curtis at wawa.ans.net
Fri Sep 2 18:56:47 UTC 1994
> I totally agree with Milo. The only problem that must be solved on day 1 is
> *STABLE* migration away from the single NSP model.
A worthy goal.
[ ... nightmares about transition problems and PSC relies heavily on
their network connectivity ... ]
> I predict that stability in the presence of congestion will prove to be far
> more important than actual pipe size.
We have been testing RFC1323 TCP under single and multiple flow fully
congested conditions. In that respect we can push the test setup
beyond what it needs to initially support. It would be a good idea to
repeat these tests with the proposed NAP equipment before it gets
deployed so we know the limits before we deploy.
We can also inject full Internet routes from multiple places (without
creating a forwarding path to the Internet) and see if the normal
(high) level of route flap affects the proposed NAP routers when
forwarding traffic. In the past this has had surprising effects.
I don't know how we could generate extremely large aggregations of
flows and the diversity of active destinations. This is a known
problem for cache architectures that was recently addressed by Cisco,
but currently requires that hardware assist be disabled. Although we
can't replicate this easily on the testnet we know we need to deploy
this (due to operational experience of at least two providers). The
testing should therefore be done with this fix in place (which will
require disabling some hardware assist in the forwarding path). We do
want to make these tests as realistic as possible.
> P.S. If you did not see my paper "Windowed Ping: An IP Layer performance
> Diagnostic" at INET'94/JENC5, get it from the ISOC www server or
> ftp.psc.edu,pub/net_tools/WPing.ps . This describes our preferred tool for
> testing connectivity to our customers. I'd prefer that it be used in October,
> rather than November.
Running mping tests would be a good idea. A word of caution though.
On the testnet, we've found you need cross traffic of similar average
pacvket size and level to get similar results to the real network with
mping (with no background you get wrong results, but they look great).
While ANS doesn't have any magic, we do have a fair amount of
operational experience with a large IP backbone and are as familiar as
anyone with the requirements it places on the routers. I do hope
Bellcore chooses to take advantage of our offer to help test.
More information about the NANOG