1 | The "Denver Airport" Protocol |
---|
2 | |
---|
3 | (discussed whilst returning robk to DEN, 12/1/06) |
---|
4 | |
---|
5 | This is a scaling improvement on the "Select Peers" phase of Tahoe2. The |
---|
6 | problem it tries to address is the storage and maintenance of the 1M-long |
---|
7 | peer list, and the relative difficulty of gathering long-term reliability |
---|
8 | information on a useful numbers of those peers. |
---|
9 | |
---|
10 | In DEN, each node maintains a Chord-style set of connections to other nodes: |
---|
11 | log2(N) "finger" connections to distant peers (the first of which is halfway |
---|
12 | across the ring, the second is 1/4 across, then 1/8th, etc). These |
---|
13 | connections need to be kept alive with relatively short timeouts (5s?), so |
---|
14 | any breaks can be rejoined quickly. In addition to the finger connections, |
---|
15 | each node must also remain aware of K "successor" nodes (those which are |
---|
16 | immediately clockwise of the starting point). The node is not required to |
---|
17 | maintain connections to these, but it should remain informed about their |
---|
18 | contact information, so that it can create connections when necessary. We |
---|
19 | probably need a connection open to the immediate successor at all times. |
---|
20 | |
---|
21 | Since inbound connections exist too, each node has something like 2*log2(N) |
---|
22 | plus up to 2*K connections. |
---|
23 | |
---|
24 | Each node keeps history of uptime/availability of the nodes that it remains |
---|
25 | connected to. Each message that is sent to these peers includes an estimate |
---|
26 | of that peer's availability from the point of view of the outside world. The |
---|
27 | receiving node will average these reports together to determine what kind of |
---|
28 | reliability they should announce to anyone they accept leases for. This |
---|
29 | reliability is expressed as a percentage uptime: P=1.0 means the peer is |
---|
30 | available 24/7, P=0.0 means it is almost never reachable. |
---|
31 | |
---|
32 | |
---|
33 | When a node wishes to publish a file, it creates a list of (verifierid, |
---|
34 | sharenum) tuples, and computes a hash of each tuple. These hashes then |
---|
35 | represent starting points for the landlord search: |
---|
36 | |
---|
37 | starting_points = [(sharenum,sha(verifierid + str(sharenum))) |
---|
38 | for sharenum in range(256)] |
---|
39 | |
---|
40 | The node then constructs a reservation message that contains enough |
---|
41 | information for the potential landlord to evaluate the lease, *and* to make a |
---|
42 | connection back to the starting node: |
---|
43 | |
---|
44 | message = [verifierid, sharesize, requestor_furl, starting_points] |
---|
45 | |
---|
46 | The node looks through its list of finger connections and splits this message |
---|
47 | into up to log2(N) smaller messages, each of which contains only the starting |
---|
48 | points that should be sent to that finger connection. Specifically we sent a |
---|
49 | starting_point to a finger A if the nodeid of that finger is <= the |
---|
50 | starting_point and if the next finger B is > starting_point. Each message |
---|
51 | sent out can contain multiple starting_points, each for a different share. |
---|
52 | |
---|
53 | When a finger node receives this message, it performs the same splitting |
---|
54 | algorithm, sending each starting_point to other fingers. Eventually a |
---|
55 | starting_point is received by a node that knows that the starting_point lies |
---|
56 | between itself and its immediate successor. At this point the message |
---|
57 | switches from the "hop" mode (following fingers) to the "search" mode |
---|
58 | (following successors). |
---|
59 | |
---|
60 | While in "search" mode, each node interprets the message as a lease request. |
---|
61 | It checks its storage pool to see if it can accomodate the reservation. If |
---|
62 | so, it uses requestor_furl to contact the originator and announces its |
---|
63 | willingness to host the given sharenum. This message will include the |
---|
64 | reliability measurement derived from the host's counterclockwise neighbors. |
---|
65 | |
---|
66 | If the recipient cannot host the share, it forwards the request on to the |
---|
67 | next successor, which repeats the cycle. Each message has a maximum hop count |
---|
68 | which limits the number of peers which may be searched before giving up. If a |
---|
69 | node sees itself to be the last such hop, it must establish a connection to |
---|
70 | the originator and let them know that this sharenum could not be hosted. |
---|
71 | |
---|
72 | The originator sends out something like 100 or 200 starting points, and |
---|
73 | expects to get back responses (positive or negative) in a reasonable amount |
---|
74 | of time. (perhaps if we receive half of the responses in time T, wait for a |
---|
75 | total of 2T for the remaining ones). If no response is received with the |
---|
76 | timeout, either re-send the requests for those shares (to different fingers) |
---|
77 | or send requests for completely different shares. |
---|
78 | |
---|
79 | Each share represents some fraction of a point "S", such that the points for |
---|
80 | enough shares to reconstruct the whole file total to 1.0 points. I.e., if we |
---|
81 | construct 100 shares such that we need 25 of them to reconstruct the file, |
---|
82 | then each share represents .04 points. |
---|
83 | |
---|
84 | As the positive responses come in, we accumulate two counters: the capacity |
---|
85 | counter (which gets a full S points for each positive response), and the |
---|
86 | reliability counter (which gets S*(reliability-of-host) points). The capacity |
---|
87 | counter is not allowed to go above some limit (like 4x), as determined by |
---|
88 | provisioning. The node keeps adding leases until the reliability counter has |
---|
89 | gone above some other threshold (larger but close to 1.0). |
---|
90 | |
---|
91 | [ at download time, each host will be able to provide the share back with |
---|
92 | probability P times an exponential decay factor related to peer death. Sum |
---|
93 | these probabilities to get the average number of shares that will be |
---|
94 | available. The interesting thing is actually the distribution of these |
---|
95 | probabilities, and what threshold you have to pick to get a sufficiently |
---|
96 | high chance of recovering the file. If there are N identical peers with |
---|
97 | probability P, the number of recovered shares will have a gaussian |
---|
98 | distribution with an average of N*P and a stddev of (??). The PMF of this |
---|
99 | function is an S-curve, with a sharper slope when N is large. The |
---|
100 | probability of recovering the file is the value of this S curve at the |
---|
101 | threshold value (the number of necessary shares). |
---|
102 | |
---|
103 | P is not actually constant across all peers, rather we assume that it has |
---|
104 | its own distribution: maybe gaussian, more likely exponential (power law). |
---|
105 | This changes the shape of the S-curve. Assuming that we can characterize |
---|
106 | the distribution of P with perhaps two parameters (say meanP and stddevP), |
---|
107 | the S-curve is a function of meanP, stddevP, N, and threshold... |
---|
108 | |
---|
109 | To get 99.99% or 99.999% recoverability, we must choose a threshold value |
---|
110 | high enough to accomodate the random variations and uncertainty about the |
---|
111 | real values of P for each of the hosts we've selected. By counting |
---|
112 | reliability points, we are trying to estimate meanP/stddevP, so we know |
---|
113 | which S-curve to look at. The threshold is fixed at 1.0, since that's what |
---|
114 | erasure coding tells us we need to recover the file. The job is then to add |
---|
115 | hosts (increasing N and possibly changing meanP/stddevP) until our |
---|
116 | recoverability probability is as high as we want. |
---|
117 | ] |
---|
118 | |
---|
119 | The originator takes all acceptance messages and adds them in order to the |
---|
120 | list of landlords that will be used to host the file. It stops when it gets |
---|
121 | enough reliability points. Note that it does *not* discriminate against |
---|
122 | unreliable hosts: they are less likely to have been found in the first place, |
---|
123 | so we don't need to discriminate against them a second time. We do, however, |
---|
124 | use the reliability points to acknowledge that sending data to an unreliable |
---|
125 | peer is not as useful as sending it to a reliable one (there is still value |
---|
126 | in doing so, though). The remaining reservation-acceptance messages are |
---|
127 | cancelled and then put aside: if we need to make a second pass, we ask those |
---|
128 | peers first. |
---|
129 | |
---|
130 | Shares are then created and published as in Tahoe2. If we lose a connection |
---|
131 | during the encoding, that share is lost. If we lose enough shares, we might |
---|
132 | want to generate more to make up for them: this is done by using the leftover |
---|
133 | acceptance messages first, then triggering a new Chord search for the |
---|
134 | as-yet-unaccepted sharenums. These new peers will get shares from all |
---|
135 | segments that have not yet been finished, then a second pass will be made to |
---|
136 | catch them up on the earlier segments. |
---|
137 | |
---|
138 | Properties of this approach: |
---|
139 | the total number of peers that each node must know anything about is bounded |
---|
140 | to something like 2*log2(N) + K, probably on the order of 50 to 100 total. |
---|
141 | This is the biggest advantage, since in tahoe2 each node must know at least |
---|
142 | the nodeid of all 1M peers. The maintenance traffic should be much less as a |
---|
143 | result. |
---|
144 | |
---|
145 | each node must maintain open (keep-alived) connections to something like |
---|
146 | 2*log2(N) peers. In tahoe2, this number is 0 (well, probably 1 for the |
---|
147 | introducer). |
---|
148 | |
---|
149 | during upload, each node must actively use 100 connections to a random set |
---|
150 | of peers to push data (just like tahoe2). |
---|
151 | |
---|
152 | The probability that any given share-request gets a response is equal to the |
---|
153 | number of hops it travels through times the chance that a peer dies while |
---|
154 | holding on to the message. This should be pretty small, as the message |
---|
155 | should only be held by a peer for a few seconds (more if their network is |
---|
156 | busy). In tahoe2, each share-request always gets a response, since they are |
---|
157 | made directly to the target. |
---|
158 | |
---|
159 | I visualize the peer-lookup process as the originator creating a |
---|
160 | message-in-a-bottle for each share. Each message says "Dear Sir/Madam, I |
---|
161 | would like to store X bytes of data for file Y (share #Z) on a system close |
---|
162 | to (but not below) nodeid STARTING_POINT. If you find this amenable, please |
---|
163 | contact me at FURL so we can make arrangements.". These messages are then |
---|
164 | bundled together according to their rough destination (STARTING_POINT) and |
---|
165 | sent somewhere in the right direction. |
---|
166 | |
---|
167 | Download happens the same way: lookup messages are disseminated towards the |
---|
168 | STARTING_POINT and then search one successor at a time from there. There are |
---|
169 | two ways that the share might go missing: if the node is now offline (or has |
---|
170 | for some reason lost its shares), or if new nodes have joined since the |
---|
171 | original upload and the search depth (maximum hop count) is too small to |
---|
172 | accomodate the churn. Both result in the same amount of localized traffic. In |
---|
173 | the latter case, a storage node might want to migrate the share closer to the |
---|
174 | starting point, or perhaps just send them a note to remember a pointer for |
---|
175 | the share. |
---|
176 | |
---|
177 | Checking: anyone who wishes to do a filecheck needs to send out a lookup |
---|
178 | message for every potential share. These lookup messages could have a higher |
---|
179 | search depth than usual. It would be useful to know how many peers each |
---|
180 | message went through before being returned: this might be useful to perform |
---|
181 | repair by instructing the old host (which is further from the starting point |
---|
182 | than you'd like) to push their share closer towards the starting point. |
---|