                break
        if new_node is None:
            #we have a phantom node?
            self.logger.error('Unable to find inventory information for nid: %s', nid)
            return
        # if we have information from the statefile we need to repopulate the
        # node with the saved data.
        # Perhaps this should be how I construct all node data anyway?
        if self.init_spec is not None:
            node_info = self.init_spec.get('node_info', {})
            try:
                new_node.reset_info(node_info[str(nid)])
                self.logger.warning('Node %s reconstructed.', nid)
            except:
                self.logger.warning("Node nid: %s not found in restart information.  Bringing up node in a clean configuration.", nid, exc_info=True)
                #put into admin_down
                new_node.admin_down = True
                new_node.status = 'down'
                self.logger.warning('Node %s marked down.', nid)
Paul Rich's avatar
Paul Rich committed
331
        new_node.managed = True
332
333
334
        self.nodes[str(nid)] = new_node
        self.logger.warning('Node %s added to tracking.', nid)

335
336
337
338
339
340
    @exposed
    def update_node_state(self):
        '''update the state of cray nodes. Check reservation status and system
        stateus as reported by ALPS

        '''
Paul Rich's avatar
Paul Rich committed
341
342
343
        #Check clenaup progress.  Check ALPS reservations.  Check allocated
        #nodes.  If there is no resource reservation and the node is not in
        #current alps reservations, the node is ready to schedule again.
344
        now = time.time()
Paul Rich's avatar
Paul Rich committed
345
346
347
348
349
350
351
352
        startup_time_to_clear = []
        #clear pending starttimes.
        for jobid, start_time in self.pending_starts.items():
            if int(now) > int(start_time):
                startup_time_to_clear.append(jobid)
        for jobid in startup_time_to_clear:
            del self.pending_starts[jobid]

Paul Rich's avatar
Paul Rich committed
353
        self.check_killing_aprun()
354
        with self._node_lock:
Paul Rich's avatar
Paul Rich committed
355
            fetch_time_start = time.time()
356
            try:
357
358
                #I have seen problems with the kitchen-sink query here, where
                #the output gets truncated on it's way into Cobalt.
Paul Rich's avatar
Paul Rich committed
359
                #inventory = ALPSBridge.fetch_inventory(resinfo=True) #This is a full-refresh,
360
                #determine if summary may be used under normal operation
361
                #updated for >= 1.6 interface
Paul Rich's avatar
Paul Rich committed
362
363
364
                inven_nodes = ALPSBridge.extract_system_node_data(ALPSBridge.system())
                reservations = ALPSBridge.fetch_reservations()
                #reserved_nodes = ALPSBridge.reserved_nodes()
365
366
367
368
            except (ALPSBridge.ALPSError, ComponentLookupError):
                _logger.warning('Error contacting ALPS for state update.  Aborting this update',
                        exc_info=True)
                return
369
            inven_reservations = reservations.get('reservations', [])
Paul Rich's avatar
Paul Rich committed
370
            fetch_time_start = time.time()
371
            #_logger.debug("time in ALPS fetch: %s seconds", (time.time() - fetch_time_start))
372
            start_time = time.time()
Paul Rich's avatar
Paul Rich committed
373
            self._detect_rereservation(inven_reservations)
374
375
376
377
378
379
            # check our reservation objects.  If a res object doesn't correspond
            # to any backend reservations, this reservation object should be
            # dropped
            alps_res_to_delete = []
            current_alps_res_ids = [int(res['reservation_id']) for res in
                    inven_reservations]
Paul Rich's avatar
Paul Rich committed
380
            res_jobid_to_delete = []
381
382
383
384
385
386
387
388
389
            if self.alps_reservations == {}:
                # if we have nodes in cleanup-pending but no alps reservations,
                # then the nodes in cleanup pending are considered idle (or
                # at least not in cleanup).  Hardware check can catch these
                # later.
                for node in self.nodes.values():
                    if node.status in ['cleanup', 'cleanup-pending']:
                        node.status = 'idle'
            for alps_res in self.alps_reservations.values():
Paul Rich's avatar
Paul Rich committed
390
391
                if alps_res.jobid in self.pending_starts.keys():
                    continue #Get to this only after timeout happens
Paul Rich's avatar
Paul Rich committed
392
                #find alps_id associated reservation
393
                if int(alps_res.alps_res_id) not in current_alps_res_ids:
Paul Rich's avatar
Paul Rich committed
394
                    for node_id in alps_res.node_ids:
395
                        if not self.nodes[str(node_id)].reserved:
Paul Rich's avatar
Paul Rich committed
396
                            #pending hardware status update
397
                            self.nodes[str(node_id)].status = 'idle'
Paul Rich's avatar
Paul Rich committed
398
                    res_jobid_to_delete.append(alps_res.jobid)
399
400
                    _logger.info('Nodes %s cleanup complete.',
                            compact_num_list(alps_res.node_ids))
Paul Rich's avatar
Paul Rich committed
401
            for jobid in res_jobid_to_delete:
402
403
                _logger.info('%s: ALPS reservation for this job complete.', jobid)
                del self.alps_reservations[str(jobid)]
Paul Rich's avatar
Paul Rich committed
404
405
            #process group should already be on the way down since cqm released the
            #resource reservation
406
407
            cleanup_nodes = [node for node in self.nodes.values()
                             if node.status == 'cleanup-pending']
Paul Rich's avatar
Paul Rich committed
408
            #If we have a block marked for cleanup, send a release message.
Paul Rich's avatar
Paul Rich committed
409
410
            released_res_jobids = []
            for node in cleanup_nodes:
411
                for alps_res in self.alps_reservations.values():
Paul Rich's avatar
Paul Rich committed
412
                    if (alps_res.jobid not in released_res_jobids and
413
414
                            str(node.node_id) in alps_res.node_ids):
                        #send only one release per iteration
Paul Rich's avatar
Paul Rich committed
415
416
417
418
                        apids = alps_res.release()
                        if apids is not None:
                            for apid in apids:
                                self.signal_aprun(apid)
419
                        released_res_jobids.append(alps_res.jobid)
Paul Rich's avatar
Paul Rich committed
420
421

        #find hardware status
422
423
424
            #so we do this only once for nodes being added.
            #full inventory fetch is expensive.
            recon_inventory = None
Paul Rich's avatar
Paul Rich committed
425
            for inven_node in inven_nodes.values():
426
427
                if self.nodes.has_key(str(inven_node['node_id'])):
                    node = self.nodes[str(inven_node['node_id'])]
428
                    node.role = inven_node['role'].upper()
Paul Rich's avatar
Paul Rich committed
429
                    if node.reserved:
430
431
                        #node marked as reserved.
                        if self.alps_reservations.has_key(str(node.reserved_jobid)):
Paul Rich's avatar
Paul Rich committed
432
433
                            node.status = 'busy'
                        else:
434
435
436
437
438
439
                            # check to see if the resource reservation should be
                            # released.
                            if node.reserved_until >= now:
                                node.status = 'allocated'
                            else:
                                node.release(user=None, jobid=None, force=True)
Paul Rich's avatar
Paul Rich committed
440
                    else:
441
                        node.status = inven_node['state'].upper()
442
443
                        if node.role.upper() not in ['BATCH'] and node.status is 'idle':
                            node.status = 'alps-interactive'
444
                else:
445
446
447
448
449
450
451
452
453
454
455
456
457
                    # Apparently, we CAN add nodes on the fly.  The node would
                    # have been disabled.  We need to add a new node and update
                    # it's state.
                    _logger.warning('Unknown node %s found. Starting reconstruction.', inven_node['node_id'])
                    try:
                        if recon_inventory is None:
                            recon_inventory = ALPSBridge.fetch_inventory()
                    except:
                        _logger.error('Failed to fetch inventory.  Will retry on next pass.', exc_info=True)
                    else:
                        self._reconstruct_node(inven_node, recon_inventory)
                   # _logger.error('UNS: ALPS reports node %s but not in our node list.',
                   #               inven_node['node_id'])
Paul Rich's avatar
Paul Rich committed
458
459
460
            #should down win over running in terms of display?
            #keep node that are marked for cleanup still in cleanup
            for node in cleanup_nodes:
461
                node.status = 'cleanup-pending'
462
        #_logger.debug("time in UNS lock: %s seconds", (time.time() - start_time))
463
464
        return

Paul Rich's avatar
Paul Rich committed
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
    def _detect_rereservation(self, inven_reservations):
        '''Detect and update the ALPS reservation associated with a running job.
        We are only concerned with BATCH reservations.  Others would be
        associated with running jobs, and should not be touched.

        '''
        def _construct_alps_res():
            with self._node_lock:
                job_nodes = [node.node_id for node in self.nodes.values()
                        if node.reserved_jobid == int(alps_res['batch_id'])]
            new_resspec = {'reserved_nodes': job_nodes,
                           'reservation_id': str(alps_res['reservation_id']),
                           'pagg_id': 0 #unknown.  Not used here.
                            }
            new_jobspec = {'jobid': int(alps_res['batch_id']),
                           'user' : alps_res['user_name']}

            return ALPSReservation(new_jobspec, new_resspec, self.nodes)

        replaced_reservation = None
        for alps_res in inven_reservations:
            try:
                #This traversal is terrible. May want to hide this in the API
                #somewhere
                if alps_res['ApplicationArray'][0]['Application'][0]['CommandArray'][0]['Command'][0]['cmd'] != 'BASIL':
                    # Not a reservation we're in direct control of.
                    continue
            except (KeyError, IndexError):
                #not a batch reservation
                continue
            if str(alps_res['batch_id']) in self.alps_reservations.keys():
                # This is a reservation we may already know about
                if (int(alps_res['reservation_id']) ==
                        self.alps_reservations[str(alps_res['batch_id'])].alps_res_id):
                    # Already know about this one
                    continue
                # we have a re-reservation.  If this has a batch id, we need
                # to add it to our list of tracked reservations, and inherit
                # other reservation details.  We can pull the reservation
                # information out of reserve_resources_until.

                # If this is a BATCH reservation and no hardware has that
                # reservation id, then this reservation needs to be released
                # Could happen if we have a job starting right at the RRU
                # boundary.
                new_alps_res = _construct_alps_res()
                tracked_res = self.alps_reservations.get(new_alps_res.jobid, None)
                if tracked_res is not None:
                    try:
Paul Rich's avatar
Paul Rich committed
514
                        apids = tracked_res.release()
Paul Rich's avatar
Paul Rich committed
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
                    except ALPSBridge.ALPSError:
                        # backend reservation probably is gone, which is why
                        # we are here in the first place.
                        pass
                self.alps_reservations[str(alps_res['batch_id'])] = new_alps_res
            else:
                #this is a basil reservation we don't know about already.
                new_alps_res = _construct_alps_res()
                if len(new_alps_res.node_ids) == 0:
                    # This reservation has no resources, so Cobalt's internal
                    # resource reservation tracking has no record.  This needs to
                    # be removed.
                    new_alps_res.release()
                else:
                    self.alps_reservations[str(alps_res['batch_id'])] = new_alps_res
Paul Rich's avatar
Paul Rich committed
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
        return

    def signal_aprun(self, aprun_id, signame='SIGINT'):
        '''Signal an aprun by aprun id (application_id).  Does not block.
        Use check_killing_aprun to determine completion/termination.  Does not
        depend on the host the aprun(s) was launched from.

        Input:
            aprun_id - integer application id number.
            signame  - string name of signal to send (default: SIGINT)
        Notes:
            Valid signals to apkill are:
            SIGHUP, SIGINT, SIGQUIT, SIGTERM, SIGABRT, SIGUSR1, SIGUSR2, SIGURG,
            and SIGWINCH (from apkill(1) man page.)  Also allowing SIGKILL.

        '''
        #Expect changes with an API updte

        #mark legal signals from docos
        if (signame not in ['SIGHUP', 'SIGINT', 'SIGQUIT', 'SIGTERM', 'SIGABRT',
            'SIGUSR1', 'SIGUSR2', 'SIGURG','SIGWINCH', 'SIGKILL']):
            raise ValueError('%s is not a legal value for signame.', signame)
        try:
            retval = Cobalt.Proxy.ComponentProxy('system_script_forker').fork(
                    [APKILL_CMD, '-%s' % signame, '%d' % int(aprun_id)],
                    'aprun_termination', '%s cleanup:'% aprun_id)
Paul Rich's avatar
Paul Rich committed
556
            _logger.info("killing backend ALPS application_id: %s", aprun_id)
Paul Rich's avatar
Paul Rich committed
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
        except xmlrpclib.Fault:
            _logger.warning("XMLRPC Error while killing backend job: %s, will retry.",
                    aprun_id, exc_info=True)
        except:
            _logger.critical("Unknown Error while killing backend job: %s, will retry.",
                    aprun_id, exc_info=True)
        else:
            self.killing_jobs[aprun_id] = retval
        return

    def check_killing_aprun(self):
        '''Check that apkill commands have completed and clean them from the
        system_script_forker.  Allows for non-blocking cleanup initiation.

        '''
572

Paul Rich's avatar
Paul Rich committed
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
        try:
            system_script_forker = Cobalt.Proxy.ComponentProxy('system_script_forker')
        except:
            self.logger.critical("Cannot connect to system_script_forker.",
                    exc_info=True)
            return
        complete_jobs = []
        rev_killing_jobs = dict([(v,k) for (k,v) in self.killing_jobs.iteritems()])
        removed_jobs = []
        current_killing_jobs = system_script_forker.get_children(None, self.killing_jobs.values())

        for job in current_killing_jobs:
            if job['complete']:
                del self.killing_jobs[rev_killing_jobs[int(job['id'])]]
                removed_jobs.append(job['id'])
        system_script_forker.cleanup_children(removed_jobs)
Paul Rich's avatar
Paul Rich committed
589
        return
Paul Rich's avatar
Paul Rich committed
590

591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
    @exposed
    def find_queue_equivalence_classes(self, reservation_dict,
            active_queue_names, passthrough_blocking_res_list=[]):
        '''Aggregate queues together that can impact eachother in the same
        general pass (both drain and backfill pass) in find_job_location.
        Equivalence classes will then be used in find_job_location to consider
        placement of jobs and resources, in separate passes.  If multiple
        equivalence classes are returned, then they must contain orthogonal sets
        of resources.

        Inputs:
        reservation_dict -- a mapping of active reservations to resrouces.
                            These will block any job in a normal queue.
        active_queue_names -- A list of queues that are currently enabled.
                              Queues that are not in the 'running' state
                              are ignored.
        passthrough_partitions -- Not used on Cray systems currently.  This is
                                  for handling hardware that supports
                                  partitioned interconnect networks.

        Output:
        A list of dictionaries of queues that may impact eachother while
        scheduling resources.

        Side effects:
        None

        Internal Data:
        queue_assignments: a mapping of queues to schedulable locations.

        '''
        equiv = []
623
        node_active_queues = set([])
624
        self.current_equivalence_classes = [] #reverse mapping of queues to nodes
625
626
627
        for node in self.nodes.values():
            if node.managed and node.schedulable:
                #only condiser nodes that we are scheduling.
628
                node_active_queues = set([])
629
630
                for queue in node.queues:
                    if queue in active_queue_names:
631
632
                        node_active_queues.add(queue)
                if node_active_queues == set([]):
633
634
635
636
637
638
639
640
641
                    #this node has nothing active.  The next check can get
                    #expensive, so skip it.
                    continue
            #determine the queues that overlap.  Hardware has to be included so
            #that reservations can be mapped into the equiv classes.
            found_a_match = False
            for e in equiv:
                for queue in node_active_queues:
                    if queue in e['queues']:
642
                        e['data'].add(node.node_id)
643
644
645
646
647
648
                        e['queues'] = e['queues'] | set(node_active_queues)
                        found_a_match = True
                        break
                if found_a_match:
                    break
            if not found_a_match:
649
                equiv.append({'queues': set(node_active_queues),
650
                              'data': set([node.node_id]),
651
                              'reservations': set()})
652
653
654
655
656
        #second pass to merge queue lists based on hardware
        real_equiv = []
        for eq_class in equiv:
            found_a_match = False
            for e in real_equiv:
657
                if e['queues'].intersection(eq_class['queues']):
658
659
660
661
662
663
664
665
666
667
                    e['queues'].update(eq_class['queues'])
                    e['data'].update(eq_class['data'])
                    found_a_match = True
                    break
            if not found_a_match:
                real_equiv.append(eq_class)
        equiv = real_equiv
        #add in reservations:
        for eq_class in equiv:
            for res_name in reservation_dict:
668
669
670
671
                for node_hunk in reservation_dict[res_name].split(":"):
                    for node_id in expand_num_list(node_hunk):
                        if str(node_id) in eq_class['data']:
                            eq_class['reservations'].add(res_name)
672
                            break
673
            #don't send what could be a large block list back in the returun
674
675
676
            for key in eq_class:
                eq_class[key] = list(eq_class[key])
            del eq_class['data']
677
            self.current_equivalence_classes.append(eq_class)
678
679
        return equiv

680
681
682
683
684
685
    @staticmethod
    def _setup_special_locaitons(job):
        forbidden = set([str(loc) for loc in chain_loc_list(job.get('forbidden', []))])
        required = set([str(loc) for loc in chain_loc_list(job.get('required', []))])
        requested_locations = set([str(n) for n in expand_num_list(job['attrs'].get('location', ''))])
        return (forbidden, required, requested_locations)
686

687
    def _assemble_queue_data(self, job, idle_only=True, drain_time=None):
688
689
        '''put together data for a queue, or queue-like reservation structure.

690
691
692
693
694
695
696
        Input:
            job - dictionary of job data.
            idle_only - [default: True] if True, return only idle nodes.
                        Otherwise return nodes in any non-down status.

        return count of idle resources, and a list of valid nodes to run on.
        if idle_only is set to false, returns a set of candidate draining nodes.
697
698
699


        '''
700
701
702
703
        # RESERVATION SUPPORT: Reservation queues are ephemeral, so we will
        # not find the queue normally. In the event of a reservation we'll
        # have to intersect required nodes with the idle and available
        # we also have to forbid a bunch of locations, in  this case.
704
        unavailable_nodes = []
705
        forbidden, required, requested_locations = self._setup_special_locaitons(job)
706
707
708
709
710
711
        requested_loc_in_forbidden = False
        for loc in requested_locations:
            if loc in forbidden:
                #don't spam the logs.
                requested_loc_in_forbidden = True
                break
712
        if job['queue'] not in self.nodes_by_queue.keys():
713
714
715
716
717
            # Either a new queue with no resources, or a possible
            # reservation need to do extra work for a reservation
            node_id_list = list(required - forbidden)
        else:
            node_id_list = list(set(self.nodes_by_queue[job['queue']]) - forbidden)
718
        if requested_locations != set([]): # handle attrs location= requests
719
            job_set = set([str(nid) for nid in requested_locations])
720
            if job['queue'] not in self.nodes_by_queue.keys():
721
722
723
                #we're in a reservation and need to further restrict nodes.
                if job_set <= set(node_id_list):
                    # We are in a reservation there are no forbidden nodes.
724
                    node_id_list = list(requested_locations)
725
726
727
728
                else:
                    # We can't run this job.  Insufficent resources in this
                    # reservation to do so.  Don't risk blocking anything.
                    node_id_list = []
729
            else:
730
731
                #normal queues.  Restrict to the non-reserved nodes.
                if job_set <= set([str(node_id) for node_id in
732
                                    self.nodes_by_queue[job['queue']]]):
733
                    node_id_list = list(requested_locations)
734
735
736
737
738
739
740
                    if not set(node_id_list).isdisjoint(forbidden):
                        # this job has requested locations that are a part of an
                        # active reservation.  Remove locaitons and drop available
                        # nodecount appropriately.
                        node_id_list = list(set(node_id_list) - forbidden)
                else:
                    node_id_list = []
741
742
                    if not requested_loc_in_forbidden:
                        raise ValueError("forbidden locations not in queue")
743
        with self._node_lock:
744
745
746
747
748
749
750
            if idle_only:
                unavailable_nodes = [node_id for node_id in node_id_list
                        if self.nodes[str(node_id)].status not in ['idle']]
            else:
                unavailable_nodes = [node_id for node_id in node_id_list
                        if self.nodes[str(node_id)].status in
                        self.nodes[str(node_id)].DOWN_STATUSES]
751
            if drain_time is not None:
752
                unavailable_nodes.extend([node_id for node_id in node_id_list
753
                    if (self.nodes[str(node_id)].draining and
754
                        (self.nodes[str(node_id)].drain_until - BACKFILL_EPSILON) < int(drain_time))])
755
        for node_id in set(unavailable_nodes):
756
            node_id_list.remove(node_id)
757
        return sorted(node_id_list, key=lambda nid: int(nid))
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776

    def _select_first_nodes(self, job, node_id_list):
        '''Given a list of nids, select the first node count nodes fromt the
        list.  This is the target for alternate allocator replacement.

        Input:
            job - dictionary of job data from the scheduler
            node_id_list - a list of possible candidate nodes

        Return:
            A list of nodes.  [] if insufficient nodes for the allocation.

        Note: hold the node lock while doing this.  We really don't want a
        update to happen while doing this.

        '''
        ret_nodes = []
        with self._node_lock:
            if int(job['nodes']) <= len(node_id_list):
777
778
                node_id_list.sort(key=lambda nid: int(nid))
                ret_nodes = node_id_list[:int(job['nodes'])]
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
        return ret_nodes

    def _associate_and_run_immediate(self, job, resource_until_time, node_id_list):
        '''Given a list of idle node ids, choose a set that can run a job
        immediately, if a set exists in the node_id_list.

        Inputs:
            job - Dictionary of job data
            node_id_list - a list of string node id values

        Side Effects:
            Will reserve resources in ALPS and will set resource reservations on
            allocated nodes.

        Return:
            None if no match, otherwise the pairing of a jobid and set of nids
            that have been allocated to a job.

        '''
        compact_locs = None
        if int(job['nodes']) <= len(node_id_list):
            #this job can be run immediately
            to_alps_list = self._select_first_nodes(job, node_id_list)
            job_locs = self._ALPS_reserve_resources(job, resource_until_time,
                    to_alps_list)
            if job_locs is not None and len(job_locs) == int(job['nodes']):
                compact_locs = compact_num_list(job_locs)
                #temporary reservation until job actually starts
                self.pending_starts[job['jobid']] = resource_until_time
                self.reserve_resources_until(compact_locs, resource_until_time, job['jobid'])
        return compact_locs
810

811
    @locking
812
813
    @exposed
    def find_job_location(self, arg_list, end_times, pt_blocking_locations=[]):
814
815
816
817
818
819
        '''Given a list of jobs, and when jobs are ending, return a set of
        locations mapped to a jobid that can be run.  Also, set up draining
        as-needed to run top-scored jobs and backfill when possible.

        Called once per equivalence class.

820
821
        Args::
            arg_list: A list of dictionaries containning information on jobs to
822
                   cosnider.
823
            end_times: list containing a mapping of locations and the times jobs
824
825
                    runninng on those locations are scheduled to end.  End times
                    are in seconds from Epoch UTC.
826
            pt_blocking_locations: Not used for this system.  Used in partitioned
827
828
829
830
                                interconnect schemes. A list of locations that
                                should not be used to prevent passthrough issues
                                with other scheduler reservations.

831
        Returns:
832
833
834
835
836
837
838
839
840
841
842
843
        A mapping of jobids to locations to run a job to run immediately.

        Side Effects:
        May set draining flags and backfill windows on nodes.
        If nodes are being returned to run, set ALPS reservations on them.

        Notes:
        The reservation set on ALPS resources is uncomfirmed at this point.
        This reservation may timeout.  The forker when it confirms will detect
        this and will re-reserve as needed.  The alps reservation id may change
        in this case post job startup.

844
845
846
        pt_blocking_locations may be used later to block out nodes that are
        impacted by warmswap operations.

847
        This function *DOES NOT* hold the component lock.
848

849
850
851
        '''
        now = time.time()
        resource_until_time = now + TEMP_RESERVATION_TIME
852
        with self._node_lock:
853
            # only valid for this scheduler iteration.
854
855
856
857
            self._clear_draining_for_queues(arg_list[0]['queue'])
            #check if we can run immedaitely, if not drain.  Keep going until all
            #nodes are marked for draining or have a pending run.
            best_match = {} #jobid: list(locations)
858
            for job in arg_list:
859
                label = '%s/%s' % (job['jobid'], job['user'])
860
861
862
                # walltime is in minutes.  We should really fix the storage of
                # that --PMR
                job_endtime = now + (int(job['walltime']) * 60)
863
                try:
864
                    node_id_list = self._assemble_queue_data(job, drain_time=job_endtime)
865
                    available_node_list = self._assemble_queue_data(job, idle_only=False)
866
867
868
                except ValueError:
                    _logger.warning('Job %s: requesting locations that are not in requested queue.',
                            job['jobid'])
869
                    continue
870
                if int(job['nodes']) > len(available_node_list):
871
                    # Insufficient operational nodes for this job at all
872
                    continue
873
874
                elif len(node_id_list) == 0:
                    pass #allow for draining pass to run.
875
                elif int(job['nodes']) <= len(node_id_list):
876
877
                    # enough nodes are in a working state to consider the job.
                    # enough nodes are idle that we can run this job
Paul Rich's avatar
Paul Rich committed
878
879
                    compact_locs = self._associate_and_run_immediate(job,
                            resource_until_time, node_id_list)
880
881
882
                    # do we want to allow multiple placements in a single
                    # pass? That would likely help startup times.
                    if compact_locs is not None:
Paul Rich's avatar
Paul Rich committed
883
                        best_match[job['jobid']] = [compact_locs]
884
                        _logger.info("%s: Job selected for running on nodes  %s",
Paul Rich's avatar
Paul Rich committed
885
                                label, compact_locs)
886
                        break #for now only select one location
887
                if DRAIN_MODE in ['backfill', 'drain-only']:
888
                    # drain sufficient nodes for this job to run
889
                    drain_node_ids = self._select_nodes_for_draining(job,
Paul Rich's avatar
Paul Rich committed
890
                            end_times)
891
892
893
                    if drain_node_ids != []:
                        _logger.info('%s: nodes %s selected for draining.', label,
                                compact_num_list(drain_node_ids))
894
        return best_match
895

896
    def _ALPS_reserve_resources(self, job, new_time, node_id_list):
897
898
899
900
901
902
903
904
905
906
907
908
909
        '''Call ALPS to reserve resrources.  Use their allocator.  We can change
        this later to substitute our own allocator if-needed.

        Input:
        Nodecount - number of nodes to reserve for  a job.

        Returns: a list of locations that ALPS has reserved.

        Side effects:
        Places an ALPS reservation on resources.  Calls reserve resources until
        on the set of nodes, and will mark nodes as allocated.

        '''
910
911
        try:
            res_info = ALPSBridge.reserve(job['user'], job['jobid'],
912
                int(job['nodes']), job['attrs'], node_id_list)
913
        except ALPSBridge.ALPSError as exc:
914
            _logger.warning('unable to reserve resources from ALPS: %s', exc.message)
915
            return None
916
917
        new_alps_res = None
        if res_info is not None:
918
919
            new_alps_res = ALPSReservation(job, res_info, self.nodes)
            self.alps_reservations[job['jobid']] = new_alps_res
920
        return new_alps_res.node_ids
921
922
923
924
925
926
927
928
929
930
931
932

    def _clear_draining_for_queues(self, queue):
        '''Given a list of queues, remove the draining flags on nodes.

        queues - a queue in an equivalence class to consider.  This will clear
        the entire equiv class

        return - none

        Note: does not acquire block lock.  Must be locked externally.

        '''
933
        now = int(time.time())
934
935
936
937
938
        current_queues = []
        for equiv_class in self.current_equivalence_classes:
            if queue in equiv_class['queues']:
                current_queues = equiv_class['queues']
        if current_queues:
939
940
941
942
943
            with self._node_lock:
                for node in self.nodes.values():
                    for q in node.queues:
                        if q in current_queues:
                            node.clear_drain()
944
945
946
947
948
949
950

    def _select_nodes_for_draining(self, job, end_times):
        '''Select nodes to be drainined.  Set backfill windows on draining
        nodes.

        Inputs:
            job - dictionary of job information to consider
Paul Rich's avatar
Paul Rich committed
951
952
953
954
955
            end_times - a list of nodes and their endtimes should be sorted
                        in order of location preference

        Side Effect:
            end_times will be sorted in ascending end-time order
956
957

        Return:
958
959
            List of node ids that have been selected for draining for this job,
            as well as the expected drain time.
960
961

        '''
962
        now = int(time.time())
Paul Rich's avatar
Paul Rich committed
963
        end_times.sort(key=lambda x: int(x[1]))
964
965
        drain_list = []
        candidate_list = []
966
        cleanup_statuses = ['cleanup', 'cleanup-pending']
967
        forbidden, required, requested_locations = self._setup_special_locaitons(job)
968
969
970
        try:
            node_id_list = self._assemble_queue_data(job, idle_only=False)
        except ValueError:
Paul Rich's avatar
Paul Rich committed
971
            _logger.warning('Job %s: requesting locations that are not in queue.', job['jobid'])
972
973
        else:
            with self._node_lock:
974
                drain_time = None
975
976
                candidate_drain_time = None
                # remove the following from the list:
977
978
979
980
981
982
983
                # 1. idle nodes that are already marked for draining.
                # 2. Nodes that are in an in-use status (busy, allocated).
                # 3. Nodes marked for cleanup that are not allocated to a real
                #    jobid. CLEANING_ID is a sentiel jobid value so we can set
                #    a drain window on cleaning nodes easiliy.  Not sure if this
                #    is the right thing to do. --PMR
                candidate_list = []
984
                candidate_list = [nid for nid in node_id_list
985
986
987
                        if (not self.nodes[str(nid)].draining and
                            (self.nodes[str(nid)].status in ['idle']) or
                            (self.nodes[str(nid)].status in cleanup_statuses)
988
989
990
                            )]
                for nid in candidate_list:
                    if self.nodes[str(nid)].status in cleanup_statuses:
991
                        candidate_drain_time = now + CLEANUP_DRAIN_WINDOW
992
                for loc_time in end_times:
993
994
                    running_nodes = [str(nid) for nid in
                            expand_num_list(",".join(loc_time[0]))
995
996
997
                            if ((job['queue'] in self.nodes[str(nid)].queues or
                                nid in required) and
                                not self.nodes[str(nid)].draining)]
998
                    for nid in running_nodes:
999
1000
1001
1002
1003
1004
1005
1006
                        # We set a drain on all running nodes for use in a later
                        # so that we can "favor" draining on the longest
                        # running set of nodes.
                        if (self.nodes[str(nid)].status != 'down' and
                                self.nodes[str(nid)].managed):
                            self.nodes[str(nid)].set_drain(loc_time[1], job['jobid'])
                    candidate_list.extend([nid for nid in running_nodes if
                        self.nodes[str(nid)].draining])
1007
                    candidate_drain_time = int(loc_time[1])
1008
1009
1010
                    if len(candidate_list) >= int(job['nodes']):
                        # Enough nodes have been found to drain for this job
                        break
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
                candidates = set(candidate_list)
                # We need to further restrict this list based on requested
                # location and reservation avoidance data:
                if forbidden != set([]):
                    candidates = candidates.difference(forbidden)
                if requested_locations != set([]):
                    candidates = candidates.intersection(requested_locations)
                candidate_list = list(candidates)
                if len(candidate_list) >= int(job['nodes']):
                    drain_time = candidate_drain_time
1021
1022
1023
                if drain_time is not None:
                    # order the node ids by id and drain-time. Longest drain
                    # first
1024
                    candidate_list.sort(key=lambda nid: int(nid))
1025
                    candidate_list.sort(reverse=True,
1026
                            key=lambda nid: self.nodes[str(nid)].drain_until)
1027
1028
1029
1030
                    drain_list = candidate_list[:int(job['nodes'])]
                    for nid in drain_list:
                        self.nodes[str(nid)].set_drain(drain_time, job['jobid'])
        return drain_list
1031
1032

    @exposed
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
    def reserve_resources_until(self, location, new_time, jobid):
        '''Place, adjust and release resource reservations.

        Input:
            location: the location to reserve [list of nodes]
            new_time: the new end time of a resource reservation
            jobid: the Cobalt jobid that this reservation is for

        Output:
            True if resource reservation is successfully placed.
            Otherwise False.

        Side Effects:
            * Sets/releases reservation on specified node list
            * Sets/releases ALPS reservation.  If set reservation is unconfirmed
              Confirmation must occur a cray_script_forker

        Notes:
            This holds the node data lock while it's running.
1052

1053
        '''
1054
        completed = False
1055
        with self._node_lock:
1056
1057
            succeeded_nodes = []
            failed_nodes = []
Paul Rich's avatar
Paul Rich committed
1058
1059
1060
            #assemble from locaion list:
            exp_location = []
            if isinstance(location, list):
1061
                exp_location = chain_loc_list(location)
Paul Rich's avatar
Paul Rich committed
1062
1063
1064
1065
            elif isinstance(location, str):
                exp_location = expand_num_list(location)
            else:
                raise TypeError("location type is %s.  Must be one of 'list' or 'str'", type(location))
1066
1067
1068
1069
1070
            if new_time is not None:
                #reserve the location. Unconfirmed reservations will have to
                #be lengthened.  Maintain a list of what we have reserved, so we
                #extend on the fly, and so that we don't accidentally get an
                #overallocation/user
Paul Rich's avatar
Paul Rich committed
1071
                for loc in exp_location:
1072
                    # node = self.nodes[self.node_name_to_id[loc]]
Paul Rich's avatar
Paul Rich committed
1073
                    node = self.nodes[str(loc)]
1074
1075
                    try:
                        node.reserve(new_time, jobid=jobid)
Paul Rich's avatar
Paul Rich committed
1076
                        succeeded_nodes.append(int(loc))
1077
1078
                    except Cobalt.Exceptions.ResourceReservationFailure as exc:
                        self.logger.error(exc)
1079
1080
1081
1082
1083
1084
1085
1086
1087
                        failed_nodes.append(loc)
                self.logger.info("job %s: nodes '%s' now reserved until %s",
                    jobid, compact_num_list(succeeded_nodes),
                    time.asctime(time.gmtime(new_time)))
                if failed_nodes != []:
                    self.logger.warning("job %s: failed to reserve nodes '%s'",
                        jobid, compact_num_list(failed_nodes))
                else:
                    completed = True
1088
1089
1090
            else:
                #release the reservation and the underlying ALPS reservation
                #and the reserration on blocks.
Paul Rich's avatar
Paul Rich committed
1091
                for loc in exp_location:
1092
                    # node = self.nodes[self.node_name_to_id[loc]]
Paul Rich's avatar
Paul Rich committed
1093
                    node = self.nodes[str(loc)]
1094
1095
                    try:
                        node.release(user=None, jobid=jobid)
Paul Rich's avatar
Paul Rich committed
1096
                        succeeded_nodes.append(int(loc))
1097
1098
                    except Cobalt.Exceptions.ResourceReservationFailure as exc:
                        self.logger.error(exc)
1099
                        failed_nodes.append(loc)
1100
1101
                    #cleanup pending has to be dealt with.  Do this in UNS for
                    #now
1102
1103
1104
                self.logger.info("job %s:  nodes '%s' released. Cleanup pending.",
                    jobid, compact_num_list(succeeded_nodes))
                if failed_nodes != []:
1105
                    self.logger.warning("job %s: failed to release nodes '%s'",
1106
1107
1108
1109
                        jobid, compact_num_list(failed_nodes))
                else:
                    completed = True
        return completed
1110

1111
1112
1113
1114
1115
    @exposed
    def add_process_groups(self, specs):
        '''Add process groups and start their runs.  Adjust the resource
        reservation time to full run time at this point.

1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
        Args:
            specs: A list of dictionaries that contain information on the Cobalt
                   job required to start the backend process.

        Returns:
            A created process group object.  This is wrapped and sent via XMLRPC
            to the caller.

        Side Effects:
            Invokes a forker component to run a user script.  In the event of a
            fatal startup error, will release resource reservations.

1128
1129
1130
1131
        '''
        start_apg_timer = int(time.time())

        for spec in specs: