[e2e] determining ingress interface?
aminil at us.ibm.com
Wed Jul 3 09:55:47 PDT 2002
> > would recommend against assumptions of either symmetric paths
> > or bgp reflecting actual traffic flow
> > unless you're writing science fiction
> always wondered about what new career i could launch into! :-)
> seriously, what's a good reference for your second point about bgp not
> reflecting actual traffic flow?
We have done some investigation of how well paths gleaned from BGP tables
reflect paths actually taken from packets. The paper is not yet published,
but you can get a copy at:
(Issues with Inferring Internet Topological Attributes)
Basically, the data was collected using Looking Glass sites, which provide
an HTTP interface to invoke traceroute commands to specified IP addresses,
and to query the site's local BGP router for the AS_PATH associated with an
IP address. We found 92 widely distributed sites in which both the BGP
facility and the traceroute facility were active. We periodically selected
pair of sites (both the period, and the pair were randomly) and requested a
traceroute from each of the pair to the other, and queried the AS_PATH from
each to the other. So for each interval we collected a forward and reverse
traceroute between a pair of nodes, and a forward and reverse BGP AS_PATH
between the same pair. We also did some comparisons to data retrieved from
the Oregon Routeviews website.
The problem (obviously) is that neither the traceroute, nor the BGP AS_PATH
be absolutely relied upon as reflecting the path a packet would actually
Some of the traceroute issues (paper discusses these in more detail) are:
resolving the AS number from the RIR's isn't necessarily accurate, and
implementations of ICMP message generation. Some of the BGP AS_PATH issues
the information is partial (filtering, route selection, aggregation), and
not all routing info is propagated into BGP (static routes, source routes,
multi-hop bgp sessions, AS stuffing).
According to the data we collected, about 33% of all measured traceroute
paths indicated a different AS path length than the corresponding BGP
(although, most were not vastly different; more on this in paper). The
information collected exhibited >50% AS path asymmetry.
More information about the end2end-interest