CraySystem.py 72 KB
Newer Older
1
"""Resource management for Cray ALPS based systems"""
2 3 4 5
import logging
import threading
import thread
import time
6
import sys
7
import xmlrpclib
8
import json
9
import ConfigParser
10 11
import Cobalt.Util
import Cobalt.Components.system.AlpsBridge as ALPSBridge
12
from Cobalt.Components.base import Component, exposed, automatic, query, locking
13 14
from Cobalt.Components.system.base_system import BaseSystem
from Cobalt.Components.system.CrayNode import CrayNode
15
from Cobalt.Components.system.base_pg_manager import ProcessGroupManager
16
from Cobalt.Components.system.ALPSProcessGroup import ALPSProcessGroup
17
from Cobalt.Exceptions import ComponentLookupError
Paul Rich's avatar
Paul Rich committed
18
from Cobalt.Exceptions import JobNotInteractive
19
from Cobalt.Components.system.ALPSProcessGroup import ALPSProcessGroup
20
from Cobalt.Exceptions import JobValidationError
21
from Cobalt.DataTypes.ProcessGroup import ProcessGroup
22
from Cobalt.Util import compact_num_list, expand_num_list
23
from Cobalt.Util import init_cobalt_config, get_config_option
24

25
_logger = logging.getLogger(__name__)
26
init_cobalt_config()
27

28 29
UPDATE_THREAD_TIMEOUT = int(get_config_option('alpssystem', 'update_thread_timeout', 10))
TEMP_RESERVATION_TIME = int(get_config_option('alpssystem', 'temp_reservation_time', 300))
30
SAVE_ME_INTERVAL = float(get_config_option('alpsssytem', 'save_me_interval', 10.0))
31 32
#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
33
APKILL_CMD = get_config_option('alps', 'apkill', '/opt/cray/alps/default/bin/apkill')
34
DRAIN_MODE = get_config_option('system', 'drain_mode', 'first-fit')
35 36 37
#cleanup time in seconds
CLEANUP_DRAIN_WINDOW = get_config_option('system', 'cleanup_drain_window', 300)

38 39
#Epsilon for backfilling.  This system does not do this on a per-node basis.
BACKFILL_EPSILON = int(get_config_option('system', 'backfill_epsilon', 120))
40
ELOGIN_HOSTS = [host for host in get_config_option('system', 'elogin_hosts', '').split(':')]
41 42
if ELOGIN_HOSTS == ['']:
    ELOGIN_HOSTS = []
Paul Rich's avatar
Paul Rich committed
43
DRAIN_MODES = ['first-fit', 'backfill']
44
CLEANING_ID = -1
45 46
DEFAULT_MCDRAM_MODE = get_config_option('alpssystem', 'default_mcdram_mode', 'cache')
DEFAULT_NUMA_MODE = get_config_option('alpssystem', 'default_numa_mode', 'quad')
47
MCDRAM_TO_HBMCACHEPCT = {'flat':'0', 'cache':'100', 'split':'25', 'equal':'50', '0':'0', '25':'25', '50':'50', '100':'100'}
48 49
VALID_MCDRAM_MODES = ['flat', 'cache', 'split', 'equal', '0', '25', '50', '100']
VALID_NUMA_MODES = ['a2a', 'hemi', 'quad', 'snc2', 'snc4']
50 51


52 53 54 55 56 57 58 59 60 61
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

62

63
class CraySystem(BaseSystem):
64 65 66
    '''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.
67

68
    '''
69 70
    name = "system"
    implementation = "alps_system"
71
    logger = _logger
72

73
    def __init__(self, *args, **kwargs):
74 75 76 77
        '''Initialize system.  Read initial states from bridge.
        Get current state

        '''
78
        start_time = time.time()
79
        super(CraySystem, self).__init__(*args, **kwargs)
80
        _logger.info('BASE SYSTEM INITIALIZED')
81 82 83 84 85 86 87 88 89 90
        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.

        '''
91 92 93 94 95
        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)
96 97 98 99 100
        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)
101
        #initilaize bridge.
102 103 104 105
        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
106 107
            # These attempts may fail due to system_script_forker not being up.
            # We don't want to trash the statefile in this case.
108 109 110 111
            try:
                ALPSBridge.init_bridge()
            except ALPSBridge.BridgeError:
                _logger.error('Bridge Initialization failed.  Retrying.')
112 113 114 115
                Cobalt.Util.sleep(10)
            except ComponentLookupError:
                _logger.warning('Error reaching forker.  Retrying.')
                Cobalt.Util.sleep(10)
116 117 118
            else:
                bridge_pending = False
                _logger.info('BRIDGE INITIALIZED')
119
        #process manager setup
120
        if spec is None:
Paul Rich's avatar
Paul Rich committed
121
            self.process_manager = ProcessGroupManager(pgroup_type=ALPSProcessGroup)
122
        else:
Paul Rich's avatar
Paul Rich committed
123
            self.process_manager = ProcessGroupManager(pgroup_type=ALPSProcessGroup).__setstate__(spec['process_manager'])
Paul Rich's avatar
Paul Rich committed
124
            self.logger.debug('pg type %s', self.process_manager.process_groups.item_cls)
125 126
        #self.process_manager.forkers.append('alps_script_forker')
        self.process_manager.update_launchers()
Paul Rich's avatar
Paul Rich committed
127
        self.pending_start_timeout = PENDING_STARTUP_TIMEOUT
128 129
        _logger.info('PROCESS MANAGER INTIALIZED')
        #resource management setup
130
        self.nodes = {} #cray node_id: CrayNode
131 132 133 134
        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']
135
        self._init_nodes_and_reservations()
136 137 138
        if spec is not None:
            node_info = spec.get('node_info', {})
            for nid, node in node_info.items():
139 140 141 142
                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
143 144 145
        #storage for pending job starts.  Allows us to handle slow starts vs
        #user deletes
        self.pending_starts = {} #jobid: time this should be cleared.
146
        self.nodes_by_queue = {} #queue:[node_ids]
147
        #populate initial state
148
        #state update thread and lock
149
        self._node_lock = threading.RLock()
150
        self._gen_node_to_queue()
Paul Rich's avatar
Paul Rich committed
151
        self.node_update_thread = thread.start_new_thread(self._run_update_state, tuple())
152
        _logger.info('UPDATE THREAD STARTED')
153
        self.current_equivalence_classes = []
Paul Rich's avatar
Paul Rich committed
154
        self.killing_jobs = {}
155 156 157 158
        #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
159

160 161 162 163 164 165 166 167
    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
168
        state['node_info'] = self.nodes
169 170 171 172 173 174 175
        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)
176
        _logger.info('ALPS SYSTEM COMPONENT READY TO RUN')
177
        _logger.info('Reinitilaization complete in %s sec.', (time.time() -
178
                start_time))
179

180 181 182 183 184
    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)
185

186 187 188 189
    def _init_nodes_and_reservations(self):
        '''Initialize nodes from ALPS bridge data'''

        retnodes = {}
190 191 192
        pending = True
        while pending:
            try:
193 194 195 196 197
                # 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
198
                _logger.info('INVENTORY FETCHED')
199
                system = ALPSBridge.extract_system_node_data(ALPSBridge.system())
Paul Rich's avatar
Paul Rich committed
200
                _logger.info('SYSTEM DATA FETCHED')
201
                reservations = ALPSBridge.fetch_reservations()
Paul Rich's avatar
Paul Rich committed
202
                _logger.info('ALPS RESERVATION DATA FETCHED')
203
                # reserved_nodes = ALPSBridge.reserved_nodes()
204 205 206
            except Exception:
                #don't crash out here.  That may trash a statefile.
                _logger.error('Possible transient encountered during initialization. Retrying.',
207
                        exc_info=True)
208 209 210 211
                Cobalt.Util.sleep(10)
            else:
                pending = False

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

221
    def _assemble_nodes(self, inventory, system):
222 223 224 225 226 227 228 229 230 231
        '''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():
232 233 234 235 236 237
            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']
238 239 240 241 242
        self.nodes = nodes

    def _assemble_reservations(self, reservations, reserved_nodes):
        # FIXME: we can recover reservations now.  Implement this.
        pass
243

244
    def _gen_node_to_queue(self):
245
        '''(Re)Generate a mapping for fast lookup of node-id's to queues.'''
246
        with self._node_lock:
247
            self.nodes_by_queue = {}
248 249 250
            for node in self.nodes.values():
                for queue in node.queues:
                    if queue in self.nodes_by_queue.keys():
251
                        self.nodes_by_queue[queue].add(node.node_id)
252
                    else:
253
                        self.nodes_by_queue[queue] = set([node.node_id])
254

255
    @exposed
256
    def get_nodes(self, as_dict=False, node_ids=None, params=None, as_json=False):
257 258
        '''fetch the node dictionary.

259 260 261 262 263 264 265
            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.
266 267 268 269 270

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

        '''
271 272 273 274 275 276 277 278 279
        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()))
280
        else:
281 282 283 284 285
            node_info = dict(filter(node_filter, self.nodes.items()))
        if as_json:
            return json.dumps(node_info)
        return node_info

286 287
    def _run_update_state(self):
        '''automated node update functions on the update timer go here.'''
288 289

        def _run_and_wrap(func):
290
            try:
291
                func()
292
            except Exception:
293
                # Prevent this thread from dying.
294
                _logger.critical('Error in _run_update_state', exc_info=True)
295 296 297 298 299 300 301 302

        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)
303

304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338
    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
339
        new_node.managed = True
340 341 342
        self.nodes[str(nid)] = new_node
        self.logger.warning('Node %s added to tracking.', nid)

343 344 345 346 347 348
    @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
349 350 351
        #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.
352
        now = time.time()
Paul Rich's avatar
Paul Rich committed
353 354 355 356 357 358 359 360
        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
361
        self.check_killing_aprun()
362
        with self._node_lock:
Paul Rich's avatar
Paul Rich committed
363
            fetch_time_start = time.time()
364
            try:
365 366
                #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
367
                #inventory = ALPSBridge.fetch_inventory(resinfo=True) #This is a full-refresh,
368
                #determine if summary may be used under normal operation
369
                #updated for >= 1.6 interface
Paul Rich's avatar
Paul Rich committed
370 371 372
                inven_nodes = ALPSBridge.extract_system_node_data(ALPSBridge.system())
                reservations = ALPSBridge.fetch_reservations()
                #reserved_nodes = ALPSBridge.reserved_nodes()
373 374 375 376
            except (ALPSBridge.ALPSError, ComponentLookupError):
                _logger.warning('Error contacting ALPS for state update.  Aborting this update',
                        exc_info=True)
                return
377
            inven_reservations = reservations.get('reservations', [])
Paul Rich's avatar
Paul Rich committed
378
            fetch_time_start = time.time()
379
            #_logger.debug("time in ALPS fetch: %s seconds", (time.time() - fetch_time_start))
380
            start_time = time.time()
Paul Rich's avatar
Paul Rich committed
381
            self._detect_rereservation(inven_reservations)
382 383 384 385 386 387
            # 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
388
            res_jobid_to_delete = []
389 390 391 392
            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
393 394
                # later. This catches leftover reservations from hard-shutdowns
                # while running.
395 396 397 398
                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
399 400
                if alps_res.jobid in self.pending_starts.keys():
                    continue #Get to this only after timeout happens
Paul Rich's avatar
Paul Rich committed
401
                #find alps_id associated reservation
402
                if int(alps_res.alps_res_id) not in current_alps_res_ids:
Paul Rich's avatar
Paul Rich committed
403
                    for node_id in alps_res.node_ids:
404
                        if not self.nodes[str(node_id)].reserved:
Paul Rich's avatar
Paul Rich committed
405
                            #pending hardware status update
406
                            self.nodes[str(node_id)].status = 'idle'
Paul Rich's avatar
Paul Rich committed
407
                    res_jobid_to_delete.append(alps_res.jobid)
Paul Rich's avatar
Paul Rich committed
408
                    _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
409
            for jobid in res_jobid_to_delete:
410
                _logger.info('%s: ALPS reservation for this job complete.', jobid)
Paul Rich's avatar
Paul Rich committed
411 412 413 414 415 416 417 418
                try:
                    del self.alps_reservations[str(jobid)]
                except KeyError:
                    _logger.warning('Job %s: Attempted to remove ALPS reservation for this job multiple times', jobid)
                if self.alps_reservations.get(int(jobid), None) is not None:
                    # in case of type leakage
                    _logger.warning('Job %s: ALPS reservation found with integer key: deleting', jobid)
                    del self.alps_reservations[jobid]
Paul Rich's avatar
Paul Rich committed
419 420
            #process group should already be on the way down since cqm released the
            #resource reservation
421
            cleanup_nodes = [node for node in self.nodes.values()
422
                             if node.status in ['cleanup-pending', 'cleanup']]
Paul Rich's avatar
Paul Rich committed
423
            #If we have a block marked for cleanup, send a release message.
Paul Rich's avatar
Paul Rich committed
424
            released_res_jobids = []
425
            cleaned_nodes = []
Paul Rich's avatar
Paul Rich committed
426
            for node in cleanup_nodes:
427
                found = False
428
                for alps_res in self.alps_reservations.values():
429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447
                    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
448 449

        #find hardware status
450 451 452
            #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
453
            for inven_node in inven_nodes.values():
454 455
                if self.nodes.has_key(str(inven_node['node_id'])):
                    node = self.nodes[str(inven_node['node_id'])]
456
                    node.role = inven_node['role'].upper()
457
                    node.attributes.update(inven_node['attrs'])
Paul Rich's avatar
Paul Rich committed
458
                    if node.reserved:
459 460
                        #node marked as reserved.
                        if self.alps_reservations.has_key(str(node.reserved_jobid)):
Paul Rich's avatar
Paul Rich committed
461 462
                            node.status = 'busy'
                        else:
463 464 465 466 467 468
                            # 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
469
                    else:
470
                        node.status = inven_node['state'].upper()
471 472
                        if node.role.upper() not in ['BATCH'] and node.status is 'idle':
                            node.status = 'alps-interactive'
473
                else:
474 475 476
                    # 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.
477
                    _logger.warning('Unknown node %s found. Starting reconstruction.', inven_node['node_id'])
478 479 480 481 482 483 484 485 486
                    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
487 488 489
            #should down win over running in terms of display?
            #keep node that are marked for cleanup still in cleanup
            for node in cleanup_nodes:
490
                node.status = 'cleanup-pending'
491
        #_logger.debug("time in UNS lock: %s seconds", (time.time() - start_time))
492 493
        return

Paul Rich's avatar
Paul Rich committed
494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542
    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
543
                        apids = tracked_res.release()
Paul Rich's avatar
Paul Rich committed
544 545 546 547 548 549 550 551 552 553 554 555 556 557 558
                    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
559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584
        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
585
            _logger.info("killing backend ALPS application_id: %s", aprun_id)
Paul Rich's avatar
Paul Rich committed
586 587 588 589 590 591 592 593 594 595 596 597 598 599 600
        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.

        '''
601

Paul Rich's avatar
Paul Rich committed
602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617
        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
618
        return
Paul Rich's avatar
Paul Rich committed
619

620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651
    @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 = []
652
        node_active_queues = set([])
653
        self.current_equivalence_classes = [] #reverse mapping of queues to nodes
654 655 656
        for node in self.nodes.values():
            if node.managed and node.schedulable:
                #only condiser nodes that we are scheduling.
657
                node_active_queues = set([])
658 659
                for queue in node.queues:
                    if queue in active_queue_names:
660 661
                        node_active_queues.add(queue)
                if node_active_queues == set([]):
662 663 664 665 666 667 668 669 670
                    #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']:
671
                        e['data'].add(node.node_id)
672 673 674 675 676 677
                        e['queues'] = e['queues'] | set(node_active_queues)
                        found_a_match = True
                        break
                if found_a_match:
                    break
            if not found_a_match:
678
                equiv.append({'queues': set(node_active_queues),
679
                              'data': set([node.node_id]),
680
                              'reservations': set()})
681 682 683 684 685
        #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:
686
                if e['queues'].intersection(eq_class['queues']):
687 688 689 690 691 692 693 694 695 696
                    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:
697 698 699 700
                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)
701
                            break
702
            #don't send what could be a large block list back in the returun
703 704 705
            for key in eq_class:
                eq_class[key] = list(eq_class[key])
            del eq_class['data']
706
            self.current_equivalence_classes.append(eq_class)
707 708
        return equiv

709 710 711 712 713 714
    @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)
715

716
    def _assemble_queue_data(self, job, idle_only=True, drain_time=None):
717 718
        '''put together data for a queue, or queue-like reservation structure.

719 720 721 722 723 724 725
        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.
726 727 728


        '''
729 730 731 732
        # 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.
733
        unavailable_nodes = []
734
        forbidden, required, requested_locations = self._setup_special_locaitons(job)
735 736 737 738 739 740
        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
741
        if job['queue'] not in self.nodes_by_queue.keys():
742 743 744 745 746
            # 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)
747
        if requested_locations != set([]): # handle attrs location= requests
748
            job_set = set([str(nid) for nid in requested_locations])
749
            if job['queue'] not in self.nodes_by_queue.keys():
750 751 752
                #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.
753
                    node_id_list = list(requested_locations)
754 755 756 757
                else:
                    # We can't run this job.  Insufficent resources in this
                    # reservation to do so.  Don't risk blocking anything.
                    node_id_list = []
758
            else:
759 760
                #normal queues.  Restrict to the non-reserved nodes.
                if job_set <= set([str(node_id) for node_id in
761
                                    self.nodes_by_queue[job['queue']]]):
762
                    node_id_list = list(requested_locations)
763 764 765 766 767 768 769
                    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 = []
770 771
                    if not requested_loc_in_forbidden:
                        raise ValueError("forbidden locations not in queue")
772
        with self._node_lock:
773 774 775 776 777 778 779
            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]
780
            if drain_time is not None:
781
                unavailable_nodes.extend([node_id for node_id in node_id_list
782
                    if (self.nodes[str(node_id)].draining and
783
                        (self.nodes[str(node_id)].drain_until - BACKFILL_EPSILON) < int(drain_time))])
784
        for node_id in set(unavailable_nodes):
785
            node_id_list.remove(node_id)
786
        return sorted(node_id_list, key=lambda nid: int(nid))
787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805

    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):
806 807
                node_id_list.sort(key=lambda nid: int(nid))
                ret_nodes = node_id_list[:int(job['nodes'])]
808 809
        return ret_nodes

810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828
    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)
829
        ret_nids = []
830 831 832
        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:
833
                if (node.attributes['hbm_cache_pct'] == MCDRAM_TO_HBMCACHEPCT[job['attrs']['mcdram']] and
834
                        node.attributes['numa_cfg'] == job['attrs']['numa']):
835 836
                    ret_nids.append(int(node.node_id))
            if len(ret_nids) < int(job['nodes']):
837 838
                node_id_list.sort(key=lambda nid: int(nid))
                for nid in node_id_list:
839 840
                    if int(nid) not in ret_nids:
                        ret_nids.append(int(nid))
841 842
                        if len(ret_nids) >= int(job['nodes']):
                            break
843 844
        return ret_nids[:int(job['nodes'])]

845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864
    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
865
            to_alps_list = self._select_first_nodes_prefer_memory_match(job, node_id_list)
866 867 868 869 870 871 872 873
            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
874

875
    @locking
876 877
    @exposed
    def find_job_location(self, arg_list, end_times, pt_blocking_locations=[]):
878 879 880 881 882 883
        '''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.

884 885
        Args::
            arg_list: A list of dictionaries containning information on jobs to
886
                   cosnider.
887
            end_times: list containing a mapping of locations and the times jobs
888 889
                    runninng on those locations are scheduled to end.  End times
                    are in seconds from Epoch UTC.
890
            pt_blocking_locations: Not used for this system.  Used in partitioned
891 892 893 894
                                interconnect schemes. A list of locations that
                                should not be used to prevent passthrough issues
                                with other scheduler reservations.

895
        Returns:
896 897 898 899 900 901 902 903 904 905 906 907
        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.

908 909 910
        pt_blocking_locations may be used later to block out nodes that are
        impacted by warmswap operations.

911
        This function *DOES NOT* hold the component lock.
912

913 914 915
        '''
        now = time.time()
        resource_until_time = now + TEMP_RESERVATION_TIME
916
        with self._node_lock:
917
            # only valid for this scheduler iteration.
918 919 920 921
            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)
922
            for job in arg_list:
923
                label = '%s/%s' % (job['jobid'], job['user'])
924 925 926
                # walltime is in minutes.  We should really fix the storage of
                # that --PMR
                job_endtime = now + (int(job['walltime']) * 60)
927
                try:
928
                    node_id_list = self._assemble_queue_data(job, drain_time=job_endtime)
929
                    available_node_list = self._assemble_queue_data(job, idle_only=False)
930 931 932
                except ValueError:
                    _logger.warning('Job %s: requesting locations that are not in requested queue.',
                            job['jobid'])
933
                    continue
934
                if int(job['nodes']) > len(available_node_list):
935
                    # Insufficient operational nodes for this job at all
936
                    continue
937 938
                elif len(node_id_list) == 0:
                    pass #allow for draining pass to run.
939
                elif int(job['nodes']) <= len(node_id_list):
940 941
                    # 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
942 943
                    compact_locs = self._associate_and_run_immediate(job,
                            resource_until_time, node_id_list)
944 945 946
                    # 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
947
                        best_match[job['jobid']] = [compact_locs]
948
                        _logger.info("%s: Job selected for running on nodes  %s",
Paul Rich's avatar
Paul Rich committed
949
                                label, compact_locs)
950
                        break #for now only select one location
951
                if DRAIN_MODE in ['backfill', 'drain-only']:
952
                    # drain sufficient nodes for this job to run
953
                    drain_node_ids = self._select_nodes_for_draining(job,
Paul Rich's avatar
Paul Rich committed
954
                            end_times)
955 956 957
                    if drain_node_ids != []:
                        _logger.info('%s: nodes %s selected for draining.', label,
                                compact_num_list(drain_node_ids))
958
        return best_match
959

960
    def _ALPS_reserve_resources(self, job, new_time, node_id_list):
961 962 963 964 965 966 967 968 969 970 971 972 973
        '''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.

        '''
974