The project is currently in a proof-of-concept prototype state. Though the source code is supposed to be functional, there is a lot of room for improvements.
While prototyping we delayed solving some minor technical issues.
Accordingly, you can find some // TODO
comments in the source
reminding to (re)consider algorithms, error handling and other details.
We already have a test environment that runs a whole network (i.e. lots of nodes) on a single host to prove that the base algorithm and its current implementation works fine in practice, e.g. the nodes discover each other and keep connected. There are some corner cases which we already identified and will work on it.
Tests should also be extended to ensure that the network has high enough graph k-connectivity and after a network split (e.g. internet is turned off in a country then back on) the network is able to merge and reunite.
Most of the network communication uses asynchronous calls already. Connecting to a host is still blocking though, this should be transformed as well. Another potential blocking problem can be the routing implementation for clients (Node::ExploreNetworkNodesByDistance, see the TODO).
All of our algorithm implementations (like discovery, relation renewal, etc) are translated from an algorithmic description in the specification thus currently they are all sequential. Though some algorithms (e.g. discovery) naturally have sequential parts, sequential execution might not scale well for huge networks and we may have to parallelize or use asynchronous execution where possible.
Our current implementation uses a local database to store just a sparse subset of all the nodes of the network, but this is not necessarily the only direction. We could also experiment with using a distributed hash table (DHT). The advantage would be having a single, full node list of the whole network, readable by every node. Each node could write only its own details (e.g. when joining the network or changing its IP address). Note that this direction has shortcomings to solve: it might not work locally if a country blocks external internet access, see e.g. Great Firewall of China. However, it might be possible to create our own solution reusing some parts of DHT protocols and forging a custom solution. We just didn't have enough time and expertise to fully discover this direction.
A reconsideration of the current layering implementation (application: Node, messaging: Dispatcher, session: Session)
also might be necessary because features "connection keepalive" and "Ip autodetection"
do not naturally fit into the picture, see the message loop implementation currently in
ProtoBufDispatchingTcpServer::AsyncAcceptHandler()
.
To make testing network behaviour easier, we should have better diagnostic tools. One possibility would be to implement sending server statistics or log files either on request or periodically to the developers.
To make administration of the Linux version easier, we could separate project files
into separate directories according to conventions of the distribution.
E.g. we could use ~/.config/iop-locnet
and ~/.local/iop-locnet
instead of current ~/.iop-locnet
.
When using config file, options could use format option=value
instead/besides
the current --option value
format. If this is needed, we might have to use
a different command line parser library, it might not be supported by ezOptionParser.
When starting the very first seed node, there are no other nodes yet,
so it has to identify itself as a seed node. As seed nodes are currently
defined as a list of domain names and host resolution is done automatically
by the network library, we have to manually set option --seednode myip
to help it identifying itself, especially as the host cannot easily identify its own
external IP address. This should be somehow automated.
We should authenticate the nodes of the network and maybe also clients. We should consider if we transfer sensitive information that should be encrypted. TLS might be a good candidate for both tasks.
The main goal of the current implementation is to prove that the concept works. So far the network is quite vulnerable as it does not implement appropriate protection from attacks. If we find no conceptual problems we should identify attack vectors and protect from malicious nodes and clients, e.g.
- Require hashes of public keys as node identifiers.
- Node identitites should be validated while negotiating a colleague or neighbour relationship.
- Check integrity of advertised information. E.g. validate if host is really accessible on the reported external address and address approximately matches to its reported location.
- Ensure that uniqueness of sensitive data (e.g. node ids and probably external addresses) is strictly enforced everywhere and cannot be impersonated by 3rd party.
- Ban misbehaving nodes.
We should improve performance and safety by expiring accepted inactive connections. Implementing this will need async timers of Asio. However, we have to keep in mind that when the localservice interface receives a GetNeighbourhood request, it must not expire but keep the connection alive and send notifications when neighbourhood changes.