dataDistribution.py 19.7 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
59
60
61
62
63
64
65
66
67
68
    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]

69
    def get_osd_list(self):
Felix Seibert's avatar
Felix Seibert committed
70
        """
71
        get a list of all existing OSD uuids.
Felix Seibert's avatar
Felix Seibert committed
72
        """
73
        osd_list = []
74
        for osd_name in self.OSDs.keys():
75
76
77
            osd_list.append(osd_name)
        return osd_list

78
    def get_containing_osd(self, folder_id):
Felix Seibert's avatar
Felix Seibert committed
79
80
81
        """
        get the OSD containing the given folder_id, or None if the folder is not assigned to any OSD.
        """
82
83
84
        for checked_osd in self.OSDs.values():
            if checked_osd.contains_folder(folder_id):
                return checked_osd
85
86
        return None

Felix Seibert's avatar
Felix Seibert committed
87
88
89
90
91
92
93
94
95
96
97
    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)

98
    def get_average_folder_size(self):
Felix Seibert's avatar
Felix Seibert committed
99
100
101
        """
        get the average folder size of all folders of all OSDs.
        """
102
        total_size = 0
103
        total_number_of_folders = 0
104
        for one_osd in self.OSDs.values():
Felix Seibert's avatar
Felix Seibert committed
105
            total_size += one_osd.total_folder_size
106
            total_number_of_folders += len(one_osd.folders)
107
108
        if total_number_of_folders == 0:
            return 0
109
110
        return total_size / total_number_of_folders

111
    def get_average_load(self):
112
        """
113
        calculate the average OSD load, that is, the average of their total_folder_size.
114
115
        """
        total_folder_size = 0
116
117
118
        for osd in self.OSDs.values():
            total_folder_size += osd.get_load()
        return total_folder_size / len(self.OSDs)
119

120
    def get_maximum_load(self):
121
        """
122
        calculate the maximum OSD load, that is, the maximum of their total_folder_size.
123
124
125
126
        """
        maximum_load = 0
        maximum_osd = None
        for osd in self.OSDs.values():
127
            load = osd.total_folder_size
128
129
130
131
132
            if maximum_osd is None or load > maximum_load:
                maximum_load = load
                maximum_osd = osd
        return maximum_osd, maximum_load

133
    def get_average_processing_time(self):
134
        """
135
136
        calculate the average OSD processing time, that is, the average of their (total_folder_size / bandwidth).
        :return:
137
        """
138
        total_processing_time = 0
139
        for osd in self.OSDs.values():
140
141
            total_processing_time += osd.get_processing_time()
        return total_processing_time / len(self.OSDs)
142

143
144
145
146
147
148
149
150
151
152
153
154
    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
155

156
157
158
    def add_folders(self, folders,
                    ignore_osd_capacities=True,
                    random_osd_assignment=False,
159
160
                    ignore_folder_sizes=False,
                    debug=False):
Felix Seibert's avatar
Felix Seibert committed
161
162
        """
        adds a list of folders to the data distribution.
Felix Seibert's avatar
Felix Seibert committed
163
        if not specified otherwise, the assignments are calculated using the LPT algorithm.
Felix Seibert's avatar
Felix Seibert committed
164
165
        returns a list of assignments from folders to OSDs, for which (folders) there was previously no assignment.

166
167
        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).
168
169
170
171
172

        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).
173
174

        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
175
        """
176
177

        # find out which folders are not assigned yet
178
        new_folders = []
179
180
        for a_folder in folders:
            containing_osd = self.get_containing_osd(a_folder.id)
181
            if containing_osd is not None:
182
                containing_osd.add_folder(a_folder.id, a_folder.size)
183
            else:
184
                new_folders.append(a_folder)
185

186
187
        if debug:
            print("dataDistribution: random_osd_assignment: " + str(random_osd_assignment))
Felix Seibert's avatar
Felix Seibert committed
188

189
190
        # keep track of which unassigned folder gets assigned to which OSD.
        # this information must be returned
191
192
        osds_for_new_folders = []

193
194
        # totally random OSD assignment, even ignoring OSD capacities
        # (might lead to I/O errors when too many groups are assigned to an OSD)
195
        if random_osd_assignment and ignore_osd_capacities and not ignore_folder_sizes:
196
197
            if debug:
                print("using totally random osd assignment")
198
            for a_folder in new_folders:
Felix Seibert's avatar
Felix Seibert committed
199
                random_osd = random.choice(list(self.OSDs.values()))
200
201
                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
202
                                             random_osd.uuid))
Felix Seibert's avatar
Felix Seibert committed
203
            return osds_for_new_folders
204

205
206
        # random OSD assignment respecting OSD capacities
        elif random_osd_assignment and not ignore_osd_capacities:
207
208
            if debug:
                print("using random osd assignment, respecting osd capacities")
209
            for a_folder in new_folders:
210
                suitable_osds = []  # list of OSDs with enough capacity
211
                for one_osd in self.OSDs.values():
212
                    if one_osd.capacity - one_osd.total_folder_size - a_folder.size >= 0:
213
214
                        suitable_osds.append(one_osd)
                suitable_random_osd = random.choice(suitable_osds)
215
216
                suitable_random_osd.add_folder(a_folder.id, a_folder.size)
                osds_for_new_folders.append((a_folder.id,
217
218
219
                                             suitable_random_osd.uuid))
            return osds_for_new_folders

220
        # random OSD assignment ignoring folder sizes // round-robin style distribution with some randomness
221
        elif random_osd_assignment and ignore_folder_sizes:
222
223
224
225
226
227
228
229
230
231
            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)
232

233
234
        # balanced deterministic OSD assignment (LPT)
        # (following largest processing time first, also called post-greedy approach)
235
236
        list.sort(new_folders, key=lambda x: x.size, reverse=True)

237
        # for each folder calculate the best OSD and add it to it
238
        for a_folder in new_folders:
239
            least_used_osd, _ = self.get_lpt_osd(a_folder.size)
240
241
            least_used_osd.add_folder(a_folder.id, a_folder.size)
            osds_for_new_folders.append((a_folder.id,
242
                                         least_used_osd.uuid))
243
244
        return osds_for_new_folders

245
    def rebalance_lpt(self, rebalance_factor=1):
246
247
248
        """
        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
249
                processing time than the average processing time of this distribution multiplied by rebalance_factor.
250
251
                2. reassign the removed folders using the LPT strategy.
        """
252
253
        movements = {}
        folders_to_be_reassigned = []
254
255
256
        # 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
257
258

        # for each OSD, remove the smallest folder until its total_folder_size does not exceed the reassignment_limit
259
        # unrolling
260
        for osd in self.OSDs.values():
261
262
263
264
265
            # 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:
266
267
268
269
270
                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)

271
        # reassignment
272
        new_assignments = self.add_folders(folders_to_be_reassigned)
273
274
275
276

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

277
        return movements, reassignment_limit
278

279
    def rebalance_one_folder(self):
280
281
        """
        rebalance folders to OSDs by assigning folders to new OSDs using the following strategy:
282
                1. find OSD with the highest processing time
283
284
                2. get folder with smallest size on this OSD
                3. find new OSD for this folder using get_lpt_osd
285
286
                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.
287
288
289
290
291
292
293
294
295
296
297
298
        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:
299
300
            # find OSD with the highest processing time (origin)
            origin_osd, maximum_processing_time = self.get_maximum_processing_time()
301
302
303
304
305
306
307
308
309

            # 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).
310
            best_osd, best_osd_processing_time = self.get_lpt_osd(smallest_folder_size)
311

312
            if best_osd_processing_time < maximum_processing_time:
313
314
315
316
317
318
319
                self.assign_new_osd(smallest_folder_id, best_osd.uuid)
                movements[smallest_folder_id] = (origin_osd.uuid, best_osd.uuid)
            else:
                break

        return movements

320
321
322
323
324
325
326
327
328
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
    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)
                    self.assign_new_osd(virtual_folder, target_osd)

        return movements

417
    def get_lpt_osd(self, folder_size):
418
        """
419
        calculate the processing time of all OSDs, using the sum of their current total_folder_size and folder_size.
420
421
        return (OSD with the smallest such value, the smallest value)
        """
422
423
        best_processing_time = None
        best_processing_time_osd = -1
424
        for one_osd in self.OSDs.values():
425
426
427
428
429
            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
430

431
    def update_folder(self, folder, size):
Felix Seibert's avatar
Felix Seibert committed
432
433
434
        """
        updates the size of a given folder
        """
435
436
437
        for one_osd in self.OSDs.values():
            if folder in one_osd.folders.keys():
                one_osd.update_folder(folder, size)
438
439
                break

440
    def description(self):
Felix Seibert's avatar
Felix Seibert committed
441
442
443
        """
        generates a string describing this data distribution
        """
444
        string = ""
445
446
        for one_osd in self.OSDs.values():
            string += str(one_osd)
447
            string += "\n"
448
            string += "folders : " + str(one_osd.folders)
449
            string += "\n"
450
        string += "average folder size: " + str(self.get_average_folder_size())
451
452
453
        return string

    def __str__(self):
454
455
456
457
458
        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