1 Clean up the khashmir actions.
3 The khashmir actions are a mess, and some cleanup is necessary. A lot
4 of the actions have most of their processing in common, so this code
5 should be put in functions that all can call. Perhaps creating a
6 base "RecurringAction" and "StaticAction" would be a good idea,
7 as then find_node and find_value could use the first, while get_value
8 and store_value could use the second. Perhaps ping and join actions
9 should also be created for consistency, and maybe inherit from a
10 "SingleNodeAction" base class.
13 Packages.diff files need to be considered.
15 The Packages.diff/Index files contain hashes of Packages.diff/rred.gz
16 files, which themselves contain diffs to the Packages files previously
17 downloaded. Apt will request these files for the testing/unstable
18 distributions. They need to either be ignored, or dealt with properly by
19 adding them to the tracking done by the AptPackages module.
22 PeerManager needs to download large files from multiple peers.
24 The PeerManager currently chooses a peer at random from the list of
25 possible peers, and downloads the entire file from there. This needs to
26 change if both a) the file is large (more than 512 KB), and b) there are
27 multiple peers with the file. The PeerManager should then break up the
28 large file into multiple pieces of size < 512 KB, and then send requests
29 to multiple peers for these pieces.
31 This can cause a problem with hash checking the returned data, as hashes
32 for the pieces are not known. Any file that fails a hash check should be
33 downloaded again, with each piece being downloaded from different peers
34 than it was previously. The peers are shifted by 1, so that if a peers
35 previously downloaded piece i, it now downloads piece i+1, and the first
36 piece is downloaded by the previous downloader of the last piece, or
37 preferably a previously unused peer. As each piece is downloaded the
38 running hash of the file should be checked to determine the place at
39 which the file differs from the previous download.
41 If the hash check then passes, then the peer who originally provided the
42 bad piece can be assessed blame for the error. Otherwise, the peer who
43 originally provided the piece is probably at fault, since he is now
44 providing a later piece. This doesn't work if the differing piece is the
45 first piece, in which case it is downloaded from a 3rd peer, with
46 consensus revealing the misbehaving peer.
49 Store and share torrent-like strings for large files.
51 In addition to storing the file download location (which would still be
52 used for small files), a bencoded dictionary containing the peer's
53 hashes of the individual pieces could be stored for the larger files
54 (20% of all the files are larger than 512 KB). This dictionary would
55 have the normal piece size, the hash length, and a string containing the
56 piece hashes of length <hash length>*<#pieces>. These piece hashes could
57 be compared ahead of time to determine which peers have the same piece
58 hashes (they all should), and then used during the download to verify
59 the downloaded pieces.
61 For very large files (5 or more pieces), the torrent strings are too
62 long to store in the DHT and retrieve (a single UDP packet should be
63 less than 1472 bytes to avoid fragmentation). Instead, the peers should
64 store the torrent-like string for large files separately, and only
65 contain a reference to it in their stored value for the hash of the
66 file. The reference would be a hash of the bencoded dictionary. If the
67 torrent-like string is short enough to store in the DHT (i.e. less than
68 1472 bytes, or about 70 pieces for the SHA1 hash), then a
69 lookup of that hash in the DHT would give the torrent-like string.
70 Otherwise, a request to the peer for the hash (just like files are
71 downloaded), should return the bencoded torrent-like string.
74 PeerManager needs to track peers' properties.
76 The PeerManager needs to keep track of the observed properties of seen
77 peers, to help determine a selection criteria for choosing peers to
78 download from. Each property will give a value from 0 to 1. The relevant
81 - hash errors in last day (1 = 0, 0 = 3+)
82 - recent download speed (1 = fastest, 0 = 0)
83 - lag time from request to download (1 = 0, 0 = 15s+)
84 - number of pending requests (1 = 0, 0 = max (10))
85 - whether a connection is open (1 = yes, 0.9 = no)
87 These should be combined (multiplied) to provide a sort order for peers
88 available to download from, which can then be used to assign new
89 downloads to peers. Pieces should be downloaded from the best peers
90 first (i.e. piece 0 from the absolute best peer).