CraySystem.py 71 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
import ConfigParser
11 12
import Cobalt.Util
import Cobalt.Components.system.AlpsBridge as ALPSBridge
13
from Cobalt.Components.base import Component, exposed, automatic, query, locking
14 15
from Cobalt.Components.system.base_system import BaseSystem
from Cobalt.Components.system.CrayNode import CrayNode
16
from Cobalt.Components.system.base_pg_manager import ProcessGroupManager
17
from Cobalt.Components.system.ALPSProcessGroup import ALPSProcessGroup
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.Exceptions import JobValidationError
22
from Cobalt.DataTypes.ProcessGroup import ProcessGroup
23
from Cobalt.Util import compact_num_list, expand_num_list
24
from Cobalt.Util import init_cobalt_config, get_config_option
25

26
_logger = logging.getLogger(__name__)
27
init_cobalt_config()
28

29 30
UPDATE_THREAD_TIMEOUT = int(get_config_option('alpssystem', 'update_thread_timeout', 10))
TEMP_RESERVATION_TIME = int(get_config_option('alpssystem', 'temp_reservation_time', 300))
31
SAVE_ME_INTERVAL = float(get_config_option('alpsssytem', 'save_me_interval', 10.0))
32 33
#default 20 minutes to account for boot.
PENDING_STARTUP_TIMEOUT = float(get_config_option('alpssystem', 'pending_startup_timeout', 1200))
Paul Rich's avatar
Paul Rich committed
34
APKILL_CMD = get_config_option('alps', 'apkill', '/opt/cray/alps/default/bin/apkill')
35
DRAIN_MODE = get_config_option('system', 'drain_mode', 'first-fit')
36 37 38
#cleanup time in seconds
CLEANUP_DRAIN_WINDOW = get_config_option('system', 'cleanup_drain_window', 300)

39 40
#Epsilon for backfilling.  This system does not do this on a per-node basis.
BACKFILL_EPSILON = int(get_config_option('system', 'backfill_epsilon', 120))
41
ELOGIN_HOSTS = [host for host in get_config_option('system', 'elogin_hosts', '').split(':')]
42 43
if ELOGIN_HOSTS == ['']:
    ELOGIN_HOSTS = []
Paul Rich's avatar
Paul Rich committed
44
DRAIN_MODES = ['first-fit', 'backfill']
45
CLEANING_ID = -1
46 47
DEFAULT_MCDRAM_MODE = get_config_option('alpssystem', 'default_mcdram_mode', 'cache')
DEFAULT_NUMA_MODE = get_config_option('alpssystem', 'default_numa_mode', 'quad')
48
MCDRAM_TO_HBMCACHEPCT = {'flat':'0', 'cache':'100', 'split':'25', 'equal':'50', '0':'0', '25':'25', '50':'50', '100':'100'}
49 50
VALID_MCDRAM_MODES = ['flat', 'cache', 'split', 'equal', '0', '25', '50', '100']
VALID_NUMA_MODES = ['a2a', 'hemi', 'quad', 'snc2', 'snc4']
51 52


53 54 55 56 57 58 59 60 61 62
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

63

64
class CraySystem(BaseSystem):
65 66 67
    '''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.
68

69
    '''
70 71
    name = "system"
    implementation = "alps_system"
72
    logger = _logger
73

74
    def __init__(self, *args, **kwargs):
75 76 77 78
        '''Initialize system.  Read initial states from bridge.
        Get current state

        '''
79
        start_time = time.time()
80
        super(CraySystem, self).__init__(*args, **kwargs)
81
        _logger.info('BASE SYSTEM INITIALIZED')
82 83 84 85 86 87 88 89 90 91
        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.

        '''
92 93 94 95 96
        try:
            self.system_size = int(get_config_option('system', 'size'))
        except (ConfigParser.NoSectionError, ConfigParser.NoOptionError):
            _logger.critical('ALPS SYSTEM: ABORT STARTUP: System size must be specified in the [system] section of the cobalt configuration file.')
            sys.exit(1)
97 98 99 100 101
        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)
102
        #initilaize bridge.
103 104 105 106
        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
107 108
            # These attempts may fail due to system_script_forker not being up.
            # We don't want to trash the statefile in this case.
109 110 111 112
            try:
                ALPSBridge.init_bridge()
            except ALPSBridge.BridgeError:
                _logger.error('Bridge Initialization failed.  Retrying.')
113 114 115 116
                Cobalt.Util.sleep(10)
            except ComponentLookupError:
                _logger.warning('Error reaching forker.  Retrying.')
                Cobalt.Util.sleep(10)
117 118 119
            else:
                bridge_pending = False
                _logger.info('BRIDGE INITIALIZED')
120
        #process manager setup
121
        if spec is None:
Paul Rich's avatar
Paul Rich committed
122
            self.process_manager = ProcessGroupManager(pgroup_type=ALPSProcessGroup)
123
        else:
Paul Rich's avatar
Paul Rich committed
124
            self.process_manager = ProcessGroupManager(pgroup_type=ALPSProcessGroup).__setstate__(spec['process_manager'])
Paul Rich's avatar
Paul Rich committed
125
            self.logger.debug('pg type %s', self.process_manager.process_groups.item_cls)
126 127
        #self.process_manager.forkers.append('alps_script_forker')
        self.process_manager.update_launchers()
Paul Rich's avatar
Paul Rich committed
128
        self.pending_start_timeout = PENDING_STARTUP_TIMEOUT
129 130
        _logger.info('PROCESS MANAGER INTIALIZED')
        #resource management setup
131
        self.nodes = {} #cray node_id: CrayNode
132 133 134 135
        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']
136
        self._init_nodes_and_reservations()
137 138 139
        if spec is not None:
            node_info = spec.get('node_info', {})
            for nid, node in node_info.items():
140 141 142 143
                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
144 145 146
        #storage for pending job starts.  Allows us to handle slow starts vs
        #user deletes
        self.pending_starts = {} #jobid: time this should be cleared.
147
        self.nodes_by_queue = {} #queue:[node_ids]
148
        #populate initial state
149
        #state update thread and lock
150
        self._node_lock = threading.RLock()
151
        self._gen_node_to_queue()
Paul Rich's avatar
Paul Rich committed
152
        self.node_update_thread = thread.start_new_thread(self._run_update_state, tuple())
153
        _logger.info('UPDATE THREAD STARTED')
154
        self.current_equivalence_classes = []
Paul Rich's avatar
Paul Rich committed
155
        self.killing_jobs = {}
156 157 158 159
        #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
160

161 162 163 164 165 166 167 168
    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
169
        state['node_info'] = self.nodes
170 171 172 173 174 175 176
        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)
177
        _logger.info('ALPS SYSTEM COMPONENT READY TO RUN')
178
        _logger.info('Reinitilaization complete in %s sec.', (time.time() -
179
                start_time))
180

181 182 183 184 185
    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)
186

187 188 189 190
    def _init_nodes_and_reservations(self):
        '''Initialize nodes from ALPS bridge data'''

        retnodes = {}
191 192 193
        pending = True
        while pending:
            try:
194 195 196 197 198
                # 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
199
                _logger.info('INVENTORY FETCHED')
200
                system = ALPSBridge.extract_system_node_data(ALPSBridge.system())
Paul Rich's avatar
Paul Rich committed
201
                _logger.info('SYSTEM DATA FETCHED')
202
                reservations = ALPSBridge.fetch_reservations()
Paul Rich's avatar
Paul Rich committed
203
                _logger.info('ALPS RESERVATION DATA FETCHED')
204
                # reserved_nodes = ALPSBridge.reserved_nodes()
205 206 207
            except Exception:
                #don't crash out here.  That may trash a statefile.
                _logger.error('Possible transient encountered during initialization. Retrying.',
208
                        exc_info=True)
209 210 211 212
                Cobalt.Util.sleep(10)
            else:
                pending = False

213
        self._assemble_nodes(inventory, system)
214 215
        #Reversing the node name to id lookup is going to save a lot of cycles.
        for node in self.nodes.values():
216
            self.node_name_to_id[node.name] = node.node_id
217 218
        _logger.info('NODE INFORMATION INITIALIZED')
        _logger.info('ALPS REPORTS %s NODES', len(self.nodes))
219 220 221 222 223 224 225 226 227 228 229 230 231 232
        # 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():
233 234 235 236 237 238
            nodes[node_id].attributes.update(nodespec['attrs'])
            # 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'
            nodes[node_id].status = nodespec['state']
239 240 241 242 243
        self.nodes = nodes

    def _assemble_reservations(self, reservations, reserved_nodes):
        # FIXME: we can recover reservations now.  Implement this.
        pass
244

245
    def _gen_node_to_queue(self):
246
        '''(Re)Generate a mapping for fast lookup of node-id's to queues.'''
247
        with self._node_lock:
248
            self.nodes_by_queue = {}
249 250 251
            for node in self.nodes.values():
                for queue in node.queues:
                    if queue in self.nodes_by_queue.keys():
252
                        self.nodes_by_queue[queue].add(node.node_id)
253
                    else:
254
                        self.nodes_by_queue[queue] = set([node.node_id])
255

256
    @exposed
257
    def get_nodes(self, as_dict=False, node_ids=None, params=None, as_json=False):
258 259
        '''fetch the node dictionary.

260 261 262 263 264 265 266
            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.
267 268 269 270 271

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

        '''
272 273 274 275 276 277 278 279 280
        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()))
281
        else:
282 283 284 285 286
            node_info = dict(filter(node_filter, self.nodes.items()))
        if as_json:
            return json.dumps(node_info)
        return node_info

287 288
    def _run_update_state(self):
        '''automated node update functions on the update timer go here.'''
289 290

        def _run_and_wrap(func):
291
            try:
292
                func()
293
            except Exception:
294
                # Prevent this thread from dying.
295
                _logger.critical('Error in _run_update_state', exc_info=True)
296 297 298 299 300 301 302 303

        while True:
            # Each of these is wrapped in it's own log-and-preserve block.
            # The outer try is there to ensure the thread update timeout happens.
            _run_and_wrap(self.process_manager.update_launchers)
            _run_and_wrap(self.update_node_state)
            _run_and_wrap(self._get_exit_status)
            Cobalt.Util.sleep(UPDATE_THREAD_TIMEOUT)
304

305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339
    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
340
        new_node.managed = True
341 342 343
        self.nodes[str(nid)] = new_node
        self.logger.warning('Node %s added to tracking.', nid)

344 345 346 347 348 349
    @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
350 351 352
        #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.
353
        now = time.time()
Paul Rich's avatar
Paul Rich committed
354 355 356 357 358 359 360 361
        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
362
        self.check_killing_aprun()
363
        with self._node_lock:
Paul Rich's avatar
Paul Rich committed
364
            fetch_time_start = time.time()
365
            try:
366 367
                #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
368
                #inventory = ALPSBridge.fetch_inventory(resinfo=True) #This is a full-refresh,
369
                #determine if summary may be used under normal operation
370
                #updated for >= 1.6 interface
Paul Rich's avatar
Paul Rich committed
371 372 373
                inven_nodes = ALPSBridge.extract_system_node_data(ALPSBridge.system())
                reservations = ALPSBridge.fetch_reservations()
                #reserved_nodes = ALPSBridge.reserved_nodes()
374 375 376 377
            except (ALPSBridge.ALPSError, ComponentLookupError):
                _logger.warning('Error contacting ALPS for state update.  Aborting this update',
                        exc_info=True)
                return
378
            inven_reservations = reservations.get('reservations', [])
Paul Rich's avatar
Paul Rich committed
379
            fetch_time_start = time.time()
380
            #_logger.debug("time in ALPS fetch: %s seconds", (time.time() - fetch_time_start))
381
            start_time = time.time()
Paul Rich's avatar
Paul Rich committed
382
            self._detect_rereservation(inven_reservations)
383 384 385 386 387 388
            # 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
389
            res_jobid_to_delete = []
390 391 392 393
            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
394 395
                # later. This catches leftover reservations from hard-shutdowns
                # while running.
396 397 398 399
                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
400 401
                if alps_res.jobid in self.pending_starts.keys():
                    continue #Get to this only after timeout happens
Paul Rich's avatar
Paul Rich committed
402
                #find alps_id associated reservation
403
                if int(alps_res.alps_res_id) not in current_alps_res_ids:
Paul Rich's avatar
Paul Rich committed
404
                    for node_id in alps_res.node_ids:
405
                        if not self.nodes[str(node_id)].reserved:
Paul Rich's avatar
Paul Rich committed
406
                            #pending hardware status update
407
                            self.nodes[str(node_id)].status = 'idle'
Paul Rich's avatar
Paul Rich committed
408
                    res_jobid_to_delete.append(alps_res.jobid)
Paul Rich's avatar
Paul Rich committed
409
                    _logger.info('job %s: Nodes %s cleanup complete.', alps_res.jobid, compact_num_list(alps_res.node_ids))
Paul Rich's avatar
Paul Rich committed
410
            for jobid in res_jobid_to_delete:
411 412
                _logger.info('%s: ALPS reservation for this job complete.', jobid)
                del self.alps_reservations[str(jobid)]
Paul Rich's avatar
Paul Rich committed
413 414
            #process group should already be on the way down since cqm released the
            #resource reservation
415
            cleanup_nodes = [node for node in self.nodes.values()
416
                             if node.status in ['cleanup-pending', 'cleanup']]
Paul Rich's avatar
Paul Rich committed
417
            #If we have a block marked for cleanup, send a release message.
Paul Rich's avatar
Paul Rich committed
418
            released_res_jobids = []
419
            cleaned_nodes = []
Paul Rich's avatar
Paul Rich committed
420
            for node in cleanup_nodes:
421
                found = False
422
                for alps_res in self.alps_reservations.values():
423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441
                    if str(node.node_id) in alps_res.node_ids:
                        found = True
                        if alps_res.jobid not in released_res_jobids:
                            #send only one release per iteration
                            apids = alps_res.release()
                            if apids is not None:
                                for apid in apids:
                                    self.signal_aprun(apid)
                            released_res_jobids.append(alps_res.jobid)
                if not found:
                    # There is no alps reservation to release, cleanup is
                    # already done.  This happens with very poorly timed
                    # qdel requests. Status will be set properly with the
                    # subsequent hardware status check.
                    _logger.info('Node %s cleanup complete.', node.node_id)
                    node.status = 'idle'
                    cleaned_nodes.append(node)
            for node in cleaned_nodes:
                cleanup_nodes.remove(node)
Paul Rich's avatar
Paul Rich committed
442 443

        #find hardware status
444 445 446
            #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
447
            for inven_node in inven_nodes.values():
448 449
                if self.nodes.has_key(str(inven_node['node_id'])):
                    node = self.nodes[str(inven_node['node_id'])]
450
                    node.role = inven_node['role'].upper()
451
                    node.attributes.update(inven_node['attrs'])
Paul Rich's avatar
Paul Rich committed
452
                    if node.reserved:
453 454
                        #node marked as reserved.
                        if self.alps_reservations.has_key(str(node.reserved_jobid)):
Paul Rich's avatar
Paul Rich committed
455 456
                            node.status = 'busy'
                        else:
457 458 459 460 461 462
                            # 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
463
                    else:
464
                        node.status = inven_node['state'].upper()
465 466
                        if node.role.upper() not in ['BATCH'] and node.status is 'idle':
                            node.status = 'alps-interactive'
467
                else:
468 469 470
                    # 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.
471
                    _logger.warning('Unknown node %s found. Starting reconstruction.', inven_node['node_id'])
472 473 474 475 476 477 478 479 480
                    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
481 482 483
            #should down win over running in terms of display?
            #keep node that are marked for cleanup still in cleanup
            for node in cleanup_nodes:
484
                node.status = 'cleanup-pending'
485
        #_logger.debug("time in UNS lock: %s seconds", (time.time() - start_time))
486 487
        return

Paul Rich's avatar
Paul Rich committed
488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536
    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
537
                        apids = tracked_res.release()
Paul Rich's avatar
Paul Rich committed
538 539 540 541 542 543 544 545 546 547 548 549 550 551 552
                    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
553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578
        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
579
            _logger.info("killing backend ALPS application_id: %s", aprun_id)
Paul Rich's avatar
Paul Rich committed
580 581 582 583 584 585 586 587 588 589 590 591 592 593 594
        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.

        '''
595

Paul Rich's avatar
Paul Rich committed
596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611
        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
612
        return
Paul Rich's avatar
Paul Rich committed
613

614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645
    @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 = []
646
        node_active_queues = set([])
647
        self.current_equivalence_classes = [] #reverse mapping of queues to nodes
648 649 650
        for node in self.nodes.values():
            if node.managed and node.schedulable:
                #only condiser nodes that we are scheduling.
651
                node_active_queues = set([])
652 653
                for queue in node.queues:
                    if queue in active_queue_names:
654 655
                        node_active_queues.add(queue)
                if node_active_queues == set([]):
656 657 658 659 660 661 662 663 664
                    #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']:
665
                        e['data'].add(node.node_id)
666 667 668 669 670 671
                        e['queues'] = e['queues'] | set(node_active_queues)
                        found_a_match = True
                        break
                if found_a_match:
                    break
            if not found_a_match:
672
                equiv.append({'queues': set(node_active_queues),
673
                              'data': set([node.node_id]),
674
                              'reservations': set()})
675 676 677 678 679
        #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:
680
                if e['queues'].intersection(eq_class['queues']):
681 682 683 684 685 686 687 688 689 690
                    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:
691 692 693 694
                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)
695
                            break
696
            #don't send what could be a large block list back in the returun
697 698 699
            for key in eq_class:
                eq_class[key] = list(eq_class[key])
            del eq_class['data']
700
            self.current_equivalence_classes.append(eq_class)
701 702
        return equiv

703 704 705 706 707 708
    @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)
709

710
    def _assemble_queue_data(self, job, idle_only=True, drain_time=None):
711 712
        '''put together data for a queue, or queue-like reservation structure.

713 714 715 716 717 718 719
        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.
720 721 722


        '''
723 724 725 726
        # 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.
727
        unavailable_nodes = []
728
        forbidden, required, requested_locations = self._setup_special_locaitons(job)
729 730 731 732 733 734
        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
735
        if job['queue'] not in self.nodes_by_queue.keys():
736 737 738 739 740
            # 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)
741
        if requested_locations != set([]): # handle attrs location= requests
742
            job_set = set([str(nid) for nid in requested_locations])
743
            if job['queue'] not in self.nodes_by_queue.keys():
744 745 746
                #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.
747
                    node_id_list = list(requested_locations)
748 749 750 751
                else:
                    # We can't run this job.  Insufficent resources in this
                    # reservation to do so.  Don't risk blocking anything.
                    node_id_list = []
752
            else:
753 754
                #normal queues.  Restrict to the non-reserved nodes.
                if job_set <= set([str(node_id) for node_id in
755
                                    self.nodes_by_queue[job['queue']]]):
756
                    node_id_list = list(requested_locations)
757 758 759 760 761 762 763
                    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 = []
764 765
                    if not requested_loc_in_forbidden:
                        raise ValueError("forbidden locations not in queue")
766
        with self._node_lock:
767 768 769 770 771 772 773
            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]
774
            if drain_time is not None:
775
                unavailable_nodes.extend([node_id for node_id in node_id_list
776
                    if (self.nodes[str(node_id)].draining and
777
                        (self.nodes[str(node_id)].drain_until - BACKFILL_EPSILON) < int(drain_time))])
778
        for node_id in set(unavailable_nodes):
779
            node_id_list.remove(node_id)
780
        return sorted(node_id_list, key=lambda nid: int(nid))
781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799

    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):
800 801
                node_id_list.sort(key=lambda nid: int(nid))
                ret_nodes = node_id_list[:int(job['nodes'])]
802 803
        return ret_nodes

804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822
    def _select_first_nodes_prefer_memory_match(self, job, node_id_list):
        '''Given a list of nids, select the first node count nodes fromt the
        list.  Prefer nodes that match the memory modes for a given job, then
        go in nid order.

        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.

        '''
        if job.get('attrs', {}).get('mcdram', None) is None or job.get('attrs', {}).get('numa', None) is None:
            # insufficient information to include a mode match
            return self._select_first_nodes(job, node_id_list)
823
        ret_nids = []
824 825 826
        with self._node_lock:
            considered_nodes = [node for node in self.nodes.values() if node.node_id in node_id_list]
            for node in considered_nodes:
827
                if (node.attributes['hbm_cache_pct'] == MCDRAM_TO_HBMCACHEPCT[job['attrs']['mcdram']] and
828
                        node.attributes['numa_cfg'] == job['attrs']['numa']):
829 830
                    ret_nids.append(int(node.node_id))
            if len(ret_nids) < int(job['nodes']):
831 832
                node_id_list.sort(key=lambda nid: int(nid))
                for nid in node_id_list:
833 834
                    if int(nid) not in ret_nids:
                        ret_nids.append(int(nid))
835 836
                        if len(ret_nids) >= int(job['nodes']):
                            break
837 838
        return ret_nids[:int(job['nodes'])]

839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858
    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
859
            to_alps_list = self._select_first_nodes_prefer_memory_match(job, node_id_list)
860 861 862 863 864 865 866 867
            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
868

869
    @locking
870 871
    @exposed
    def find_job_location(self, arg_list, end_times, pt_blocking_locations=[]):
872 873 874 875 876 877
        '''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.

878 879
        Args::
            arg_list: A list of dictionaries containning information on jobs to
880
                   cosnider.
881
            end_times: list containing a mapping of locations and the times jobs
882 883
                    runninng on those locations are scheduled to end.  End times
                    are in seconds from Epoch UTC.
884
            pt_blocking_locations: Not used for this system.  Used in partitioned
885 886 887 888
                                interconnect schemes. A list of locations that
                                should not be used to prevent passthrough issues
                                with other scheduler reservations.

889
        Returns:
890 891 892 893 894 895 896 897 898 899 900 901
        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.

902 903 904
        pt_blocking_locations may be used later to block out nodes that are
        impacted by warmswap operations.

905
        This function *DOES NOT* hold the component lock.
906

907 908 909
        '''
        now = time.time()
        resource_until_time = now + TEMP_RESERVATION_TIME
910
        with self._node_lock:
911
            # only valid for this scheduler iteration.
912 913 914 915
            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)
916
            for job in arg_list:
917
                label = '%s/%s' % (job['jobid'], job['user'])
918 919 920
                # walltime is in minutes.  We should really fix the storage of
                # that --PMR
                job_endtime = now + (int(job['walltime']) * 60)
921
                try:
922
                    node_id_list = self._assemble_queue_data(job, drain_time=job_endtime)
923
                    available_node_list = self._assemble_queue_data(job, idle_only=False)
924 925 926
                except ValueError:
                    _logger.warning('Job %s: requesting locations that are not in requested queue.',
                            job['jobid'])
927
                    continue
928
                if int(job['nodes']) > len(available_node_list):
929
                    # Insufficient operational nodes for this job at all
930
                    continue
931 932
                elif len(node_id_list) == 0:
                    pass #allow for draining pass to run.
933
                elif int(job['nodes']) <= len(node_id_list):
934 935
                    # 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
936 937
                    compact_locs = self._associate_and_run_immediate(job,
                            resource_until_time, node_id_list)
938 939 940
                    # 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
941
                        best_match[job['jobid']] = [compact_locs]
942
                        _logger.info("%s: Job selected for running on nodes  %s",
Paul Rich's avatar
Paul Rich committed
943
                                label, compact_locs)
944
                        break #for now only select one location
945
                if DRAIN_MODE in ['backfill', 'drain-only']:
946
                    # drain sufficient nodes for this job to run
947
                    drain_node_ids = self._select_nodes_for_draining(job,
Paul Rich's avatar
Paul Rich committed
948
                            end_times)
949 950 951
                    if drain_node_ids != []:
                        _logger.info('%s: nodes %s selected for draining.', label,
                                compact_num_list(drain_node_ids))
952
        return best_match
953

954
    def _ALPS_reserve_resources(self, job, new_time, node_id_list):
955 956 957 958 959 960 961 962 963 964 965 966 967
        '''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.

        '''
968 969
        try:
            res_info = ALPSBridge.reserve(job['user'], job['jobid'],
970
                int(job['nodes']), job['attrs'], node_id_list)
971
        except ALPSBridge.ALPSError as exc:
972
            _logger.warning('unable to reserve resources from ALPS: %s', exc.message)