dataDistribution.py 20 KB
Newer Older
1
import random
2
3
4
import copy

from ortools.graph import pywrapgraph
5

Felix Seibert's avatar
Felix Seibert committed
6
from xtreemfs_client import osd
7
from xtreemfs_client import folder
8

9

Felix Seibert's avatar
Felix Seibert committed
10
11
12
13
class DataDistribution(object):
    """
    class to keep track of the osd (object storage device) locations of different folders, i.e.,
    their physical location.
14

Felix Seibert's avatar
Felix Seibert committed
15
16
    this class also allows to calculate several data distributions, e.g., mappings from folders to OSDs (each folder
    gets mapped to one OSD).
Felix Seibert's avatar
Felix Seibert committed
17
    """
18
19

    def __init__(self):
20
        self.OSDs = {}
21

22
    def add_new_osd(self, osd_uuid):
Felix Seibert's avatar
Felix Seibert committed
23
24
25
        """
        create a new empty osd and add it to the existing OSDs.
        """
26
27
        if osd_uuid in self.OSDs:
            print("key: " + osd_uuid + " is already present!")
28
            return
29
30
        new_osd = osd.OSD(osd_uuid)
        self.OSDs[osd_uuid] = new_osd
31

32
    def add_osd(self, new_osd):
Felix Seibert's avatar
Felix Seibert committed
33
34
35
        """
        add the given OSD (object) to the existing OSDs.
        """
36
37
        if new_osd.uuid in self.OSDs:
            print("key: " + new_osd.uuid + " is already present!")
38
            return
39
        self.OSDs[new_osd.uuid] = new_osd
40

41
    def add_osd_list(self, osd_list):
Felix Seibert's avatar
Felix Seibert committed
42
43
44
        """
        add the given list of OSDs (objects) to the existing OSDs.
        """
45
46
47
48
        for osd_uuid in osd_list:
            if osd_uuid not in self.OSDs:
                new_osd = osd.OSD(osd_uuid)
                self.OSDs[osd_uuid] = new_osd
49

50
51
52
53
54
55
56
57
58
    def replace_osd(self, new_osd):
        """
        replaces the osd with uuid new_osd.uuid by new_osd
        :param new_osd:
        :return:
        """
        assert new_osd.uuid in self.OSDs.keys()
        self.OSDs[new_osd.uuid] = new_osd

59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
    def set_osd_capacities(self, osd_capacities):
        """
        set osd capacities
        :param osd_capacities: map from osd uuids to osd capacities
        :return:
        """
        for one_osd in self.OSDs.values():
            assert type(osd_capacities[one_osd.uuid]) is int
            one_osd.capacity = osd_capacities[one_osd.uuid]

    def set_osd_bandwidths(self, osd_bandwidths):
        """
        set osd bandwidths
        :param osd_bandwidths:
        :return:
        """
        for one_osd in self.OSDs.values():
            one_osd.bandwidth = osd_bandwidths[one_osd.uuid]

78
    def get_osd_list(self):
Felix Seibert's avatar
Felix Seibert committed
79
        """
80
        get a list of all existing OSD uuids.
Felix Seibert's avatar
Felix Seibert committed
81
        """
82
        osd_list = []
83
        for osd_name in self.OSDs.keys():
84
85
86
            osd_list.append(osd_name)
        return osd_list

87
    def get_containing_osd(self, folder_id):
Felix Seibert's avatar
Felix Seibert committed
88
89
90
        """
        get the OSD containing the given folder_id, or None if the folder is not assigned to any OSD.
        """
91
92
93
        for checked_osd in self.OSDs.values():
            if checked_osd.contains_folder(folder_id):
                return checked_osd
94
95
        return None

Felix Seibert's avatar
Felix Seibert committed
96
97
98
99
100
101
102
103
104
105
106
    def assign_new_osd(self, folder_id, new_osd):
        """
        assign folder_id to new_osd. if folder_id already is assigned to an OSD, this old assignment is deleted.
        """
        old_osd = self.get_containing_osd(folder_id)
        if old_osd is None:
            self.OSDs[new_osd].add_folder(folder_id, self.get_average_folder_size())
        else:
            self.OSDs[new_osd].add_folder(folder_id, self.OSDs[old_osd.uuid].folders[folder_id])
            self.OSDs[old_osd.uuid].remove_folder(folder_id)

107
    def get_average_folder_size(self):
Felix Seibert's avatar
Felix Seibert committed
108
109
110
        """
        get the average folder size of all folders of all OSDs.
        """
111
        total_size = 0
112
        total_number_of_folders = 0
113
        for one_osd in self.OSDs.values():
Felix Seibert's avatar
Felix Seibert committed
114
            total_size += one_osd.total_folder_size
115
            total_number_of_folders += len(one_osd.folders)
116
117
        if total_number_of_folders == 0:
            return 0
118
119
        return total_size / total_number_of_folders

120
    def get_average_load(self):
121
        """
122
        calculate the average OSD load, that is, the average of their total_folder_size.
123
124
        """
        total_folder_size = 0
125
126
127
        for osd in self.OSDs.values():
            total_folder_size += osd.get_load()
        return total_folder_size / len(self.OSDs)
128

129
    def get_maximum_load(self):
130
        """
131
        calculate the maximum OSD load, that is, the maximum of their total_folder_size.
132
133
134
135
        """
        maximum_load = 0
        maximum_osd = None
        for osd in self.OSDs.values():
136
            load = osd.total_folder_size
137
138
139
140
141
            if maximum_osd is None or load > maximum_load:
                maximum_load = load
                maximum_osd = osd
        return maximum_osd, maximum_load

142
    def get_average_processing_time(self):
143
        """
144
145
        calculate the average OSD processing time, that is, the average of their (total_folder_size / bandwidth).
        :return:
146
        """
147
        total_processing_time = 0
148
        for osd in self.OSDs.values():
149
150
            total_processing_time += osd.get_processing_time()
        return total_processing_time / len(self.OSDs)
151

152
153
154
155
156
157
158
159
160
161
162
163
    def get_maximum_processing_time(self):
        """
        calculate the maximum OSD processing time, also known as makespan
        """
        maximum_processing_time = 0
        maximum_osd = None
        for osd in self.OSDs.values():
            processing_time = osd.get_processing_time()
            if maximum_osd is None or processing_time > maximum_processing_time:
                maximum_processing_time = processing_time
                maximum_osd = osd
        return maximum_osd, maximum_processing_time
164

165
166
167
    def add_folders(self, folders,
                    ignore_osd_capacities=True,
                    random_osd_assignment=False,
168
169
                    ignore_folder_sizes=False,
                    debug=False):
Felix Seibert's avatar
Felix Seibert committed
170
171
        """
        adds a list of folders to the data distribution.
Felix Seibert's avatar
Felix Seibert committed
172
        if not specified otherwise, the assignments are calculated using the LPT algorithm.
Felix Seibert's avatar
Felix Seibert committed
173
174
        returns a list of assignments from folders to OSDs, for which (folders) there was previously no assignment.

175
176
        if capacities and bandwidths are set for the OSDs, folders are assigned accordingly
        (capacities are respected and OSDs with higher bandwidth obtain more/larger files).
177
178
179
180
181

        if random_osd_assignment=True and ignore_osd_capacities=True, a totally random OSD assignment generated.

        if random_osd_assignment=True and ignore_folder_sizes=True,
        folders are randomly assigned to OSDs such that all OSDs have the same number of folders (if possible).
182
183

        the assignment is stable (i.e., folders already assigned to an OSD are not reassigned to another OSD).
Felix Seibert's avatar
Felix Seibert committed
184
        """
185
186

        # find out which folders are not assigned yet
187
        new_folders = []
188
189
        for a_folder in folders:
            containing_osd = self.get_containing_osd(a_folder.id)
190
            if containing_osd is not None:
191
                containing_osd.add_folder(a_folder.id, a_folder.size)
192
            else:
193
                new_folders.append(a_folder)
194

195
196
        if debug:
            print("dataDistribution: random_osd_assignment: " + str(random_osd_assignment))
Felix Seibert's avatar
Felix Seibert committed
197

198
199
        # keep track of which unassigned folder gets assigned to which OSD.
        # this information must be returned
200
201
        osds_for_new_folders = []

202
203
        # totally random OSD assignment, even ignoring OSD capacities
        # (might lead to I/O errors when too many groups are assigned to an OSD)
204
        if random_osd_assignment and ignore_osd_capacities and not ignore_folder_sizes:
205
206
            if debug:
                print("using totally random osd assignment")
207
            for a_folder in new_folders:
Felix Seibert's avatar
Felix Seibert committed
208
                random_osd = random.choice(list(self.OSDs.values()))
209
210
                random_osd.add_folder(a_folder.id, a_folder.size)
                osds_for_new_folders.append((a_folder.id,
Felix Seibert's avatar
Felix Seibert committed
211
                                             random_osd.uuid))
Felix Seibert's avatar
Felix Seibert committed
212
            return osds_for_new_folders
213

214
215
        # random OSD assignment respecting OSD capacities
        elif random_osd_assignment and not ignore_osd_capacities:
216
217
            if debug:
                print("using random osd assignment, respecting osd capacities")
218
            for a_folder in new_folders:
219
                suitable_osds = []  # list of OSDs with enough capacity
220
                for one_osd in self.OSDs.values():
221
                    if one_osd.capacity - one_osd.total_folder_size - a_folder.size >= 0:
222
223
                        suitable_osds.append(one_osd)
                suitable_random_osd = random.choice(suitable_osds)
224
225
                suitable_random_osd.add_folder(a_folder.id, a_folder.size)
                osds_for_new_folders.append((a_folder.id,
226
227
228
                                             suitable_random_osd.uuid))
            return osds_for_new_folders

229
        # random OSD assignment ignoring folder sizes // round-robin style distribution with some randomness
230
        elif random_osd_assignment and ignore_folder_sizes:
231
232
233
234
235
236
237
238
239
240
            if debug:
                print("using random osd assignment ignoring folder sizes")

            average_folder_size = self.get_average_folder_size()
            if average_folder_size == 0:
                average_folder_size = 1

            modified_folders = list(map(lambda f: folder.Folder(f.id, average_folder_size, f.origin), folders))
            random.shuffle(modified_folders)
            return self.add_folders(modified_folders)
241

242
243
        # balanced deterministic OSD assignment (LPT)
        # (following largest processing time first, also called post-greedy approach)
244
245
        list.sort(new_folders, key=lambda x: x.size, reverse=True)

246
        # for each folder calculate the best OSD and add it to it
247
        for a_folder in new_folders:
248
            least_used_osd, _ = self.get_lpt_osd(a_folder.size)
249
250
            least_used_osd.add_folder(a_folder.id, a_folder.size)
            osds_for_new_folders.append((a_folder.id,
251
                                         least_used_osd.uuid))
252
253
        return osds_for_new_folders

254
    def rebalance_lpt(self, rebalance_factor=1):
255
256
257
        """
        rebalance folders to OSDs by assigning folders to new OSDs using the following strategy:
                1. 'unroll' the assignment. this means that, for each OSD, folders are removed until the OSD has less
258
                processing time than the average processing time of this distribution multiplied by rebalance_factor.
259
260
                2. reassign the removed folders using the LPT strategy.
        """
261
262
        movements = {}
        folders_to_be_reassigned = []
263
264
265
        # TODO reassignment factor based on load or based on processing time?
        # as long as we use OSDs with the same bandwidth, there is no difference.
        reassignment_limit = self.get_average_processing_time() * rebalance_factor
266
267

        # for each OSD, remove the smallest folder until its total_folder_size does not exceed the reassignment_limit
268
        # unrolling
269
        for osd in self.OSDs.values():
270
271
272
273
274
            # TODO how to calculate the 'unrolling limit' for each individual OSD?
            # again, as long as all OSDs have the same bandwidth, there is no difference.
            # BUT it should definitely not depend on the load or pt of the OSD.
            # so for now we use a static limit, the same for all OSDs.
            while osd.get_processing_time() > reassignment_limit:
275
276
277
278
279
                folder_id, folder_size = osd.get_smallest_folder()
                folders_to_be_reassigned.append(folder.Folder(folder_id, folder_size, None))
                movements[folder_id] = osd.uuid
                osd.remove_folder(folder_id)

280
        # reassignment
281
        new_assignments = self.add_folders(folders_to_be_reassigned)
282
283
284
285

        for folder_id, target in new_assignments:
            movements[folder_id] = (movements[folder_id], target)

286
        return movements, reassignment_limit
287

288
    def rebalance_one_folder(self):
289
290
        """
        rebalance folders to OSDs by assigning folders to new OSDs using the following strategy:
291
                1. find OSD with the highest processing time
292
293
                2. get folder with smallest size on this OSD
                3. find new OSD for this folder using get_lpt_osd
294
295
                4. if the processing time on the new OSD is lower than on the original OSD,
                move the folder to the new OSD. otherwise, return.
296
297
298
299
300
301
302
303
304
305
306
307
        one open question is whether getting the folder with smallest size in step 2 is a clever choice
        (in principle, all folders of the OSD with the highest load are eligible).

        this optimization scheme classifies as local search. two distributions are neighbors if one can be transformed
        into the other by moving one folder from one OSD to another. note, however, that we do not search the whole
        neighborhood of a distribution.
        but it might be possible to show that if there is no improvement step of the type that we check for,
        there is no improvement step at all.
        """
        movements = {}

        while True:
308
309
            # find OSD with the highest processing time (origin)
            origin_osd, maximum_processing_time = self.get_maximum_processing_time()
310
311
312
313
314
315
316
317
318

            # pick a folder of this OSD
            # there are several ways to pick a folder (like largest, smallest, constrained by the resulting load of the
            # origin OSD, random...), it is not clear which way is a good way
            # for now pick the smallest folder on origin OSD
            smallest_folder_id, smallest_folder_size = self.OSDs[origin_osd.uuid].get_smallest_folder()

            # find other OSD best suited for the picked folder (target)
            # check whether moving folder from origin to target decreases the maximum load of all OSDs (makespan).
319
            best_osd, best_osd_processing_time = self.get_lpt_osd(smallest_folder_size)
320

321
            if best_osd_processing_time < maximum_processing_time:
322
323
324
325
326
327
328
                self.assign_new_osd(smallest_folder_id, best_osd.uuid)
                movements[smallest_folder_id] = (origin_osd.uuid, best_osd.uuid)
            else:
                break

        return movements

329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
    def rebalance_two_steps(self):
        """
        rebalance the distribution in two steps:
            1. calculate new distribution, independently of the current one
            2. use a minimum weight matching to transform the current distribution into the new distribution.
            minimum weight perfect matching on bipartite graphs can be solved using the successive shortest path
            algorithm.
        while any algorithm (solving/approximating that kind of problem) could be used for the first step,
        we here only implement the LPT algorithm, as it is a pretty good approximation with extremely good running time.
        :return:
        """
        virtual_distribution = copy.deepcopy(self)
        virtual_distribution.rebalance_lpt(rebalance_factor=0)

        # create a mincostflow object
        min_cost_flow = pywrapgraph.SimpleMinCostFlow()

        # define the directed graph for the flow
        # arcs are added individually, and are added implicitly
        # nodes (OSDs) have to be given by numeric id
        # so we need some conversion logic between current/virtual osds and node ids

        current_osds_list = list(self.OSDs.values())
        current_osds_list.sort(key=lambda x: x.uuid)
        virtual_osds_list = list(virtual_distribution.OSDs.values())
        virtual_osds_list.sort(key=lambda x: x.uuid)

        # conversion logic:
        # n = len(current_osd_list) = len(virtual_osd_list)
        # 0 = source, 1 = sink
        # 2, ..., n + 1: current OSDs
        # n + 2, ..., 2n + 1: virtual OSDs
        num_osds = len(current_osds_list)
        assert num_osds == len(virtual_osds_list)

        # edges between the two partitions
        for i in range(0, num_osds):
            for j in range(0, num_osds):
                current_osd = current_osds_list[i]
                virtual_osd = virtual_osds_list[j]
                # calculate the total size of folders that the current OSD has to fetch if the virtual OSD is assigned
                # to it
                edge_cost = 0
                for folder_id in virtual_osd.folders.keys():
                    if not current_osd.contains_folder(folder_id):
                        edge_cost += virtual_osd.folders[folder_id]
                tail = 2 + i  # current OSD
                head = num_osds + 2 + j  # virtual OSD
                min_cost_flow.AddArcWithCapacityAndUnitCost(tail, head, 1, edge_cost)

        # (artificial) edges between the source node and the current OSDs
        for i in range(0, num_osds):
            edge_cost = 0
            tail = 0
            head = i + 2
            min_cost_flow.AddArcWithCapacityAndUnitCost(tail, head, 1, edge_cost)

        # (artificial) edges between the virtual OSDs and the sink node
        for j in range(0, num_osds):
            edge_cost = 0
            tail = num_osds + 2 + j
            head = 1
            min_cost_flow.AddArcWithCapacityAndUnitCost(tail, head, 1, edge_cost)

        # define the supplies (which equals the number of OSDs)
        min_cost_flow.SetNodeSupply(0, num_osds)
        min_cost_flow.SetNodeSupply(1, -num_osds)

        # solve the min cost flow problem
        min_cost_flow.Solve()

        # we need to transform the calculated optimal assignment into a rebalanced distribution, including the necessary
        # movements
        current_to_virtual_osd_matching = []
        for arc in range(min_cost_flow.NumArcs()):
            tail = min_cost_flow.Tail(arc)
            head = min_cost_flow.Head(arc)
            if tail != 0 and head != 1 and min_cost_flow.Flow(arc) == 1:
                current_osd = current_osds_list[tail - 2]
                virtual_osd = virtual_osds_list[head - num_osds - 2]
                current_to_virtual_osd_matching.append((current_osd, virtual_osd))

        movements = {}
        for current_osd, virtual_osd in current_to_virtual_osd_matching:
            # iterate over virtual folders and check whether they are on the correct OSD.
            # the correct OSD is current_osd, as it is the one that is matched with virtual_osd.
            # if it is not present on current_osd, assign it to it.
            # this also removes it from the origin osd.
            for virtual_folder in virtual_osd.folders.keys():
                if not current_osd.contains_folder(virtual_folder):
                    origin_osd = self.get_containing_osd(virtual_folder).uuid
                    target_osd = current_osd.uuid
                    movements[virtual_folder] = (origin_osd, target_osd)
422
423
424
425

        for current_osd, virtual_osd in current_to_virtual_osd_matching:
            virtual_osd.uuid = current_osd.uuid
            self.replace_osd(virtual_osd)
426
427
428

        return movements

429
    def get_lpt_osd(self, folder_size):
430
        """
431
        calculate the processing time of all OSDs, using the sum of their current total_folder_size and folder_size.
432
433
        return (OSD with the smallest such value, the smallest value)
        """
434
435
        best_processing_time = None
        best_processing_time_osd = -1
436
        for one_osd in self.OSDs.values():
437
438
439
440
441
            processing_time = (one_osd.total_folder_size + folder_size) / one_osd.bandwidth
            if (best_processing_time is None) or processing_time < best_processing_time_osd:
                best_processing_time = one_osd
                best_processing_time_osd = processing_time
        return best_processing_time, best_processing_time_osd
442

443
    def update_folder(self, folder, size):
Felix Seibert's avatar
Felix Seibert committed
444
445
446
        """
        updates the size of a given folder
        """
447
448
449
        for one_osd in self.OSDs.values():
            if folder in one_osd.folders.keys():
                one_osd.update_folder(folder, size)
450
451
                break

452
    def description(self):
Felix Seibert's avatar
Felix Seibert committed
453
454
455
        """
        generates a string describing this data distribution
        """
456
        string = ""
457
458
        for one_osd in self.OSDs.values():
            string += str(one_osd)
459
            string += "\n"
460
            string += "folders : " + str(one_osd.folders)
461
            string += "\n"
462
        string += "average folder size: " + str(self.get_average_folder_size())
463
464
465
        return string

    def __str__(self):
466
467
468
469
470
        string_representation = "DataDistribution has " + str(len(self.OSDs)) \
                                + " osds: \n"
        for key, value in self.OSDs.items():
            string_representation += str(value) + " \n"
        return string_representation