1 | .. -*- coding: utf-8-with-signature -*- |
---|
2 | |
---|
3 | ==================== |
---|
4 | Servers of Happiness |
---|
5 | ==================== |
---|
6 | |
---|
7 | When you upload a file to a Tahoe-LAFS grid, you expect that it will |
---|
8 | stay there for a while, and that it will do so even if a few of the |
---|
9 | peers on the grid stop working, or if something else goes wrong. An |
---|
10 | upload health metric helps to make sure that this actually happens. |
---|
11 | An upload health metric is a test that looks at a file on a Tahoe-LAFS |
---|
12 | grid and says whether or not that file is healthy; that is, whether it |
---|
13 | is distributed on the grid in such a way as to ensure that it will |
---|
14 | probably survive in good enough shape to be recoverable, even if a few |
---|
15 | things go wrong between the time of the test and the time that it is |
---|
16 | recovered. Our current upload health metric for immutable files is called |
---|
17 | 'servers-of-happiness'; its predecessor was called 'shares-of-happiness'. |
---|
18 | |
---|
19 | shares-of-happiness used the number of encoded shares generated by a |
---|
20 | file upload to say whether or not it was healthy. If there were more |
---|
21 | shares than a user-configurable threshold, the file was reported to be |
---|
22 | healthy; otherwise, it was reported to be unhealthy. In normal |
---|
23 | situations, the upload process would distribute shares fairly evenly |
---|
24 | over the peers in the grid, and in that case shares-of-happiness |
---|
25 | worked fine. However, because it only considered the number of shares, |
---|
26 | and not where they were on the grid, it could not detect situations |
---|
27 | where a file was unhealthy because most or all of the shares generated |
---|
28 | from the file were stored on one or two peers. |
---|
29 | |
---|
30 | servers-of-happiness addresses this by extending the share-focused |
---|
31 | upload health metric to also consider the location of the shares on |
---|
32 | grid. servers-of-happiness looks at the mapping of peers to the shares |
---|
33 | that they hold, and compares the cardinality of the largest happy subset |
---|
34 | of those to a user-configurable threshold. A happy subset of peers has |
---|
35 | the property that any k (where k is as in k-of-n encoding) peers within |
---|
36 | the subset can reconstruct the source file. This definition of file |
---|
37 | health provides a stronger assurance of file availability over time; |
---|
38 | with 3-of-10 encoding, and happy=7, a healthy file is still guaranteed |
---|
39 | to be available even if 4 peers fail. |
---|
40 | |
---|
41 | Measuring Servers of Happiness |
---|
42 | ============================== |
---|
43 | |
---|
44 | We calculate servers-of-happiness by computing a matching on a |
---|
45 | bipartite graph that is related to the layout of shares on the grid. |
---|
46 | One set of vertices is the peers on the grid, and one set of vertices is |
---|
47 | the shares. An edge connects a peer and a share if the peer will (or |
---|
48 | does, for existing shares) hold the share. The size of the maximum |
---|
49 | matching on this graph is the size of the largest happy peer set that |
---|
50 | exists for the upload. |
---|
51 | |
---|
52 | First, note that a bipartite matching of size n corresponds to a happy |
---|
53 | subset of size n. This is because a bipartite matching of size n implies |
---|
54 | that there are n peers such that each peer holds a share that no other |
---|
55 | peer holds. Then any k of those peers collectively hold k distinct |
---|
56 | shares, and can restore the file. |
---|
57 | |
---|
58 | A bipartite matching of size n is not necessary for a happy subset of |
---|
59 | size n, however (so it is not correct to say that the size of the |
---|
60 | maximum matching on this graph is the size of the largest happy subset |
---|
61 | of peers that exists for the upload). For example, consider a file with |
---|
62 | k = 3, and suppose that each peer has all three of those pieces. Then, |
---|
63 | since any peer from the original upload can restore the file, if there |
---|
64 | are 10 peers holding shares, and the happiness threshold is 7, the |
---|
65 | upload should be declared happy, because there is a happy subset of size |
---|
66 | 10, and 10 > 7. However, since a maximum matching on the bipartite graph |
---|
67 | related to this layout has only 3 edges, Tahoe-LAFS declares the upload |
---|
68 | unhealthy. Though it is not unhealthy, a share layout like this example |
---|
69 | is inefficient; for k = 3, and if there are n peers, it corresponds to |
---|
70 | an expansion factor of 10x. Layouts that are declared healthy by the |
---|
71 | bipartite graph matching approach have the property that they correspond |
---|
72 | to uploads that are either already relatively efficient in their |
---|
73 | utilization of space, or can be made to be so by deleting shares; and |
---|
74 | that place all of the shares that they generate, enabling redistribution |
---|
75 | of shares later without having to re-encode the file. Also, it is |
---|
76 | computationally reasonable to compute a maximum matching in a bipartite |
---|
77 | graph, and there are well-studied algorithms to do that. |
---|
78 | |
---|
79 | Issues |
---|
80 | ====== |
---|
81 | |
---|
82 | The uploader is good at detecting unhealthy upload layouts, but it |
---|
83 | doesn't always know how to make an unhealthy upload into a healthy |
---|
84 | upload if it is possible to do so; it attempts to redistribute shares to |
---|
85 | achieve happiness, but only in certain circumstances. The redistribution |
---|
86 | algorithm isn't optimal, either, so even in these cases it will not |
---|
87 | always find a happy layout if one can be arrived at through |
---|
88 | redistribution. We are investigating improvements to address these |
---|
89 | issues. |
---|
90 | |
---|
91 | We don't use servers-of-happiness for mutable files yet; this fix will |
---|
92 | likely come in Tahoe-LAFS version 1.13. |
---|
93 | |
---|
94 | |
---|
95 | ============================ |
---|
96 | Upload Strategy of Happiness |
---|
97 | ============================ |
---|
98 | |
---|
99 | As mentioned above, the uploader is good at detecting instances which |
---|
100 | do not pass the servers-of-happiness test, but the share distribution algorithm |
---|
101 | is not always successful in instances where happiness can be achieved. A new |
---|
102 | placement algorithm designed to pass the servers-of-happiness test, titled |
---|
103 | 'Upload Strategy of Happiness', is meant to fix these instances where the uploader |
---|
104 | is unable to achieve happiness. |
---|
105 | |
---|
106 | Calculating Share Placements |
---|
107 | ============================ |
---|
108 | |
---|
109 | We calculate share placement like so: |
---|
110 | |
---|
111 | 0. Start with an ordered list of servers. Maybe *2N* of them. |
---|
112 | |
---|
113 | 1. Query all servers for existing shares. |
---|
114 | |
---|
115 | 1a. Query remaining space from all servers. Every server that has |
---|
116 | enough free space is considered "readwrite" and every server with too |
---|
117 | little space is "readonly". |
---|
118 | |
---|
119 | 2. Construct a bipartite graph G1 of *readonly* servers to pre-existing |
---|
120 | shares, where an edge exists between an arbitrary readonly server S and an |
---|
121 | arbitrary share T if and only if S currently holds T. |
---|
122 | |
---|
123 | 3. Calculate a maximum matching graph of G1 (a set of S->T edges that has or |
---|
124 | is-tied-for the highest "happiness score"). There is a clever efficient |
---|
125 | algorithm for this, named "Ford-Fulkerson". There may be more than one |
---|
126 | maximum matching for this graph; we choose one of them arbitrarily, but |
---|
127 | prefer earlier servers. Call this particular placement M1. The placement |
---|
128 | maps shares to servers, where each share appears at most once, and each |
---|
129 | server appears at most once. |
---|
130 | |
---|
131 | 4. Construct a bipartite graph G2 of readwrite servers to pre-existing |
---|
132 | shares. Then remove any edge (from G2) that uses a server or a share found |
---|
133 | in M1. Let an edge exist between server S and share T if and only if S |
---|
134 | already holds T. |
---|
135 | |
---|
136 | 5. Calculate a maximum matching graph of G2, call this M2, again preferring |
---|
137 | earlier servers. |
---|
138 | |
---|
139 | 6. Construct a bipartite graph G3 of (only readwrite) servers to |
---|
140 | shares (some shares may already exist on a server). Then remove |
---|
141 | (from G3) any servers and shares used in M1 or M2 (note that we |
---|
142 | retain servers/shares that were in G1/G2 but *not* in the M1/M2 |
---|
143 | subsets) |
---|
144 | |
---|
145 | 7. Calculate a maximum matching graph of G3, call this M3, preferring earlier |
---|
146 | servers. The final placement table is the union of M1+M2+M3. |
---|
147 | |
---|
148 | 8. Renew the shares on their respective servers from M1 and M2. |
---|
149 | |
---|
150 | 9. Upload share T to server S if an edge exists between S and T in M3. |
---|
151 | |
---|
152 | 10. If any placements from step 9 fail, mark the server as read-only. Go back |
---|
153 | to step 2 (since we may discover a server is/has-become read-only, or has |
---|
154 | failed, during step 9). |
---|
155 | |
---|
156 | Rationale (Step 4): when we see pre-existing shares on read-only servers, we |
---|
157 | prefer to rely upon those (rather than the ones on read-write servers), so we |
---|
158 | can maybe use the read-write servers for new shares. If we picked the |
---|
159 | read-write server's share, then we couldn't re-use that server for new ones |
---|
160 | (we only rely upon each server for one share, more or less). |
---|
161 | |
---|
162 | Properties of Upload Strategy of Happiness |
---|
163 | ========================================== |
---|
164 | |
---|
165 | The size of the maximum bipartite matching is bounded by the size of the smaller |
---|
166 | set of vertices. Therefore in a situation where the set of servers is smaller |
---|
167 | than the set of shares, placement is not generated for a subset of shares. In |
---|
168 | this case the remaining shares are distributed as evenly as possible across the |
---|
169 | set of writable servers. |
---|
170 | |
---|
171 | If the servers-of-happiness criteria can be met, the upload strategy of |
---|
172 | happiness guarantees that H shares will be placed on the network. During file |
---|
173 | repair, if the set of servers is larger than N, the algorithm will only attempt |
---|
174 | to spread shares over N distinct servers. For both initial file upload and file |
---|
175 | repair, N should be viewed as the maximum number of distinct servers shares |
---|
176 | can be placed on, and H as the minimum amount. The uploader will fail if |
---|
177 | the number of distinct servers is less than H, and it will never attempt to |
---|
178 | exceed N. |
---|