2 I contain utilities useful for calculating servers_of_happiness, and for
3 reporting it in messages
6 def failure_message(peer_count, k, happy, effective_happy):
7 # If peer_count < needed_shares, this error message makes more
8 # sense than any of the others, so use it.
10 msg = ("shares could be placed or found on only %d "
12 "We were asked to place shares on at least %d "
13 "server(s) such that any %d of them have "
14 "enough shares to recover the file." %
15 (peer_count, happy, k))
16 # Otherwise, if we've placed on at least needed_shares
17 # peers, but there isn't an x-happy subset of those peers
18 # for x >= needed_shares, we use this error message.
19 elif effective_happy < k:
20 msg = ("shares could be placed or found on %d "
21 "server(s), but they are not spread out evenly "
22 "enough to ensure that any %d of these servers "
23 "would have enough shares to recover the file. "
24 "We were asked to place "
25 "shares on at least %d servers such that any "
26 "%d of them have enough shares to recover the "
28 (peer_count, k, happy, k))
29 # Otherwise, if there is an x-happy subset of peers where
30 # x >= needed_shares, but x < servers_of_happiness, then
31 # we use this message.
33 msg = ("shares could be placed on only %d server(s) "
34 "such that any %d of them have enough shares "
35 "to recover the file, but we were asked to "
36 "place shares on at least %d such servers." %
37 (effective_happy, k, happy))
41 def shares_by_server(servermap):
43 I accept a dict of shareid -> set(peerid) mappings, and return a
44 dict of peerid -> set(shareid) mappings. My argument is a dictionary
45 with sets of peers, indexed by shares, and I transform that into a
46 dictionary of sets of shares, indexed by peerids.
49 for shareid, peers in servermap.iteritems():
50 assert isinstance(peers, set)
52 ret.setdefault(peerid, set()).add(shareid)
55 def merge_peers(servermap, used_peers=None):
57 I accept a dict of shareid -> set(peerid) mappings, and optionally a
58 set of PeerTrackers. If no set of PeerTrackers is provided, I return
59 my first argument unmodified. Otherwise, I update a copy of my first
60 argument to include the shareid -> peerid mappings implied in the
61 set of PeerTrackers, returning the resulting dict.
66 assert(isinstance(servermap, dict))
67 assert(isinstance(used_peers, set))
69 # Since we mutate servermap, and are called outside of a
70 # context where it is okay to do that, make a copy of servermap and
72 servermap = servermap.copy()
73 for peer in used_peers:
74 for shnum in peer.buckets:
75 servermap.setdefault(shnum, set()).add(peer.peerid)
78 def servers_of_happiness(sharemap):
80 I accept 'sharemap', a dict of shareid -> set(peerid) mappings. I
81 return the 'servers_of_happiness' number that sharemap results in.
83 To calculate the 'servers_of_happiness' number for the sharemap, I
84 construct a bipartite graph with servers in one partition of vertices
85 and shares in the other, and with an edge between a server s and a share t
86 if s is to store t. I then compute the size of a maximum matching in
87 the resulting graph; this is then returned as the 'servers_of_happiness'
90 For example, consider the following layout:
92 server 1: shares 1, 2, 3, 4
98 From this, we can construct the following graph:
100 L = {server 1, server 2, server 3, server 4, server 5}
101 R = {share 1, share 2, share 3, share 4, share 6}
103 E = {(server 1, share 1), (server 1, share 2), (server 1, share 3),
104 (server 1, share 4), (server 2, share 6), (server 3, share 3),
105 (server 4, share 4), (server 5, share 2)}
108 Note that G is bipartite since every edge in e has one endpoint in L
109 and one endpoint in R.
111 A matching in a graph G is a subset M of E such that, for any vertex
112 v in V, v is incident to at most one edge of M. A maximum matching
113 in G is a matching that is no smaller than any other matching. For
114 this graph, a matching of cardinality 5 is:
116 M = {(server 1, share 1), (server 2, share 6),
117 (server 3, share 3), (server 4, share 4),
120 Since G is bipartite, and since |L| = 5, we cannot have an M' such
121 that |M'| > |M|. Then M is a maximum matching in G. Intuitively, and
122 as long as k <= 5, we can see that the layout above has
123 servers_of_happiness = 5, which matches the results here.
127 sharemap = shares_by_server(sharemap)
128 graph = flow_network_for(sharemap)
129 # This is an implementation of the Ford-Fulkerson method for finding
130 # a maximum flow in a flow network applied to a bipartite graph.
131 # Specifically, it is the Edmonds-Karp algorithm, since it uses a
132 # BFS to find the shortest augmenting path at each iteration, if one
135 # The implementation here is an adapation of an algorithm described in
136 # "Introduction to Algorithms", Cormen et al, 2nd ed., pp 658-662.
138 flow_function = [[0 for sh in xrange(dim)] for s in xrange(dim)]
139 residual_graph, residual_function = residual_network(graph, flow_function)
140 while augmenting_path_for(residual_graph):
141 path = augmenting_path_for(residual_graph)
142 # Delta is the largest amount that we can increase flow across
143 # all of the edges in path. Because of the way that the residual
144 # function is constructed, f[u][v] for a particular edge (u, v)
145 # is the amount of unused capacity on that edge. Taking the
146 # minimum of a list of those values for each edge in the
147 # augmenting path gives us our delta.
148 delta = min(map(lambda (u, v): residual_function[u][v], path))
150 flow_function[u][v] += delta
151 flow_function[v][u] -= delta
152 residual_graph, residual_function = residual_network(graph,
154 num_servers = len(sharemap)
155 # The value of a flow is the total flow out of the source vertex
156 # (vertex 0, in our graph). We could just as well sum across all of
157 # f[0], but we know that vertex 0 only has edges to the servers in
158 # our graph, so we can stop after summing flow across those. The
159 # value of a flow computed in this way is the size of a maximum
160 # matching on the bipartite graph described above.
161 return sum([flow_function[0][v] for v in xrange(1, num_servers+1)])
163 def flow_network_for(sharemap):
165 I take my argument, a dict of peerid -> set(shareid) mappings, and
166 turn it into a flow network suitable for use with Edmonds-Karp. I
167 then return the adjacency list representation of that network.
169 Specifically, I build G = (V, E), where:
170 V = { peerid in sharemap } U { shareid in sharemap } U {s, t}
171 E = {(s, peerid) for each peerid}
172 U {(peerid, shareid) if peerid is to store shareid }
173 U {(shareid, t) for each shareid}
175 s and t will be source and sink nodes when my caller starts treating
176 the graph I return like a flow network. Without s and t, the
177 returned graph is bipartite.
179 # Servers don't have integral identifiers, and we can't make any
180 # assumptions about the way shares are indexed -- it's possible that
181 # there are missing shares, for example. So before making a graph,
182 # we re-index so that all of our vertices have integral indices, and
183 # that there aren't any holes. We start indexing at 1, so that we
184 # can add a source node at index 0.
185 sharemap, num_shares = reindex(sharemap, base_index=1)
186 num_servers = len(sharemap)
187 graph = [] # index -> [index], an adjacency list
188 # Add an entry at the top (index 0) that has an edge to every server
190 graph.append(sharemap.keys())
191 # For each server, add an entry that has an edge to every share that it
192 # contains (or will contain).
194 graph.append(sharemap[k])
195 # For each share, add an entry that has an edge to the sink.
196 sink_num = num_servers + num_shares + 1
197 for i in xrange(num_shares):
198 graph.append([sink_num])
199 # Add an empty entry for the sink, which has no outbound edges.
203 def reindex(sharemap, base_index):
205 Given sharemap, I map peerids and shareids to integers that don't
206 conflict with each other, so they're useful as indices in a graph. I
207 return a sharemap that is reindexed appropriately, and also the
208 number of distinct shares in the resulting sharemap as a convenience
209 for my caller. base_index tells me where to start indexing.
211 shares = {} # shareid -> vertex index
213 ret = {} # peerid -> [shareid], a reindexed sharemap.
214 # Number the servers first
216 ret[num] = sharemap[k]
221 if not shares.has_key(shnum):
224 ret[k] = map(lambda x: shares[x], ret[k])
225 return (ret, len(shares))
227 def residual_network(graph, f):
229 I return the residual network and residual capacity function of the
230 flow network represented by my graph and f arguments. graph is a
231 flow network in adjacency-list form, and f is a flow in graph.
233 new_graph = [[] for i in xrange(len(graph))]
234 cf = [[0 for s in xrange(len(graph))] for sh in xrange(len(graph))]
235 for i in xrange(len(graph)):
238 # We add an edge (v, i) with cf[v,i] = 1. This means
239 # that we can remove 1 unit of flow from the edge (i, v)
240 new_graph[v].append(i)
244 # We add the edge (i, v), since we're not using it right
246 new_graph[i].append(v)
249 return (new_graph, cf)
251 def augmenting_path_for(graph):
253 I return an augmenting path, if there is one, from the source node
254 to the sink node in the flow network represented by my graph argument.
255 If there is no augmenting path, I return False. I assume that the
256 source node is at index 0 of graph, and the sink node is at the last
257 index. I also assume that graph is a flow network in adjacency list
260 bfs_tree = bfs(graph, 0)
261 if bfs_tree[len(graph) - 1]:
263 path = [] # [(u, v)], where u and v are vertices in the graph
265 path.insert(0, (bfs_tree[n], n))
272 Perform a BFS on graph starting at s, where graph is a graph in
273 adjacency list form, and s is a node in graph. I return the
274 predecessor table that the BFS generates.
276 # This is an adaptation of the BFS described in "Introduction to
277 # Algorithms", Cormen et al, 2nd ed., p. 532.
278 # WHITE vertices are those that we haven't seen or explored yet.
280 # GRAY vertices are those we have seen, but haven't explored yet
282 # BLACK vertices are those we have seen and explored
284 color = [WHITE for i in xrange(len(graph))]
285 predecessor = [None for i in xrange(len(graph))]
286 distance = [-1 for i in xrange(len(graph))]
287 queue = [s] # vertices that we haven't explored yet.
293 if color[v] == WHITE:
295 distance[v] = distance[n] + 1