src/ceph/src/pybind/mgr/balancer/module.py

   1
   2 """
   3 Balance PG distribution across OSDs.
   4 """
   5
   6 import copy
   7 import errno
   8 import json
   9 import math
  10 import random
  11 import time
  12 from mgr_module import MgrModule, CommandResult
  13 from threading import Event
  14
  15 # available modes: 'none', 'crush', 'crush-compat', 'upmap', 'osd_weight'
  16 default_mode = 'none'
  17 default_sleep_interval = 60   # seconds
  18 default_max_misplaced = .05    # max ratio of pgs replaced at a time
  19
  20 TIME_FORMAT = '%Y-%m-%d_%H:%M:%S'
  21
  22
  23 class MappingState:
  24     def __init__(self, osdmap, pg_dump, desc=''):
  25         self.desc = desc
  26         self.osdmap = osdmap
  27         self.osdmap_dump = self.osdmap.dump()
  28         self.crush = osdmap.get_crush()
  29         self.crush_dump = self.crush.dump()
  30         self.pg_dump = pg_dump
  31         self.pg_stat = {
  32             i['pgid']: i['stat_sum'] for i in pg_dump.get('pg_stats', [])
  33         }
  34         self.poolids = [p['pool'] for p in self.osdmap_dump.get('pools', [])]
  35         self.pg_up = {}
  36         self.pg_up_by_poolid = {}
  37         for poolid in self.poolids:
  38             self.pg_up_by_poolid[poolid] = osdmap.map_pool_pgs_up(poolid)
  39             for a,b in self.pg_up_by_poolid[poolid].iteritems():
  40                 self.pg_up[a] = b
  41
  42     def calc_misplaced_from(self, other_ms):
  43         num = len(other_ms.pg_up)
  44         misplaced = 0
  45         for pgid, before in other_ms.pg_up.iteritems():
  46             if before != self.pg_up.get(pgid, []):
  47                 misplaced += 1
  48         if num > 0:
  49             return float(misplaced) / float(num)
  50         return 0.0
  51
  52 class Plan:
  53     def __init__(self, name, ms):
  54         self.mode = 'unknown'
  55         self.name = name
  56         self.initial = ms
  57
  58         self.osd_weights = {}
  59         self.compat_ws = {}
  60         self.inc = ms.osdmap.new_incremental()
  61
  62     def final_state(self):
  63         self.inc.set_osd_reweights(self.osd_weights)
  64         self.inc.set_crush_compat_weight_set_weights(self.compat_ws)
  65         return MappingState(self.initial.osdmap.apply_incremental(self.inc),
  66                             self.initial.pg_dump,
  67                             'plan %s final' % self.name)
  68
  69     def dump(self):
  70         return json.dumps(self.inc.dump(), indent=4)
  71
  72     def show(self):
  73         ls = []
  74         ls.append('# starting osdmap epoch %d' % self.initial.osdmap.get_epoch())
  75         ls.append('# starting crush version %d' %
  76                   self.initial.osdmap.get_crush_version())
  77         ls.append('# mode %s' % self.mode)
  78         if len(self.compat_ws) and \
  79            '-1' not in self.initial.crush_dump.get('choose_args', {}):
  80             ls.append('ceph osd crush weight-set create-compat')
  81         for osd, weight in self.compat_ws.iteritems():
  82             ls.append('ceph osd crush weight-set reweight-compat %s %f' %
  83                       (osd, weight))
  84         for osd, weight in self.osd_weights.iteritems():
  85             ls.append('ceph osd reweight osd.%d %f' % (osd, weight))
  86         incdump = self.inc.dump()
  87         for pgid in incdump.get('old_pg_upmap_items', []):
  88             ls.append('ceph osd rm-pg-upmap-items %s' % pgid)
  89         for item in incdump.get('new_pg_upmap_items', []):
  90             osdlist = []
  91             for m in item['mappings']:
  92                 osdlist += [m['from'], m['to']]
  93             ls.append('ceph osd pg-upmap-items %s %s' %
  94                       (item['pgid'], ' '.join([str(a) for a in osdlist])))
  95         return '\n'.join(ls)
  96
  97
  98 class Eval:
  99     root_ids = {}        # root name -> id
 100     pool_name = {}       # pool id -> pool name
 101     pool_id = {}         # pool name -> id
 102     pool_roots = {}      # pool name -> root name
 103     root_pools = {}      # root name -> pools
 104     target_by_root = {}  # root name -> target weight map
 105     count_by_pool = {}
 106     count_by_root = {}
 107     actual_by_pool = {}  # pool -> by_* -> actual weight map
 108     actual_by_root = {}  # pool -> by_* -> actual weight map
 109     total_by_pool = {}   # pool -> by_* -> total
 110     total_by_root = {}   # root -> by_* -> total
 111     stats_by_pool = {}   # pool -> by_* -> stddev or avg -> value
 112     stats_by_root = {}   # root -> by_* -> stddev or avg -> value
 113
 114     score_by_pool = {}
 115     score_by_root = {}
 116
 117     score = 0.0
 118
 119     def __init__(self, ms):
 120         self.ms = ms
 121
 122     def show(self, verbose=False):
 123         if verbose:
 124             r = self.ms.desc + '\n'
 125             r += 'target_by_root %s\n' % self.target_by_root
 126             r += 'actual_by_pool %s\n' % self.actual_by_pool
 127             r += 'actual_by_root %s\n' % self.actual_by_root
 128             r += 'count_by_pool %s\n' % self.count_by_pool
 129             r += 'count_by_root %s\n' % self.count_by_root
 130             r += 'total_by_pool %s\n' % self.total_by_pool
 131             r += 'total_by_root %s\n' % self.total_by_root
 132             r += 'stats_by_root %s\n' % self.stats_by_root
 133             r += 'score_by_pool %s\n' % self.score_by_pool
 134             r += 'score_by_root %s\n' % self.score_by_root
 135         else:
 136             r = self.ms.desc + ' '
 137         r += 'score %f (lower is better)\n' % self.score
 138         return r
 139
 140     def calc_stats(self, count, target, total):
 141         num = max(len(target), 1)
 142         r = {}
 143         for t in ('pgs', 'objects', 'bytes'):
 144             avg = float(total[t]) / float(num)
 145             dev = 0.0
 146
 147             # score is a measure of how uneven the data distribution is.
 148             # score lies between [0, 1), 0 means perfect distribution.
 149             score = 0.0
 150             sum_weight = 0.0
 151
 152             for k, v in count[t].iteritems():
 153                 # adjust/normalize by weight
 154                 if target[k]:
 155                     adjusted = float(v) / target[k] / float(num)
 156                 else:
 157                     adjusted = 0.0
 158
 159                 # Overweighted devices and their weights are factors to calculate reweight_urgency.
 160                 # One 10% underfilled device with 5 2% overfilled devices, is arguably a better
 161                 # situation than one 10% overfilled with 5 2% underfilled devices
 162                 if adjusted > avg:
 163                     '''
 164                     F(x) = 2*phi(x) - 1, where phi(x) = cdf of standard normal distribution
 165                     x = (adjusted - avg)/avg.
 166                     Since, we're considering only over-weighted devices, x >= 0, and so phi(x) lies in [0.5, 1).
 167                     To bring range of F(x) in range [0, 1), we need to make the above modification.
 168
 169                     In general, we need to use a function F(x), where x = (adjusted - avg)/avg
 170                     1. which is bounded between 0 and 1, so that ultimately reweight_urgency will also be bounded.
 171                     2. A larger value of x, should imply more urgency to reweight.
 172                     3. Also, the difference between F(x) when x is large, should be minimal.
 173                     4. The value of F(x) should get close to 1 (highest urgency to reweight) with steeply.
 174
 175                     Could have used F(x) = (1 - e^(-x)). But that had slower convergence to 1, compared to the one currently in use.
 176
 177                     cdf of standard normal distribution: https://stackoverflow.com/a/29273201
 178                     '''
 179                     score += target[k] * (math.erf(((adjusted - avg)/avg) / math.sqrt(2.0)))
 180                     sum_weight += target[k]
 181                 dev += (avg - adjusted) * (avg - adjusted)
 182             stddev = math.sqrt(dev / float(max(num - 1, 1)))
 183             score = score / max(sum_weight, 1)
 184             r[t] = {
 185                 'avg': avg,
 186                 'stddev': stddev,
 187                 'sum_weight': sum_weight,
 188                 'score': score,
 189             }
 190         return r
 191
 192 class Module(MgrModule):
 193     COMMANDS = [
 194         {
 195             "cmd": "balancer status",
 196             "desc": "Show balancer status",
 197             "perm": "r",
 198         },
 199         {
 200             "cmd": "balancer mode name=mode,type=CephChoices,strings=none|crush-compat|upmap",
 201             "desc": "Set balancer mode",
 202             "perm": "rw",
 203         },
 204         {
 205             "cmd": "balancer on",
 206             "desc": "Enable automatic balancing",
 207             "perm": "rw",
 208         },
 209         {
 210             "cmd": "balancer off",
 211             "desc": "Disable automatic balancing",
 212             "perm": "rw",
 213         },
 214         {
 215             "cmd": "balancer eval name=plan,type=CephString,req=false",
 216             "desc": "Evaluate data distribution for the current cluster or specific plan",
 217             "perm": "r",
 218         },
 219         {
 220             "cmd": "balancer eval-verbose name=plan,type=CephString,req=false",
 221             "desc": "Evaluate data distribution for the current cluster or specific plan (verbosely)",
 222             "perm": "r",
 223         },
 224         {
 225             "cmd": "balancer optimize name=plan,type=CephString",
 226             "desc": "Run optimizer to create a new plan",
 227             "perm": "rw",
 228         },
 229         {
 230             "cmd": "balancer show name=plan,type=CephString",
 231             "desc": "Show details of an optimization plan",
 232             "perm": "r",
 233         },
 234         {
 235             "cmd": "balancer rm name=plan,type=CephString",
 236             "desc": "Discard an optimization plan",
 237             "perm": "rw",
 238         },
 239         {
 240             "cmd": "balancer reset",
 241             "desc": "Discard all optimization plans",
 242             "perm": "rw",
 243         },
 244         {
 245             "cmd": "balancer dump name=plan,type=CephString",
 246             "desc": "Show an optimization plan",
 247             "perm": "r",
 248         },
 249         {
 250             "cmd": "balancer execute name=plan,type=CephString",
 251             "desc": "Execute an optimization plan",
 252             "perm": "r",
 253         },
 254     ]
 255     active = False
 256     run = True
 257     plans = {}
 258     mode = ''
 259
 260     def __init__(self, *args, **kwargs):
 261         super(Module, self).__init__(*args, **kwargs)
 262         self.event = Event()
 263
 264     def handle_command(self, command):
 265         self.log.warn("Handling command: '%s'" % str(command))
 266         if command['prefix'] == 'balancer status':
 267             s = {
 268                 'plans': self.plans.keys(),
 269                 'active': self.active,
 270                 'mode': self.get_config('mode', default_mode),
 271             }
 272             return (0, json.dumps(s, indent=4), '')
 273         elif command['prefix'] == 'balancer mode':
 274             self.set_config('mode', command['mode'])
 275             return (0, '', '')
 276         elif command['prefix'] == 'balancer on':
 277             if not self.active:
 278                 self.set_config('active', '1')
 279                 self.active = True
 280             self.event.set()
 281             return (0, '', '')
 282         elif command['prefix'] == 'balancer off':
 283             if self.active:
 284                 self.set_config('active', '')
 285                 self.active = False
 286             self.event.set()
 287             return (0, '', '')
 288         elif command['prefix'] == 'balancer eval' or command['prefix'] == 'balancer eval-verbose':
 289             verbose = command['prefix'] == 'balancer eval-verbose'
 290             if 'plan' in command:
 291                 plan = self.plans.get(command['plan'])
 292                 if not plan:
 293                     return (-errno.ENOENT, '', 'plan %s not found' %
 294                             command['plan'])
 295                 ms = plan.final_state()
 296             else:
 297                 ms = MappingState(self.get_osdmap(),
 298                                   self.get("pg_dump"),
 299                                   'current cluster')
 300             return (0, self.evaluate(ms, verbose=verbose), '')
 301         elif command['prefix'] == 'balancer optimize':
 302             plan = self.plan_create(command['plan'])
 303             self.optimize(plan)
 304             return (0, '', '')
 305         elif command['prefix'] == 'balancer rm':
 306             self.plan_rm(command['name'])
 307             return (0, '', '')
 308         elif command['prefix'] == 'balancer reset':
 309             self.plans = {}
 310             return (0, '', '')
 311         elif command['prefix'] == 'balancer dump':
 312             plan = self.plans.get(command['plan'])
 313             if not plan:
 314                 return (-errno.ENOENT, '', 'plan %s not found' % command['plan'])
 315             return (0, plan.dump(), '')
 316         elif command['prefix'] == 'balancer show':
 317             plan = self.plans.get(command['plan'])
 318             if not plan:
 319                 return (-errno.ENOENT, '', 'plan %s not found' % command['plan'])
 320             return (0, plan.show(), '')
 321         elif command['prefix'] == 'balancer execute':
 322             plan = self.plans.get(command['plan'])
 323             if not plan:
 324                 return (-errno.ENOENT, '', 'plan %s not found' % command['plan'])
 325             self.execute(plan)
 326             self.plan_rm(plan)
 327             return (0, '', '')
 328         else:
 329             return (-errno.EINVAL, '',
 330                     "Command not found '{0}'".format(command['prefix']))
 331
 332     def shutdown(self):
 333         self.log.info('Stopping')
 334         self.run = False
 335         self.event.set()
 336
 337     def time_in_interval(self, tod, begin, end):
 338         if begin <= end:
 339             return tod >= begin and tod < end
 340         else:
 341             return tod >= begin or tod < end
 342
 343     def serve(self):
 344         self.log.info('Starting')
 345         while self.run:
 346             self.active = self.get_config('active', '') is not ''
 347             begin_time = self.get_config('begin_time') or '0000'
 348             end_time = self.get_config('end_time') or '2400'
 349             timeofday = time.strftime('%H%M', time.localtime())
 350             self.log.debug('Waking up [%s, scheduled for %s-%s, now %s]',
 351                            "active" if self.active else "inactive",
 352                            begin_time, end_time, timeofday)
 353             sleep_interval = float(self.get_config('sleep_interval',
 354                                                    default_sleep_interval))
 355             if self.active and self.time_in_interval(timeofday, begin_time, end_time):
 356                 self.log.debug('Running')
 357                 name = 'auto_%s' % time.strftime(TIME_FORMAT, time.gmtime())
 358                 plan = self.plan_create(name)
 359                 if self.optimize(plan):
 360                     self.execute(plan)
 361                 self.plan_rm(name)
 362             self.log.debug('Sleeping for %d', sleep_interval)
 363             self.event.wait(sleep_interval)
 364             self.event.clear()
 365
 366     def plan_create(self, name):
 367         plan = Plan(name, MappingState(self.get_osdmap(),
 368                                        self.get("pg_dump"),
 369                                        'plan %s initial' % name))
 370         self.plans[name] = plan
 371         return plan
 372
 373     def plan_rm(self, name):
 374         if name in self.plans:
 375             del self.plans[name]
 376
 377     def calc_eval(self, ms):
 378         pe = Eval(ms)
 379         pool_rule = {}
 380         pool_info = {}
 381         for p in ms.osdmap_dump.get('pools',[]):
 382             pe.pool_name[p['pool']] = p['pool_name']
 383             pe.pool_id[p['pool_name']] = p['pool']
 384             pool_rule[p['pool_name']] = p['crush_rule']
 385             pe.pool_roots[p['pool_name']] = []
 386             pool_info[p['pool_name']] = p
 387         pools = pe.pool_id.keys()
 388         if len(pools) == 0:
 389             return pe
 390         self.log.debug('pool_name %s' % pe.pool_name)
 391         self.log.debug('pool_id %s' % pe.pool_id)
 392         self.log.debug('pools %s' % pools)
 393         self.log.debug('pool_rule %s' % pool_rule)
 394
 395         osd_weight = { a['osd']: a['weight']
 396                        for a in ms.osdmap_dump.get('osds',[]) }
 397
 398         # get expected distributions by root
 399         actual_by_root = {}
 400         rootids = ms.crush.find_takes()
 401         roots = []
 402         for rootid in rootids:
 403             root = ms.crush.get_item_name(rootid)
 404             pe.root_ids[root] = rootid
 405             roots.append(root)
 406             ls = ms.osdmap.get_pools_by_take(rootid)
 407             pe.root_pools[root] = []
 408             for poolid in ls:
 409                 pe.pool_roots[pe.pool_name[poolid]].append(root)
 410                 pe.root_pools[root].append(pe.pool_name[poolid])
 411             weight_map = ms.crush.get_take_weight_osd_map(rootid)
 412             adjusted_map = {
 413                 osd: cw * osd_weight.get(osd, 1.0)
 414                 for osd,cw in weight_map.iteritems()
 415             }
 416             sum_w = sum(adjusted_map.values()) or 1.0
 417             pe.target_by_root[root] = { osd: w / sum_w
 418                                         for osd,w in adjusted_map.iteritems() }
 419             actual_by_root[root] = {
 420                 'pgs': {},
 421                 'objects': {},
 422                 'bytes': {},
 423             }
 424             for osd in pe.target_by_root[root].iterkeys():
 425                 actual_by_root[root]['pgs'][osd] = 0
 426                 actual_by_root[root]['objects'][osd] = 0
 427                 actual_by_root[root]['bytes'][osd] = 0
 428             pe.total_by_root[root] = {
 429                 'pgs': 0,
 430                 'objects': 0,
 431                 'bytes': 0,
 432             }
 433         self.log.debug('pool_roots %s' % pe.pool_roots)
 434         self.log.debug('root_pools %s' % pe.root_pools)
 435         self.log.debug('target_by_root %s' % pe.target_by_root)
 436
 437         # pool and root actual
 438         for pool, pi in pool_info.iteritems():
 439             poolid = pi['pool']
 440             pm = ms.pg_up_by_poolid[poolid]
 441             pgs = 0
 442             objects = 0
 443             bytes = 0
 444             pgs_by_osd = {}
 445             objects_by_osd = {}
 446             bytes_by_osd = {}
 447             for root in pe.pool_roots[pool]:
 448                 for osd in pe.target_by_root[root].iterkeys():
 449                     pgs_by_osd[osd] = 0
 450                     objects_by_osd[osd] = 0
 451                     bytes_by_osd[osd] = 0
 452             for pgid, up in pm.iteritems():
 453                 for osd in [int(osd) for osd in up]:
 454                     pgs_by_osd[osd] += 1
 455                     objects_by_osd[osd] += ms.pg_stat[pgid]['num_objects']
 456                     bytes_by_osd[osd] += ms.pg_stat[pgid]['num_bytes']
 457                     # pick a root to associate this pg instance with.
 458                     # note that this is imprecise if the roots have
 459                     # overlapping children.
 460                     # FIXME: divide bytes by k for EC pools.
 461                     for root in pe.pool_roots[pool]:
 462                         if osd in pe.target_by_root[root]:
 463                             actual_by_root[root]['pgs'][osd] += 1
 464                             actual_by_root[root]['objects'][osd] += ms.pg_stat[pgid]['num_objects']
 465                             actual_by_root[root]['bytes'][osd] += ms.pg_stat[pgid]['num_bytes']
 466                             pgs += 1
 467                             objects += ms.pg_stat[pgid]['num_objects']
 468                             bytes += ms.pg_stat[pgid]['num_bytes']
 469                             pe.total_by_root[root]['pgs'] += 1
 470                             pe.total_by_root[root]['objects'] += ms.pg_stat[pgid]['num_objects']
 471                             pe.total_by_root[root]['bytes'] += ms.pg_stat[pgid]['num_bytes']
 472                             break
 473             pe.count_by_pool[pool] = {
 474                 'pgs': {
 475                     k: v
 476                     for k, v in pgs_by_osd.iteritems()
 477                 },
 478                 'objects': {
 479                     k: v
 480                     for k, v in objects_by_osd.iteritems()
 481                 },
 482                 'bytes': {
 483                     k: v
 484                     for k, v in bytes_by_osd.iteritems()
 485                 },
 486             }
 487             pe.actual_by_pool[pool] = {
 488                 'pgs': {
 489                     k: float(v) / float(max(pgs, 1))
 490                     for k, v in pgs_by_osd.iteritems()
 491                 },
 492                 'objects': {
 493                     k: float(v) / float(max(objects, 1))
 494                     for k, v in objects_by_osd.iteritems()
 495                 },
 496                 'bytes': {
 497                     k: float(v) / float(max(bytes, 1))
 498                     for k, v in bytes_by_osd.iteritems()
 499                 },
 500             }
 501             pe.total_by_pool[pool] = {
 502                 'pgs': pgs,
 503                 'objects': objects,
 504                 'bytes': bytes,
 505             }
 506         for root, m in pe.total_by_root.iteritems():
 507             pe.count_by_root[root] = {
 508                 'pgs': {
 509                     k: float(v)
 510                     for k, v in actual_by_root[root]['pgs'].iteritems()
 511                 },
 512                 'objects': {
 513                     k: float(v)
 514                     for k, v in actual_by_root[root]['objects'].iteritems()
 515                 },
 516                 'bytes': {
 517                     k: float(v)
 518                     for k, v in actual_by_root[root]['bytes'].iteritems()
 519                 },
 520             }
 521             pe.actual_by_root[root] = {
 522                 'pgs': {
 523                     k: float(v) / float(max(pe.total_by_root[root]['pgs'], 1))
 524                     for k, v in actual_by_root[root]['pgs'].iteritems()
 525                 },
 526                 'objects': {
 527                     k: float(v) / float(max(pe.total_by_root[root]['objects'], 1))
 528                     for k, v in actual_by_root[root]['objects'].iteritems()
 529                 },
 530                 'bytes': {
 531                     k: float(v) / float(max(pe.total_by_root[root]['bytes'], 1))
 532                     for k, v in actual_by_root[root]['bytes'].iteritems()
 533                 },
 534             }
 535         self.log.debug('actual_by_pool %s' % pe.actual_by_pool)
 536         self.log.debug('actual_by_root %s' % pe.actual_by_root)
 537
 538         # average and stddev and score
 539         pe.stats_by_root = {
 540             a: pe.calc_stats(
 541                 b,
 542                 pe.target_by_root[a],
 543                 pe.total_by_root[a]
 544             ) for a, b in pe.count_by_root.iteritems()
 545         }
 546
 547         # the scores are already normalized
 548         pe.score_by_root = {
 549             r: {
 550                 'pgs': pe.stats_by_root[r]['pgs']['score'],
 551                 'objects': pe.stats_by_root[r]['objects']['score'],
 552                 'bytes': pe.stats_by_root[r]['bytes']['score'],
 553             } for r in pe.total_by_root.keys()
 554         }
 555
 556         # total score is just average of normalized stddevs
 557         pe.score = 0.0
 558         for r, vs in pe.score_by_root.iteritems():
 559             for k, v in vs.iteritems():
 560                 pe.score += v
 561         pe.score /= 3 * len(roots)
 562         return pe
 563
 564     def evaluate(self, ms, verbose=False):
 565         pe = self.calc_eval(ms)
 566         return pe.show(verbose=verbose)
 567
 568     def optimize(self, plan):
 569         self.log.info('Optimize plan %s' % plan.name)
 570         plan.mode = self.get_config('mode', default_mode)
 571         max_misplaced = float(self.get_config('max_misplaced',
 572                                               default_max_misplaced))
 573         self.log.info('Mode %s, max misplaced %f' %
 574                       (plan.mode, max_misplaced))
 575
 576         info = self.get('pg_status')
 577         unknown = info.get('unknown_pgs_ratio', 0.0)
 578         degraded = info.get('degraded_ratio', 0.0)
 579         inactive = info.get('inactive_pgs_ratio', 0.0)
 580         misplaced = info.get('misplaced_ratio', 0.0)
 581         self.log.debug('unknown %f degraded %f inactive %f misplaced %g',
 582                        unknown, degraded, inactive, misplaced)
 583         if unknown > 0.0:
 584             self.log.info('Some PGs (%f) are unknown; waiting', unknown)
 585         elif degraded > 0.0:
 586             self.log.info('Some objects (%f) are degraded; waiting', degraded)
 587         elif inactive > 0.0:
 588             self.log.info('Some PGs (%f) are inactive; waiting', inactive)
 589         elif misplaced >= max_misplaced:
 590             self.log.info('Too many objects (%f > %f) are misplaced; waiting',
 591                           misplaced, max_misplaced)
 592         else:
 593             if plan.mode == 'upmap':
 594                 return self.do_upmap(plan)
 595             elif plan.mode == 'crush-compat':
 596                 return self.do_crush_compat(plan)
 597             elif plan.mode == 'none':
 598                 self.log.info('Idle')
 599             else:
 600                 self.log.info('Unrecognized mode %s' % plan.mode)
 601         return False
 602
 603         ##
 604
 605     def do_upmap(self, plan):
 606         self.log.info('do_upmap')
 607         max_iterations = self.get_config('upmap_max_iterations', 10)
 608         max_deviation = self.get_config('upmap_max_deviation', .01)
 609
 610         ms = plan.initial
 611         pools = [str(i['pool_name']) for i in ms.osdmap_dump.get('pools',[])]
 612         if len(pools) == 0:
 613             self.log.info('no pools, nothing to do')
 614             return False
 615         # shuffle pool list so they all get equal (in)attention
 616         random.shuffle(pools)
 617         self.log.info('pools %s' % pools)
 618
 619         inc = plan.inc
 620         total_did = 0
 621         left = max_iterations
 622         for pool in pools:
 623             did = ms.osdmap.calc_pg_upmaps(inc, max_deviation, left, [pool])
 624             total_did += did
 625             left -= did
 626             if left <= 0:
 627                 break
 628         self.log.info('prepared %d/%d changes' % (total_did, max_iterations))
 629         return True
 630
 631     def do_crush_compat(self, plan):
 632         self.log.info('do_crush_compat')
 633         max_iterations = self.get_config('crush_compat_max_iterations', 25)
 634         if max_iterations < 1:
 635             return False
 636         step = self.get_config('crush_compat_step', .5)
 637         if step <= 0 or step >= 1.0:
 638             return False
 639         max_misplaced = float(self.get_config('max_misplaced',
 640                                               default_max_misplaced))
 641         min_pg_per_osd = 2
 642
 643         ms = plan.initial
 644         osdmap = ms.osdmap
 645         crush = osdmap.get_crush()
 646         pe = self.calc_eval(ms)
 647         if pe.score == 0:
 648             self.log.info('Distribution is already perfect')
 649             return False
 650
 651         # get current osd reweights
 652         orig_osd_weight = { a['osd']: a['weight']
 653                             for a in ms.osdmap_dump.get('osds',[]) }
 654         reweighted_osds = [ a for a,b in orig_osd_weight.iteritems()
 655                             if b < 1.0 and b > 0.0 ]
 656
 657         # get current compat weight-set weights
 658         orig_ws = self.get_compat_weight_set_weights()
 659         orig_ws = { a: b for a, b in orig_ws.iteritems() if a >= 0 }
 660
 661         # Make sure roots don't overlap their devices.  If so, we
 662         # can't proceed.
 663         roots = pe.target_by_root.keys()
 664         self.log.debug('roots %s', roots)
 665         visited = {}
 666         overlap = {}
 667         root_ids = {}
 668         for root, wm in pe.target_by_root.iteritems():
 669             for osd in wm.iterkeys():
 670                 if osd in visited:
 671                     overlap[osd] = 1
 672                 visited[osd] = 1
 673         if len(overlap) > 0:
 674             self.log.err('error: some osds belong to multiple subtrees: %s' %
 675                          overlap)
 676             return False
 677
 678         key = 'pgs'  # pgs objects or bytes
 679
 680         # go
 681         best_ws = copy.deepcopy(orig_ws)
 682         best_ow = copy.deepcopy(orig_osd_weight)
 683         best_pe = pe
 684         left = max_iterations
 685         bad_steps = 0
 686         next_ws = copy.deepcopy(best_ws)
 687         next_ow = copy.deepcopy(best_ow)
 688         while left > 0:
 689             # adjust
 690             self.log.debug('best_ws %s' % best_ws)
 691             random.shuffle(roots)
 692             for root in roots:
 693                 pools = best_pe.root_pools[root]
 694                 pgs = len(best_pe.target_by_root[root])
 695                 min_pgs = pgs * min_pg_per_osd
 696                 if best_pe.total_by_root[root] < min_pgs:
 697                     self.log.info('Skipping root %s (pools %s), total pgs %d '
 698                                   '< minimum %d (%d per osd)',
 699                                   root, pools, pgs, min_pgs, min_pg_per_osd)
 700                     continue
 701                 self.log.info('Balancing root %s (pools %s) by %s' %
 702                               (root, pools, key))
 703                 target = best_pe.target_by_root[root]
 704                 actual = best_pe.actual_by_root[root][key]
 705                 queue = sorted(actual.keys(),
 706                                key=lambda osd: -abs(target[osd] - actual[osd]))
 707                 for osd in queue:
 708                     if orig_osd_weight[osd] == 0:
 709                         self.log.debug('skipping out osd.%d', osd)
 710                     else:
 711                         deviation = target[osd] - actual[osd]
 712                         if deviation == 0:
 713                             break
 714                         self.log.debug('osd.%d deviation %f', osd, deviation)
 715                         weight = best_ws[osd]
 716                         ow = orig_osd_weight[osd]
 717                         if actual[osd] > 0:
 718                             calc_weight = target[osd] / actual[osd] * weight * ow
 719                         else:
 720                             # not enough to go on here... keep orig weight
 721                             calc_weight = weight / orig_osd_weight[osd]
 722                         new_weight = weight * (1.0 - step) + calc_weight * step
 723                         self.log.debug('Reweight osd.%d %f -> %f', osd, weight,
 724                                        new_weight)
 725                         next_ws[osd] = new_weight
 726                         if ow < 1.0:
 727                             new_ow = min(1.0, max(step + (1.0 - step) * ow,
 728                                                   ow + .005))
 729                             self.log.debug('Reweight osd.%d reweight %f -> %f',
 730                                            osd, ow, new_ow)
 731                             next_ow[osd] = new_ow
 732
 733                 # normalize weights under this root
 734                 root_weight = crush.get_item_weight(pe.root_ids[root])
 735                 root_sum = sum(b for a,b in next_ws.iteritems()
 736                                if a in target.keys())
 737                 if root_sum > 0 and root_weight > 0:
 738                     factor = root_sum / root_weight
 739                     self.log.debug('normalizing root %s %d, weight %f, '
 740                                    'ws sum %f, factor %f',
 741                                    root, pe.root_ids[root], root_weight,
 742                                    root_sum, factor)
 743                     for osd in actual.keys():
 744                         next_ws[osd] = next_ws[osd] / factor
 745
 746             # recalc
 747             plan.compat_ws = copy.deepcopy(next_ws)
 748             next_ms = plan.final_state()
 749             next_pe = self.calc_eval(next_ms)
 750             next_misplaced = next_ms.calc_misplaced_from(ms)
 751             self.log.debug('Step result score %f -> %f, misplacing %f',
 752                            best_pe.score, next_pe.score, next_misplaced)
 753
 754             if next_misplaced > max_misplaced:
 755                 if best_pe.score < pe.score:
 756                     self.log.debug('Step misplaced %f > max %f, stopping',
 757                                    next_misplaced, max_misplaced)
 758                     break
 759                 step /= 2.0
 760                 next_ws = copy.deepcopy(best_ws)
 761                 next_ow = copy.deepcopy(best_ow)
 762                 self.log.debug('Step misplaced %f > max %f, reducing step to %f',
 763                                next_misplaced, max_misplaced, step)
 764             else:
 765                 if next_pe.score > best_pe.score * 1.0001:
 766                     if bad_steps < 5 and random.randint(0, 100) < 70:
 767                         self.log.debug('Score got worse, taking another step')
 768                     else:
 769                         step /= 2.0
 770                         next_ws = copy.deepcopy(best_ws)
 771                         next_ow = copy.deepcopy(best_ow)
 772                         self.log.debug('Score got worse, trying smaller step %f',
 773                                        step)
 774                 else:
 775                     bad_steps = 0
 776                     best_pe = next_pe
 777                     best_ws = next_ws
 778                     best_ow = next_ow
 779                     if best_pe.score == 0:
 780                         break
 781             left -= 1
 782
 783         # allow a small regression if we are phasing out osd weights
 784         fudge = 0
 785         if next_ow != orig_osd_weight:
 786             fudge = .001
 787
 788         if best_pe.score < pe.score + fudge:
 789             self.log.info('Success, score %f -> %f', pe.score, best_pe.score)
 790             plan.compat_ws = best_ws
 791             for osd, w in best_ow.iteritems():
 792                 if w != orig_osd_weight[osd]:
 793                     self.log.debug('osd.%d reweight %f', osd, w)
 794                     plan.osd_weights[osd] = w
 795             return True
 796         else:
 797             self.log.info('Failed to find further optimization, score %f',
 798                           pe.score)
 799             return False
 800
 801     def get_compat_weight_set_weights(self):
 802         # enable compat weight-set
 803         self.log.debug('ceph osd crush weight-set create-compat')
 804         result = CommandResult('')
 805         self.send_command(result, 'mon', '', json.dumps({
 806             'prefix': 'osd crush weight-set create-compat',
 807             'format': 'json',
 808         }), '')
 809         r, outb, outs = result.wait()
 810         if r != 0:
 811             self.log.error('Error creating compat weight-set')
 812             return
 813
 814         result = CommandResult('')
 815         self.send_command(result, 'mon', '', json.dumps({
 816             'prefix': 'osd crush dump',
 817             'format': 'json',
 818         }), '')
 819         r, outb, outs = result.wait()
 820         if r != 0:
 821             self.log.error('Error dumping crush map')
 822             return
 823         try:
 824             crushmap = json.loads(outb)
 825         except:
 826             raise RuntimeError('unable to parse crush map')
 827
 828         raw = crushmap.get('choose_args',{}).get('-1', [])
 829         weight_set = {}
 830         for b in raw:
 831             bucket = None
 832             for t in crushmap['buckets']:
 833                 if t['id'] == b['bucket_id']:
 834                     bucket = t
 835                     break
 836             if not bucket:
 837                 raise RuntimeError('could not find bucket %s' % b['bucket_id'])
 838             self.log.debug('bucket items %s' % bucket['items'])
 839             self.log.debug('weight set %s' % b['weight_set'][0])
 840             if len(bucket['items']) != len(b['weight_set'][0]):
 841                 raise RuntimeError('weight-set size does not match bucket items')
 842             for pos in range(len(bucket['items'])):
 843                 weight_set[bucket['items'][pos]['id']] = b['weight_set'][0][pos]
 844
 845         self.log.debug('weight_set weights %s' % weight_set)
 846         return weight_set
 847
 848     def do_crush(self):
 849         self.log.info('do_crush (not yet implemented)')
 850
 851     def do_osd_weight(self):
 852         self.log.info('do_osd_weight (not yet implemented)')
 853
 854     def execute(self, plan):
 855         self.log.info('Executing plan %s' % plan.name)
 856
 857         commands = []
 858
 859         # compat weight-set
 860         if len(plan.compat_ws) and \
 861            '-1' not in plan.initial.crush_dump.get('choose_args', {}):
 862             self.log.debug('ceph osd crush weight-set create-compat')
 863             result = CommandResult('')
 864             self.send_command(result, 'mon', '', json.dumps({
 865                 'prefix': 'osd crush weight-set create-compat',
 866                 'format': 'json',
 867             }), '')
 868             r, outb, outs = result.wait()
 869             if r != 0:
 870                 self.log.error('Error creating compat weight-set')
 871                 return
 872
 873         for osd, weight in plan.compat_ws.iteritems():
 874             self.log.info('ceph osd crush weight-set reweight-compat osd.%d %f',
 875                           osd, weight)
 876             result = CommandResult('foo')
 877             self.send_command(result, 'mon', '', json.dumps({
 878                 'prefix': 'osd crush weight-set reweight-compat',
 879                 'format': 'json',
 880                 'item': 'osd.%d' % osd,
 881                 'weight': [weight],
 882             }), 'foo')
 883             commands.append(result)
 884
 885         # new_weight
 886         reweightn = {}
 887         for osd, weight in plan.osd_weights.iteritems():
 888             reweightn[str(osd)] = str(int(weight * float(0x10000)))
 889         if len(reweightn):
 890             self.log.info('ceph osd reweightn %s', reweightn)
 891             result = CommandResult('foo')
 892             self.send_command(result, 'mon', '', json.dumps({
 893                 'prefix': 'osd reweightn',
 894                 'format': 'json',
 895                 'weights': json.dumps(reweightn),
 896             }), 'foo')
 897             commands.append(result)
 898
 899         # upmap
 900         incdump = plan.inc.dump()
 901         for pgid in incdump.get('old_pg_upmap_items', []):
 902             self.log.info('ceph osd rm-pg-upmap-items %s', pgid)
 903             result = CommandResult('foo')
 904             self.send_command(result, 'mon', '', json.dumps({
 905                 'prefix': 'osd rm-pg-upmap-items',
 906                 'format': 'json',
 907                 'pgid': pgid,
 908             }), 'foo')
 909             commands.append(result)
 910
 911         for item in incdump.get('new_pg_upmap_items', []):
 912             self.log.info('ceph osd pg-upmap-items %s mappings %s', item['pgid'],
 913                           item['mappings'])
 914             osdlist = []
 915             for m in item['mappings']:
 916                 osdlist += [m['from'], m['to']]
 917             result = CommandResult('foo')
 918             self.send_command(result, 'mon', '', json.dumps({
 919                 'prefix': 'osd pg-upmap-items',
 920                 'format': 'json',
 921                 'pgid': item['pgid'],
 922                 'id': osdlist,
 923             }), 'foo')
 924             commands.append(result)
 925
 926         # wait for commands
 927         self.log.debug('commands %s' % commands)
 928         for result in commands:
 929             r, outb, outs = result.wait()
 930             if r != 0:
 931                 self.log.error('Error on command')
 932                 return
 933         self.log.debug('done')