4 The following is a list of issues with GNUnet 0.11.0 that will need
5 to be addressed before we might consider GNUnet deployable to larger
6 audiences. Please keep this in mind when trying out GNUnet 0.11.0!
8 GNUnet may however work fine for applications that aren't impeded by
9 these known deficiencies. Help would of course be welcome to reduce
10 this list, so an estimate of how much work is needed and the main
11 impact are given with each item.
15 * We currently select one transport per peer, but we should allow the
16 use of multiple channels concurrently (i.e. UDP + TCP at the same
17 time). Neither ATS nor transport support this today, and this requires
18 a major change in the ATS plugins and the ATS API and the overall
19 ATS/transport logic. [6-12 PM, robustness, performance]
21 * The current proportional ATS heuristic does not devalue increasing
22 the number of connections once we have reached saturation point,
23 possibly resulting in an excessive number of connections.
26 * The mlp/ril ATS heuristics are highly unstable and keep crashing or
27 worse, and have horrific code quality. [3-12 PM, stability,
31 transport: [12-48 PM overall]
32 * We should move plugins into separate processes to break an overly
33 complicated subsystem into more manageable bits. [maintainability]
35 * Transport plugins are currently expected to be bi-directional; the
36 API should be changed to make them uni-directional, so we could have
37 say UDP for sending but receive via say SMTP, simply because NAT
38 punching and other transports (especially SAT) simply are not
39 bi-directional. Fragmentation and ACKs should then not be done in UDP
40 plugin but at transport level. This should result in significantly:
41 - better NAT traversal
42 - faster transports (especially with the multi-transport of ATS)
43 - simplified transport plugins
44 [stability, robustness, performance]
46 * Transport currently does not encrypt. This has the disadvantage that
47 TCP/UDP traffic is easily identified as "GNUnet"-traffic. It would be
48 better to_also use a simple cipher (ECDHE+AEAD) in a plugin-specific
49 way (i.e. HTTPS is fine already) to minimize information leakage, even
50 if for efficiency that cipher is is not replay-protected at this level
51 (and leave true replay-protected OTR to CORE). [privacy, censorship-resistance]
53 * transport's 'manipulation' functions should be moved into
54 a plugin-proxy, simplifying the code. This may have
55 modest implications for testbed due to the API change.
56 [maintainability, security]
58 * testcases are plenty but insufficiently systematic, the changes
59 described above should also enable us to create more systematic tests.
60 [maintainability, correctness]
64 * The current code may leak LAN IPs (in particular IPv6 with
65 MAC) globally. We have started to put in some provisions to tag
66 addresses as loopback/LAN/WAN, but need to systematically ensure
67 that addresses are only propagated in a useful scope and avoid
68 leaking "sensitive" address data globally.
73 * core needs to be able to communicate to other peers whether this
74 peer is high-bandwidth or on battery and thus either great for
75 relaying or really bad for relaying. Higher-level subsystems
76 could then bias their peer selection to more capable peers --
77 this is key for going on battery-operated systems.
80 * we currently use timestamps (based on roughly-synchronized clocks),
81 challenge-response AND sequence numbers to protect against
82 replay-protection. This is overkill, and the use of timestamps
83 causes issues on OpenWRT where RTCs are often unavailable. We
84 might want to increase the nonce for challenges and get rid of
85 the roughly-synchronized clocks assumption, at least for CORE
86 (we can't avoid it for NSE). [1-3 PM, usability]
88 * once transport encrypts (ECDH+AES), we should simplify CORE level
89 encryption to use ECDH+Twofish(+SHA512) instead of the current
90 ECDH+AES+Twofish double-encryption.
92 * Mobile peers currently allow adversaries to track their users as the
93 peer's identity does not change with locations. We need a mechanism to
94 notify a mobile peer about a location change and then systematically
95 change the public key we use depending on our location. [6-12 PM,
98 * Core should support protocol versioning for higher-level services
103 * needs more systematic testing, ideally with a mock up of DHT
104 and core. [6-12 PM, correctness]
106 * flow and congestion control implementation remains incomplete;
107 should borrow more from net2o [3-6 PM, performance]
109 * Various optimizations, such as key material pre-computation
110 and avoiding unnecessary DHT queries should be implemented.
111 [3-9 PM, performance]
113 * Once transport/core use AES/Twofish for encryption, we should
114 switch CADET to use KECCAK-AEAD.
116 * Generally needs more benchmarking to identify performance
117 bottlenecks tune accordingly. [3-12 PM, performance]
121 * many known NAT traversal methods are not yet supported
122 [6-12 PM, connectivity, usability]
126 * dv is still not working, but currently based on SET and we
127 should simplify the design to not use SET. [3-6 PM, connectivity,
132 * Event loop instantiations for various event loops (like glib,
133 libev, libevent, etc.) should be created (and tested). This
134 allows applications to be built with GNUnet using those various
135 styles of event loop handling. General support for abstracting
136 the event loop is new in 0.11.0, we just need to make broader use
137 of it. [1-3 PM, performance, usability]
139 * gnunet-service-resolver should use asynchronous DNS queries if
140 available. This will ensure that name resolution is not super-slow
141 when GNUnet makes various parallel DNS queries (rare, but DNS
142 queries can happen to resolve peer IP addresses for visualization).
143 [1 PM, performance, usability]
147 * various minor tweaks should be explored, but first we need to
148 define and find a good benchmark to see where our performance
149 bottlenecks really are today; datacache (IO subsystem) might
150 be one of them. Here, picking appropriate transaction scopes
151 and isolation levels should be an easy way to boost performance,
152 but we also need more benchmarking logic. [3-12 PM, performance]
156 * Current implementation is not yet Byzantine fault-tolerant
157 against stuffing attack. [1-3 PM, fault-tolerance]
161 * quite a few different functions are squished together in one
162 subsystem here, will require some changes to address changes
163 to transport and hello discussed above [1-3 PM, correctness]
165 * there are more robust known ways for maintaining a random mesh, we
166 might want to explore using those [3-6 PM, robustness]
170 * The service should be split into two, one for user-specific and one
171 for system-specific components [3-6 PM, security, usability]
173 * Currently, data and files are stored in a disassociated way,
174 resulting in search results that then fail to download; need OR/RPS
175 (see below) before we can think about fixing this properly. [6-18 PM,
178 * datastore (IO subsystem) is bad at keeping the database within the
179 quota set by the user. [1-3 PM, correctness]
183 * dropping 'silence' on large-buffers during replay should be
184 used to avoid accumulating latency (standard VoIP technique,
185 crucial to get it to work for real calls) [1-3 PM, usability]
187 * needs ring tones [1 PM, usability]
191 * should support hijacking arbitrary gtlds [1 PM, usability]
193 * should import ".fr" ccTLD and ensure implementation performs [3-6 PM]
197 * implementation currently "trivial" (aka doing round-robin, not multicast)
200 * still uses old NTR-style API, needs to transition to MQ API [DONE]
204 * implementation under-tested, database performance likely in
205 need for optimization [3-12 PM]
209 * code in need for clean up and more systematic testing [6-12 PM]
213 * Open research, still experimental. [3-24 PM]
217 * Open research, still non-existent. Symmetric cipher should
218 be CHACHA. [48-120 PM]
221 No serious problems are known in:
251 We reserve judgement (due to lack of data) on: