Changes from Version 1 of ViniVeritas

Author:

trac (IP: 127.0.0.1)

Timestamp:

03/26/08 18:05:12 (2 years ago)

Comment:

--

ViniVeritas

@@ +1 @@
+
+
+Home page: [http://www.vini-veritas.net/ http://www.vini-veritas.net/]
+
+Mike notes:
+
+=== Points about the architecture. ===
+
+   * Links between virtual router nodes in VINI are intended to mirror the
+     physical links they run over; i.e., they are '''not''' intended to be
+     transparent, robust links between nodes.
+
+   * Existing implementations use IP-encapsulated ethernet protocols.
+
+   * Has mechanisms for routing "real" traffic through the VINI network
+     (i.e., to direct traffic into and out of the network).
+
+   * Able to inject network events such as link failures.  However, there
+     are no specific tools mentioned in either paper for doing this.
+
+=== [http://www.cs.princeton.edu/nsg/papers/vini_sigcomm_06/ PL-VINI] or VINI version 1. ===
+
+Runs on largely unmodified PlanetLab kernel and nodes (possibly changes to
+support tun/tap device?)
+
+Each sliver in a PL-VINI slice has two components at user level:
+
+   * the data plane: a click router instance consisting of UDP tunnels
+     (home brew implementation?) to other slivers, a local tap interface
+     to inject packets from the local node, a forwarding (routing) table
+     and a switch interface to UML
+   * the control plane: a UML (User-mode Linux) instance presenting
+     multiple virtual ethernet interfaces to XORP running within.
+
+In theory, can support forwarding of arbitrary packets (not just IP) depending
+on the implementation of the forwarding table in Click and the routing protocol
+in XORP.  In practice, the use has been as an "Internet in a Slice" which is
+IPV4 with node mirroring Internet2 backbone routers.
+
+Packets enter and leave PL-VINI via OpenVPN and NAT.
+
+Clients that "opt in" to PL-VINI use OpenVPN.  A client running on an arbitrary
+internet machine opens a VPN tunnel to an ingress/egress point for PL-VINI
+(nodes running an OpenVPN server in addition to UML and Click).  The server
+feeds packets into PL-VINI via the local tap interface.
+
+For talking to hosts that have not "opted in", PL-VINI runs NAT within the
+Click instance at certain egress points.  These points serve as proxies that
+not only allow traffic to get to outside servers, but allow the return traffic
+to get back in.
+
+=== Trellis or VINI version 2. ===
+
+Replaces user-level components of PL-VINI with Linux kernel features.
+The function of Click as per-instance-forwarding-table-provider is replaced
+by NetNS in the Linux kernel.
+
+It uses ethernet over [http://www.faqs.org/rfcs/rfc2784.html GRE] for
+inter-node tunnels which are terminated in the kernel.  These endpoints are
+connected through tc traffic shaping to a bridge device (either the standard
+Linux bridge or the custom 2-way bridge "shortbridge") and then to virtual
+ethernet devices which appear inside the vservers.
+
+This optimized implementation can only do IP routing but, they claim, can
+fall back on a PL-VINI style implementation to get non-IP routing.
+
+Can forward minimum-sized packets at about 67% of raw Linux Gb Ethernet speed.
+
+-- Main.MikeHibler - 14 Nov 2007