1 G : We should focus on clearing the TODO List, then go to testing phase,
2 since the end is nearing... ( I finish on august, 3 )
5 Catch a more precise exception thant Exception at line 108 in vifibnet.py
8 Upgrade the logging function in order to be able to log message like
9 "Refreshing peers DB ... done", or add log messages to specify that an
10 action advertised by a previous log message has been completed
12 use the server as a bootstrap node -> switch peer discovery to be done
13 by vifibnet directly ?
15 Use an algorithm to choose which connections to keep and/or establish
16 instead of pure randomness
18 Find a name for the project : the projet is ( or should be ) independant
19 from vifib, openvpn, babel, and so should the name; btw we aim to build a
20 distributed, scalable, resilient VPN.... ( if it helps with the name )
22 Replace comments at the beginning of functions with docstrings & give all
25 In peers DB, flag the dead peers so we only choose them if necessary and we
26 can remove them if we have enought peers
28 Use a timeout for the server peersDB so we can flag unreachable peers and
29 remove the peers whose certificate is no longer valid
31 Specify a lease duration in ForwardViaUPnP
33 Handle LAN internally in order not to have catastrophic results ....
34 ( avahi could be used )
37 G, J : To get traffic stats ( bytes in/out ), you can use
38 /sys/class/net/interface/statistics/rx_bytes, etc...
39 or /proc/net/dev/snmp6/interface ( all in one file ). This can be enough
40 if used as follows: get traffic diff from last time we checked in order
41 to choose which connection is significantly unused compared to others,
42 and close it. Of course, too recent connections (i.e. those for which we
43 have no previous stat) would be always kept.
44 This should be combined with routing table (i.e. how many nodes are
45 served by each tunnel), which is possibly redundant.
46 ip6tables should be avoided if possible.
49 U : Babel seems to be very long to establish the routes : maybe we should
50 tell him thant we are not on a wired network but on a mobile network ?
51 G : babel establish routes quickly enough i'd say. There are two new
52 options : hello and wireless, for hello_interval and treating all
53 interfaces as wireless. However, treating an interface as wireless
54 doesn't lessen the hello_interval, it only changes how babel estimates
55 quality link, and cost.
56 U : from babel web page : "When the Babel daemon detects a wired network,
57 it will use a larger interval between hellos".
58 Moreover, it seems that the wireless option only means
59 "hostile environment" which seems best for a resilient network.
60 30 sec of hello interval seams also too much. The default value for
61 babel is 4 sec (from babel man page).
62 According to raphael's stats on the nexedi's server downtime,
63 45% of the problems dont last more than 2 minutes, 55% no more than
64 3 minutes If it takes 2 min to detect a dead connection, then we wont be
65 solving many problems with our overlay network
66 G : ok, so babel hello-interval should be set to a lower value,
67 we should do some tests to pinpoint the best compromise between
68 speed and bandwith usage.
69 Btw, is there a doc ( pdf, image, file ) resuming Raphael's stats
70 on nexedi's server downtime ? it could be useful for the internship
73 U : The peer DB size should depend on the number of connection and the
75 G : ?! I don't agree, the db size should be proportional ( with a factor
76 like 0.01 or less ) to the total number of peers in the entire network,
77 with maybe a max size.
78 U : what we need to do is to keep the randomness. For this, we need a big
79 enought DB to ensure we can still choose a peer as if it was choosen
80 directly from the server. The requiered db size can be calculated from
81 the number of connections and the refresh time.
82 G : ok, you can erase this talk
84 U : Why are --ip and internal-port mutually exclusive ?
85 Currently upnp only forward via UDP. Should he also forward via TCP ?
86 Why dont we only use UDP ?
87 No error should be raised when no upnp is detected : we should allow
88 machines having public IP to do an automatic configuration using the
89 discovery by an other peer
90 G : Actually, i was wrong, --ip and internal-port are no longer exclusive
91 Julien said udp might not be used by some people because of
92 restrictions imposed by the ISP ( FAI in french ), so we should
93 allow both, and act according to the options specifying which servers
94 to start (upd, tcp-server)
96 G : I think the number of route going through an interface should be a
97 Connection attribute, not a dict in tunnelManager