CraySystem.py 65.3 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.Exceptions import ComponentLookupError
Paul Rich's avatar
Paul Rich committed
19
from Cobalt.Exceptions import JobNotInteractive
20
from Cobalt.Components.system.ALPSProcessGroup import ALPSProcessGroup
21
from Cobalt.Util import compact_num_list, expand_num_list
22
from Cobalt.Util import init_cobalt_config, get_config_option
23

24
25
_logger = logging.getLogger(__name__)

26
init_cobalt_config()
27

28
29
30
31
32
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
33
34
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
35
APKILL_CMD = get_config_option('alps', 'apkill', '/opt/cray/alps/default/bin/apkill')
36
DRAIN_MODE = get_config_option('system', 'drain_mode', 'first-fit')
37
38
39
#cleanup time in seconds
CLEANUP_DRAIN_WINDOW = get_config_option('system', 'cleanup_drain_window', 300)

40
41
42
#Epsilon for backfilling.  This system does not do this on a per-node basis.
BACKFILL_EPSILON = int(get_config_option('system', 'backfill_epsilon', 120))

Paul Rich's avatar
Paul Rich committed
43
DRAIN_MODES = ['first-fit', 'backfill']
44
CLEANING_ID = -1
45

46
47
48
49
50
51
52
53
54
55
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

56

57
class CraySystem(BaseSystem):
58
59
60
    '''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.
61

62
    '''
63
64
    name = "system"
    implementation = "alps_system"
65
    logger = _logger
66

67
    def __init__(self, *args, **kwargs):
68
69
70
71
        '''Initialize system.  Read initial states from bridge.
        Get current state

        '''
72
        start_time = time.time()
73
        super(CraySystem, self).__init__(*args, **kwargs)
74
        _logger.info('BASE SYSTEM INITIALIZED')
75
76
77
78
79
80
81
82
83
84
        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.

        '''
85
86
87
88
89
        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)
90
        #initilaize bridge.
91
92
93
94
        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
95
96
            # These attempts may fail due to system_script_forker not being up.
            # We don't want to trash the statefile in this case.
97
98
99
100
            try:
                ALPSBridge.init_bridge()
            except ALPSBridge.BridgeError:
                _logger.error('Bridge Initialization failed.  Retrying.')
101
102
103
104
                Cobalt.Util.sleep(10)
            except ComponentLookupError:
                _logger.warning('Error reaching forker.  Retrying.')
                Cobalt.Util.sleep(10)
105
106
107
            else:
                bridge_pending = False
                _logger.info('BRIDGE INITIALIZED')
108
        #process manager setup
109
        if spec is None:
Paul Rich's avatar
Paul Rich committed
110
            self.process_manager = ProcessGroupManager(pgroup_type=ALPSProcessGroup)
111
        else:
Paul Rich's avatar
Paul Rich committed
112
            self.process_manager = ProcessGroupManager(pgroup_type=ALPSProcessGroup).__setstate__(spec['process_manager'])
Paul Rich's avatar
Paul Rich committed
113
            self.logger.debug('pg type %s', self.process_manager.process_groups.item_cls)
114
115
        #self.process_manager.forkers.append('alps_script_forker')
        self.process_manager.update_launchers()
Paul Rich's avatar
Paul Rich committed
116
        self.pending_start_timeout = PENDING_STARTUP_TIMEOUT
117
118
        _logger.info('PROCESS MANAGER INTIALIZED')
        #resource management setup
119
        self.nodes = {} #cray node_id: CrayNode
120
121
122
123
        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']
124
        self._init_nodes_and_reservations()
125
126
127
        if spec is not None:
            node_info = spec.get('node_info', {})
            for nid, node in node_info.items():
128
129
130
131
                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
132
133
134
        #storage for pending job starts.  Allows us to handle slow starts vs
        #user deletes
        self.pending_starts = {} #jobid: time this should be cleared.
135
        self.nodes_by_queue = {} #queue:[node_ids]
136
        #populate initial state
137
        #state update thread and lock
138
        self._node_lock = threading.RLock()
139
        self._gen_node_to_queue()
Paul Rich's avatar
Paul Rich committed
140
        self.node_update_thread = thread.start_new_thread(self._run_update_state, tuple())
141
        _logger.info('UPDATE THREAD STARTED')
142
        self.current_equivalence_classes = []
Paul Rich's avatar
Paul Rich committed
143
        self.killing_jobs = {}
144
145
146
147
        #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
148

149
150
151
152
153
154
155
156
    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
157
        state['node_info'] = self.nodes
158
159
160
161
162
163
164
        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)
165
        _logger.info('ALPS SYSTEM COMPONENT READY TO RUN')
166
        _logger.info('Reinitilaization complete in %s sec.', (time.time() -
167
                start_time))
168

169
170
171
172
173
    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)
174

175
176
177
178
    def _init_nodes_and_reservations(self):
        '''Initialize nodes from ALPS bridge data'''

        retnodes = {}
179
180
181
        pending = True
        while pending:
            try:
182
183
184
185
186
                # 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
187
                _logger.info('INVENTORY FETCHED')
188
                system = ALPSBridge.extract_system_node_data(ALPSBridge.system())
Paul Rich's avatar
Paul Rich committed
189
                _logger.info('SYSTEM DATA FETCHED')
190
                reservations = ALPSBridge.fetch_reservations()
Paul Rich's avatar
Paul Rich committed
191
                _logger.info('ALPS RESERVATION DATA FETCHED')
192
                # reserved_nodes = ALPSBridge.reserved_nodes()
193
194
195
            except Exception:
                #don't crash out here.  That may trash a statefile.
                _logger.error('Possible transient encountered during initialization. Retrying.',
196
                        exc_info=True)
197
198
199
200
                Cobalt.Util.sleep(10)
            else:
                pending = False

201
        self._assemble_nodes(inventory, system)
202
203
        #Reversing the node name to id lookup is going to save a lot of cycles.
        for node in self.nodes.values():
204
            self.node_name_to_id[node.name] = node.node_id
205
206
        _logger.info('NODE INFORMATION INITIALIZED')
        _logger.info('ALPS REPORTS %s NODES', len(self.nodes))
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
        # 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'])
222
223
224
225
            # 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'
226
227
228
229
230
231
            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
232

233
    def _gen_node_to_queue(self):
234
        '''(Re)Generate a mapping for fast lookup of node-id's to queues.'''
235
        with self._node_lock:
236
            self.nodes_by_queue = {}
237
238
239
            for node in self.nodes.values():
                for queue in node.queues:
                    if queue in self.nodes_by_queue.keys():
240
                        self.nodes_by_queue[queue].add(node.node_id)
241
                    else:
242
                        self.nodes_by_queue[queue] = set([node.node_id])
243

244
    @exposed
245
    def get_nodes(self, as_dict=False, node_ids=None, params=None, as_json=False):
246
247
        '''fetch the node dictionary.

248
249
250
251
252
253
254
            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.
255
256
257
258
259

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

        '''
260
261
262
263
264
265
266
267
268
        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()))
269
        else:
270
271
272
273
274
            node_info = dict(filter(node_filter, self.nodes.items()))
        if as_json:
            return json.dumps(node_info)
        return node_info

275
276
277
    def _run_update_state(self):
        '''automated node update functions on the update timer go here.'''
        while True:
278
279
280
281
282
283
284
285
286
            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)
287

288
289
290
291
292
293
294
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
    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
323
        new_node.managed = True
324
325
326
        self.nodes[str(nid)] = new_node
        self.logger.warning('Node %s added to tracking.', nid)

327
328
329
330
331
332
    @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
333
334
335
        #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.
336
        now = time.time()
Paul Rich's avatar
Paul Rich committed
337
338
339
340
341
342
343
344
        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
345
        self.check_killing_aprun()
346
        with self._node_lock:
Paul Rich's avatar
Paul Rich committed
347
            fetch_time_start = time.time()
348
            try:
349
350
                #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
351
                #inventory = ALPSBridge.fetch_inventory(resinfo=True) #This is a full-refresh,
352
                #determine if summary may be used under normal operation
353
                #updated for >= 1.6 interface
Paul Rich's avatar
Paul Rich committed
354
355
356
                inven_nodes = ALPSBridge.extract_system_node_data(ALPSBridge.system())
                reservations = ALPSBridge.fetch_reservations()
                #reserved_nodes = ALPSBridge.reserved_nodes()
357
358
359
360
            except (ALPSBridge.ALPSError, ComponentLookupError):
                _logger.warning('Error contacting ALPS for state update.  Aborting this update',
                        exc_info=True)
                return
361
            inven_reservations = reservations.get('reservations', [])
Paul Rich's avatar
Paul Rich committed
362
            fetch_time_start = time.time()
363
            #_logger.debug("time in ALPS fetch: %s seconds", (time.time() - fetch_time_start))
364
            start_time = time.time()
Paul Rich's avatar
Paul Rich committed
365
            self._detect_rereservation(inven_reservations)
366
367
368
369
370
371
            # 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
372
            res_jobid_to_delete = []
373
374
375
376
377
378
379
380
381
            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
382
383
                if alps_res.jobid in self.pending_starts.keys():
                    continue #Get to this only after timeout happens
Paul Rich's avatar
Paul Rich committed
384
                #find alps_id associated reservation
385
                if int(alps_res.alps_res_id) not in current_alps_res_ids:
Paul Rich's avatar
Paul Rich committed
386
                    for node_id in alps_res.node_ids:
387
                        if not self.nodes[str(node_id)].reserved:
Paul Rich's avatar
Paul Rich committed
388
                            #pending hardware status update
389
                            self.nodes[str(node_id)].status = 'idle'
Paul Rich's avatar
Paul Rich committed
390
                    res_jobid_to_delete.append(alps_res.jobid)
391
392
                    _logger.info('Nodes %s cleanup complete.',
                            compact_num_list(alps_res.node_ids))
Paul Rich's avatar
Paul Rich committed
393
            for jobid in res_jobid_to_delete:
394
395
                _logger.info('%s: ALPS reservation for this job complete.', jobid)
                del self.alps_reservations[str(jobid)]
Paul Rich's avatar
Paul Rich committed
396
397
            #process group should already be on the way down since cqm released the
            #resource reservation
398
399
            cleanup_nodes = [node for node in self.nodes.values()
                             if node.status == 'cleanup-pending']
Paul Rich's avatar
Paul Rich committed
400
            #If we have a block marked for cleanup, send a release message.
Paul Rich's avatar
Paul Rich committed
401
402
            released_res_jobids = []
            for node in cleanup_nodes:
403
                for alps_res in self.alps_reservations.values():
Paul Rich's avatar
Paul Rich committed
404
                    if (alps_res.jobid not in released_res_jobids and
405
406
                            str(node.node_id) in alps_res.node_ids):
                        #send only one release per iteration
Paul Rich's avatar
Paul Rich committed
407
408
409
410
                        apids = alps_res.release()
                        if apids is not None:
                            for apid in apids:
                                self.signal_aprun(apid)
411
                        released_res_jobids.append(alps_res.jobid)
Paul Rich's avatar
Paul Rich committed
412
413

        #find hardware status
414
415
416
            #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
417
            for inven_node in inven_nodes.values():
418
419
                if self.nodes.has_key(str(inven_node['node_id'])):
                    node = self.nodes[str(inven_node['node_id'])]
420
                    node.role = inven_node['role'].upper()
Paul Rich's avatar
Paul Rich committed
421
                    if node.reserved:
422
423
                        #node marked as reserved.
                        if self.alps_reservations.has_key(str(node.reserved_jobid)):
Paul Rich's avatar
Paul Rich committed
424
425
                            node.status = 'busy'
                        else:
426
427
428
429
430
431
                            # 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
432
                    else:
433
                        node.status = inven_node['state'].upper()
434
435
                        if node.role.upper() not in ['BATCH'] and node.status is 'idle':
                            node.status = 'alps-interactive'
436
                else:
437
438
439
440
441
442
443
444
445
446
447
448
449
                    # 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
450
451
452
            #should down win over running in terms of display?
            #keep node that are marked for cleanup still in cleanup
            for node in cleanup_nodes:
453
                node.status = 'cleanup-pending'
454
        #_logger.debug("time in UNS lock: %s seconds", (time.time() - start_time))
455
456
        return

Paul Rich's avatar
Paul Rich committed
457
458
459
460
461
462
463
464
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
    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
506
                        apids = tracked_res.release()
Paul Rich's avatar
Paul Rich committed
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
                    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
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
        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
548
            _logger.info("killing backend ALPS application_id: %s", aprun_id)
Paul Rich's avatar
Paul Rich committed
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
        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.

        '''
564

Paul Rich's avatar
Paul Rich committed
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
        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
581
        return
Paul Rich's avatar
Paul Rich committed
582

583
584
585
586
587
588
589
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
    @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 = []
615
        node_active_queues = set([])
616
        self.current_equivalence_classes = [] #reverse mapping of queues to nodes
617
618
619
        for node in self.nodes.values():
            if node.managed and node.schedulable:
                #only condiser nodes that we are scheduling.
620
                node_active_queues = set([])
621
622
                for queue in node.queues:
                    if queue in active_queue_names:
623
624
                        node_active_queues.add(queue)
                if node_active_queues == set([]):
625
626
627
628
629
630
631
632
633
                    #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']:
634
                        e['data'].add(node.node_id)
635
636
637
638
639
640
                        e['queues'] = e['queues'] | set(node_active_queues)
                        found_a_match = True
                        break
                if found_a_match:
                    break
            if not found_a_match:
641
                equiv.append({'queues': set(node_active_queues),
642
                              'data': set([node.node_id]),
643
                              'reservations': set()})
644
645
646
647
648
        #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:
649
                if e['queues'].intersection(eq_class['queues']):
650
651
652
653
654
655
656
657
658
659
                    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:
660
661
662
663
                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)
664
                            break
665
            #don't send what could be a large block list back in the returun
666
667
668
            for key in eq_class:
                eq_class[key] = list(eq_class[key])
            del eq_class['data']
669
            self.current_equivalence_classes.append(eq_class)
670
671
        return equiv

672
673
674
675
676
677
    @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)
678

679
    def _assemble_queue_data(self, job, idle_only=True, drain_time=None):
680
681
        '''put together data for a queue, or queue-like reservation structure.

682
683
684
685
686
687
688
        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.
689
690
691


        '''
692
693
694
695
        # 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.
696
        unavailable_nodes = []
697
        forbidden, required, requested_locations = self._setup_special_locaitons(job)
698
699
700
701
702
703
        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
704
        if job['queue'] not in self.nodes_by_queue.keys():
705
706
707
708
709
            # 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)
710
        if requested_locations != set([]): # handle attrs location= requests
711
            job_set = set([str(nid) for nid in requested_locations])
712
            if job['queue'] not in self.nodes_by_queue.keys():
713
714
715
                #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.
716
                    node_id_list = list(requested_locations)
717
718
719
720
                else:
                    # We can't run this job.  Insufficent resources in this
                    # reservation to do so.  Don't risk blocking anything.
                    node_id_list = []
721
            else:
722
723
                #normal queues.  Restrict to the non-reserved nodes.
                if job_set <= set([str(node_id) for node_id in
724
                                    self.nodes_by_queue[job['queue']]]):
725
                    node_id_list = list(requested_locations)
726
727
728
729
730
731
732
                    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 = []
733
734
                    if not requested_loc_in_forbidden:
                        raise ValueError("forbidden locations not in queue")
735
        with self._node_lock:
736
737
738
739
740
741
742
            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]
743
            if drain_time is not None:
744
                unavailable_nodes.extend([node_id for node_id in node_id_list
745
                    if (self.nodes[str(node_id)].draining and
746
                        (self.nodes[str(node_id)].drain_until - BACKFILL_EPSILON) < int(drain_time))])
747
        for node_id in set(unavailable_nodes):
748
            node_id_list.remove(node_id)
749
        return sorted(node_id_list, key=lambda nid: int(nid))
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768

    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):
769
770
                node_id_list.sort(key=lambda nid: int(nid))
                ret_nodes = node_id_list[:int(job['nodes'])]
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
        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
802

803
    @locking
804
805
    @exposed
    def find_job_location(self, arg_list, end_times, pt_blocking_locations=[]):
806
807
808
809
810
811
        '''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.

812
813
        Args::
            arg_list: A list of dictionaries containning information on jobs to
814
                   cosnider.
815
            end_times: list containing a mapping of locations and the times jobs
816
817
                    runninng on those locations are scheduled to end.  End times
                    are in seconds from Epoch UTC.
818
            pt_blocking_locations: Not used for this system.  Used in partitioned
819
820
821
822
                                interconnect schemes. A list of locations that
                                should not be used to prevent passthrough issues
                                with other scheduler reservations.

823
        Returns:
824
825
826
827
828
829
830
831
832
833
834
835
        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.

836
837
838
        pt_blocking_locations may be used later to block out nodes that are
        impacted by warmswap operations.

839
        This function *DOES NOT* hold the component lock.
840

841
842
843
        '''
        now = time.time()
        resource_until_time = now + TEMP_RESERVATION_TIME
844
        with self._node_lock:
845
            # only valid for this scheduler iteration.
846
847
848
849
            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)
850
            for job in arg_list:
851
                label = '%s/%s' % (job['jobid'], job['user'])
852
853
854
                # walltime is in minutes.  We should really fix the storage of
                # that --PMR
                job_endtime = now + (int(job['walltime']) * 60)
855
                try:
856
                    node_id_list = self._assemble_queue_data(job, drain_time=job_endtime)
857
                    available_node_list = self._assemble_queue_data(job, idle_only=False)
858
859
860
                except ValueError:
                    _logger.warning('Job %s: requesting locations that are not in requested queue.',
                            job['jobid'])
861
                    continue
862
                if int(job['nodes']) > len(available_node_list):
863
                    # Insufficient operational nodes for this job at all
864
                    continue
865
866
                elif len(node_id_list) == 0:
                    pass #allow for draining pass to run.
867
                elif int(job['nodes']) <= len(node_id_list):
868
869
                    # 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
870
871
                    compact_locs = self._associate_and_run_immediate(job,
                            resource_until_time, node_id_list)
872
873
874
                    # 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
875
                        best_match[job['jobid']] = [compact_locs]
876
                        _logger.info("%s: Job selected for running on nodes  %s",
Paul Rich's avatar
Paul Rich committed
877
                                label, compact_locs)
878
                        break #for now only select one location
879
                if DRAIN_MODE in ['backfill', 'drain-only']:
880
                    # drain sufficient nodes for this job to run
881
                    drain_node_ids = self._select_nodes_for_draining(job,
Paul Rich's avatar
Paul Rich committed
882
                            end_times)
883
884
885
                    if drain_node_ids != []:
                        _logger.info('%s: nodes %s selected for draining.', label,
                                compact_num_list(drain_node_ids))
886
        return best_match
887

888
    def _ALPS_reserve_resources(self, job, new_time, node_id_list):
889
890
891
892
893
894
895
896
897
898
899
900
901
        '''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.

        '''
902
903
        try:
            res_info = ALPSBridge.reserve(job['user'], job['jobid'],
904
                int(job['nodes']), job['attrs'], node_id_list)
905
        except ALPSBridge.ALPSError as exc:
906
            _logger.warning('unable to reserve resources from ALPS: %s', exc.message)
907
            return None
908
909
        new_alps_res = None
        if res_info is not None:
910
911
            new_alps_res = ALPSReservation(job, res_info, self.nodes)
            self.alps_reservations[job['jobid']] = new_alps_res
912
        return new_alps_res.node_ids
913
914
915
916
917
918
919
920
921
922
923
924

    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.

        '''
925
        now = int(time.time())
926
927
928
929
930
        current_queues = []
        for equiv_class in self.current_equivalence_classes:
            if queue in equiv_class['queues']:
                current_queues = equiv_class['queues']
        if current_queues:
931
932
933
934
935
            with self._node_lock:
                for node in self.nodes.values():
                    for q in node.queues:
                        if q in current_queues:
                            node.clear_drain()
936
937
938
939
940
941
942

    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
943
944
945
946
947
            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
948
949

        Return:
950
951
            List of node ids that have been selected for draining for this job,
            as well as the expected drain time.
952
953

        '''
954
        now = int(time.time())
Paul Rich's avatar
Paul Rich committed
955
        end_times.sort(key=lambda x: int(x[1]))
956
957
        drain_list = []
        candidate_list = []
958
        cleanup_statuses = ['cleanup', 'cleanup-pending']
959
        forbidden, required, requested_locations = self._setup_special_locaitons(job)
960
961
962
        try:
            node_id_list = self._assemble_queue_data(job, idle_only=False)
        except ValueError:
Paul Rich's avatar
Paul Rich committed
963
            _logger.warning('Job %s: requesting locations that are not in queue.', job['jobid'])
964
965
        else:
            with self._node_lock:
966
                drain_time = None
967
968
                candidate_drain_time = None
                # remove the following from the list:
969
970
971
972
973
974
975
                # 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 = []
976
                candidate_list = [nid for nid in node_id_list
977
978
979
                        if (not self.nodes[str(nid)].draining and
                            (self.nodes[str(nid)].status in ['idle']) or
                            (self.nodes[str(nid)].status in cleanup_statuses)
980
981
982
                            )]
                for nid in candidate_list:
                    if self.nodes[str(nid)].status in cleanup_statuses:
983
                        candidate_drain_time = now + CLEANUP_DRAIN_WINDOW
984
                for loc_time in end_times:
985
986
                    running_nodes = [str(nid) for nid in
                            expand_num_list(",".join(loc_time[0]))
987
988
989
                            if ((job['queue'] in self.nodes[str(nid)].queues or
                                nid in required) and
                                not self.nodes[str(nid)].draining)]
990
                    for nid in running_nodes:
991
992
993
994
995
996
997
998
                        # 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])
999
                    candidate_drain_time = int(loc_time[1])
1000
1001
1002
                    if len(candidate_list) >= int(job['nodes']):
                        # Enough nodes have been found to drain for this job
                        break
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
                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
1013
1014
1015
                if drain_time is not None:
                    # order the node ids by id and drain-time. Longest drain
                    # first
1016
                    candidate_list.sort(key=lambda nid: int(nid))
1017
                    candidate_list.sort(reverse=True,
1018
                            key=lambda nid: self.nodes[str(nid)].drain_until)
1019
1020
1021
1022
                    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
1023
1024

    @exposed
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
    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.
1044

1045
        '''
1046
        completed = False
1047
        with self._node_lock:
1048
1049
            succeeded_nodes = []
            failed_nodes = []
Paul Rich's avatar
Paul Rich committed
1050
1051
1052
            #assemble from locaion list:
            exp_location = []
            if isinstance(location, list):
1053
                exp_location = chain_loc_list(location)
Paul Rich's avatar
Paul Rich committed
1054
1055
1056
1057
            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))
1058
1059
1060
1061
1062
            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
1063
                for loc in exp_location:
1064
                    # node = self.nodes[self.node_name_to_id[loc]]
Paul Rich's avatar
Paul Rich committed
1065
                    node = self.nodes[str(loc)]
1066
1067
                    try:
                        node.reserve(new_time, jobid=jobid)
Paul Rich's avatar
Paul Rich committed
1068
                        succeeded_nodes.append(int(loc))
1069
1070
                    except Cobalt.Exceptions.ResourceReservationFailure as exc:
                        self.logger.error(exc)
1071
1072
1073
1074
1075
1076
1077
1078
1079
                        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
1080
1081
1082
            else:
                #release the reservation and the underlying ALPS reservation
                #and the reserration on blocks.
Paul Rich's avatar
Paul Rich committed
1083
                for loc in exp_location:
1084
                    # node = self.nodes[self.node_name_to_id[loc]]
Paul Rich's avatar
Paul Rich committed
1085
                    node = self.nodes[str(loc)]
1086
1087
                    try:
                        node.release(user=None, jobid=jobid)
Paul Rich's avatar
Paul Rich committed
1088
                        succeeded_nodes.append(int(loc))
1089
1090
                    except Cobalt.Exceptions.ResourceReservationFailure as exc:
                        self.logger.error(exc)
1091
                        failed_nodes.append(loc)
1092
1093
                    #cleanup pending has to be dealt with.  Do this in UNS for
                    #now
1094
1095
1096
                self.logger.info("job %s:  nodes '%s' released. Cleanup pending.",
                    jobid, compact_num_list(succeeded_nodes))
                if failed_nodes != []:
1097
                    self.logger.warning("job %s: failed to release nodes '%s'",
1098
1099
1100
1101
                        jobid, compact_num_list(failed_nodes))
                else:
                    completed = True
        return completed
1102

1103
1104
1105
1106
1107
    @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.

1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
        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.

1120
1121
1122
1123
        '''
        start_apg_timer = int(time.time())

        for spec in specs:
1124
            spec['forker'] = None
1125
1126
1127
            alps_res = self.alps_reservations.get(str(spec['jobid']), None)
            if alps_res is not None:
                spec['alps_res_id'] = alps_res.alps_res_id
1128
1129
1130
1131
1132
1133
1134
            new_pgroups = self.process_manager.init_groups(specs)
        for pgroup in new_pgroups:
            _logger.info('%s: process group %s created to track job status',
                    pgroup.label, pgroup.id)
            #check resource reservation, and attempt to start.  If there's a
            #failure here, set exit status in process group to a sentinel value.
            try:
1135
                started = self.process_manager.start_groups([pgroup.id])
1136
            except ComponentLookupError:
Paul Rich's avatar