-
Notifications
You must be signed in to change notification settings - Fork 42
Roadmap
This page centralizes the list of projects being worked on, ideas for future projects, list of implemented projects, and projects and ideas that have failed due to some unknown reason at the time they were conceived.
The intent of this roadmap is to guide our community to better allocate our efforts. In addition, this roadmap should give newcomers an opportunity to understand where Linux XIA is, how our efforts are organized, and what you can do to help.
If you are looking for something to work on, and find this roadmap too rough, check out our Ideas list.
Depend on: reaching enough code maturity
Developer: Michel Machado
Status: started
Linux XIA is not yet mature enough for production, but we do think that broader exposure will help us to move forward faster, and get more in tune with the interested community. This project will figure out how we can submit our code upstream, do all required work, and, hopefully, go mainline.
Depends on: no dependencies
Developer: Michel Machado
Status: Not started
This principal will highlight Linux XIA as a layerless network stack, and enable a cleaner implementation of NWP in userland.
Ether principal should follow the example of Linux's "neighbour" (British spelling) subsystem, which links network devices to network protocols, and cache link-layer headers. In order to cache L2 headers, HID principal must support MAC-address changes like the one shown in this example:
ip link set eth0 lladrr 01:02:03:04:05:06
Depends on: TCP-like transport
Developer: Alexander Breen
Status: started
Bringing iperf, nginx, and Firefox to XIA allows one to test and demo a lot of things.
Depends on: UDP- and TCP-like transports
Developer: Yuguang Li
Status: started
VLC and Live555 are client and server video applications. Their port to XIA will enhance our demos, and allow one to explore principals that can improve video transmissions. The importance of this effort comes from the large fraction of the total Internet traffic that is nowadays dedicated to video.
The following projects implement good-to-have features that we are not working on right now, but we'll come back to them at an opportune time. If you choose to work on one of them, please announce it on our mailing list.
Depends on: LPM principal
Status: Not started
Providing intra-domain routing is fundamental to allow users to build testbeds comprised of more than a few machines. Retrofitting BGP/OSPF to XIA would provide XIA users with a "simple" way to build large XIA networks. Bird or Quagga are good starting points to avoid developing from scratch.
Status: Not started
Simplify the work that applications have to do to create XIP addresses.
Status: Not started
These principals offer a gradual migration plan from IPv4 to pure XIA networks.
Status: Not started
The idea is to have a name resolution for XIA. A new resource record type is needed for XIP addresses.
Depends on: none
Status: Not started
Finding a way to reuse TCP's code in the Linux kernel such that TCP can be a principal in Linux XIA. This would simplify porting applications, offer a reliable transport for XIA applications, and benefit XIA from all the long-accumulated experience that Linux's implementation of TCP has.
Depends on: Ether principal
Status: Not started
A very simple version of NWP is implemented inside of HID's kernel module. Once Ether principal is implemented, NWP could be implemented in userland. A userland implementation of NWP would simplify the code, and lower the effort to bring out the potential of the protocol. For example:
- Identifying and removing unavailable hosts efficiently;
- Low convergence time of the neighbor tables;
- Levering instrinsic security to avoid cache poisoning;
- Taking advantage of Ethernet multicast to reduce broadcast frames;
- Supporting zero configuration in local networks.
The last item would eliminate the need for DHCP, and can be seem is just an extension of what NDP does in IPv6.
Once NWP is in userland, the current NWP in the kernel should be removed, and all communication with the kernel to maintain the neighbor table must go through Netlink.
Currently, HIDs are bound to a machine and thus available on all network interfaces, however, this makes testing forwarding harder, and limits network administrators' control on the topology of their networks. The solution is to scope HIDs to one or more network interfaces in a similar way to IPv6, in which an IPv6 is bound to a machine, but scoped to network interfaces. The new incarnation of NWP should scope HIDs.
Depends on: Userland NWP
Status: Not started
Once NWP is in userland and flesh out, the current HID principal will lose meaning, and should be deprecated.
Instead of dropping HID principal entirely, one should implement a new HID principal that simply redirects to Ether principal in order to keep compatibility with Click XIA.
Depends on: TCP-like transport
Status: Not started
Linux kernel does not include a full RSA implementation because its maintainers believe that all current demands can be properly handled in userland. This page holds some links about this issue. The current implementation only supports signature verification. Without RSA, implementing intrinsic security is daunting.
Depends on: None
Status: Not started
Having Linux XIA on Raspberry Pi would open the door to have Linux XIA on embedded solutions. For example, home and car networks. An idea to showcase this port is to send videos and pictures captured with a Pi camera using the Digital fountain.
Depends on: Fast-path
Status: Not started
Implementing CID principal for Linux XIA.
The following projects explore less understood aspects of XIA. Their outcomes can lead to discoveries and publications, or painful failures. If you dare work on one of them, the general recommendation is as follows:
- Start researching the literature. Google scholar is a really good starting point. Read as many related papers as you can find. Write a "related work" page that summarizes the papers you read.
- If you decide to keep going, announce your project on our mailing list.
- Present the most relevant papers you found in our group meetings. Remember to include the link to your related-work page when you request meeting slots to present the papers. You may get good feedback and more references from our group, and someone may even join your effort.
Make the journey your destination. Research is in general rewarding, but it can also be painful. Do not give up on your project because you may not have the qualifications at beginning, or the challenges may seem daunting, or after the first failures. While persistence is going to be key to overcome the difficulties, keep your eyes open to opportunities that you may find during the project. This last point is well captured by Nobel prize-winning biochemist Albert Szent-Györgyi: "A discovery is said to be an accident meeting a prepared mind."
Depends on: TCP-like transport
Status: Not started
If a congestion control solution could be implemented just above XIP and below all principals, it would automatically make everything congestion control aware. There are plenty of challenges here, TCP/IP offers no example, and there is little literature about this. TCP/IP, for example, does not enforce congestion control, which is left to protocol as a measure of good citizenship. Moreover, any congestion control protocol should be friendly to TCP's congestion control in order to avoid disrupting most of the Internet traffic.
Status: Done
The application will take advantage of zFilter principal to multicast videos, and erasure codes to make the transmission reliable. Read about the Digital fountain.
Depends on: none
Status: Done
Here is the list of implemented items necessary for the foundation of XIP:
- XIP routing tables. It was implemented with a sophisticated lock mechanism that supports concurrent writes, lockless reads, and on-demand expansion.
- Support for RTNetlink API. RTNetlink API allows one to write routing daemons in userland, and applications to manage the XIP stack. Notice that RTNetlink API is the same API daemons like Quagga (http://www.quagga.net/) use to manage IP's routing tables.
- Applications to manage configuration (package xiaconf). Package xiaconf is a analogue of iproute2 for XIA, that is, xiaconf is a collection of utilities for controlling an XIA stack.
- Principals as independent kernel modules. Not only does it make developing new principals easier, it also bounds the complexity of the whole code since all principals are independent.
- Support for LXC (http://lxc.sourceforge.net/). LXC is relevant for two reasons, it simplifies sophisticate testing/debugging, and is a cutting-edge virtualization mechanism supported in Linux that makes it important to push XIA for cloud providers/users. Although LXC is not trivial to understand, it turned out that supporting LXC took little coding effort and helped to organize the code.
- Forwarding API for principals with support for fast path (Linux DST subsystem) and positive dependency. Positive dependency is important to avoid flushing routing cache whenever an modification in the routing table is made; otherwise the performance could be even worse than slow path, and would produce traffic hiccups in busy routers.
Depends on: Foundation
Status: Done
HID principal integrates with XIP's routing table, application xip, and packet forwarding. HID also implements a very simple version of Neighbor Watch Protocol (NWP) that fills in a role similar to ARP's one in IPv4.
Depends on: Foundation
Status: Done
AD principal integrates with XIP's routing table, application xip, and packet forwarding.
Depends on: XIPNIC
Status: Done
XDP implements a UDP-like protocol, but it's header less. XDP doesn't add checksum, and since XIP doesn't support fragmentation, neither does XDP.
Depends on: Foundation and principals AD, HID, and XDP
Status: Done
XIP DST entries must be invalidated when they rely on the non-existence of either a principal (i.e. the principal isn't loaded into the kernel), or an XID of a known principal; a principal becomes known when it's loaded in the kernel. Negative dependency is essential in XIP because routing cache will end up making wrong decisions without it.
Depends on: UDP-like transport
Status: Done
The idea here is to have connection state, reliable delivery, flow control, and optional ordered delivery. We ported Serval to XIA; Serval offers all these features and more.
Status: Done
These principals allow Linux XIA to interoperate with IPv4 networks, and Click XIA's variation of U4ID.
Status: Done
It is multicast principal that does not keep state per multicast group, but router's neighbors.
XIP addresses are not going to be associated to network devices like IP addresses are. XIP addresses do not have the same semantic of an IP address. Even XIDs are not assigned to interfaces, for example, an HID is associated to a machine, not an interface.
This new protocol must support dynamic extensions in order to allow unforseen principals to use it for unforseen messages.
It turned out that to support evolution, each principal must implement its own messages when need, and that a principal dedicated to diagnosis is the way to fill the remaining gaps of not having an ICMP-like protocol.
Depends on: Foundation
Status: Not started
Raw sockets was thought as a way of first testing ideas for principals in userland. However, Implementing principals in userland is not on the roadmap anymore because it turned out that each principal makes use of different kernel APIs to work, and restricting userland principals to a set of kernel APIs would limit userland principals' scope. Moreover, Linux prototype has successfully been able to componentize principals as kernel modules.
All grants that have generously supported the development of Linux XIA are listed on our Funding page.