Re: Weird 'net connection problem

Tim Streater <timstreater@xxxxxxxxxxxx> wrote:

real-address-in-sig@xxxxxxxxxxxxxxx (Rowland McDonnell) wrote:

Tim Streater <timstreater@xxxxxxxxxxxx> wrote:

Okay - so what's the benefit in reducing the MTU to `anything in
particular'? Or, for that matter, in any given value of MTU?

A larger MTU will in general improve throughput on a TCP connection
(where you get feedback from the receiver to the transmitter (i.e.,
across the Internet) to throttle or not the traffic).

Okay.

That's why a path
containing two devices that can only communicate with a small (-ish) MTU
will cause problems, cos they'll fragment the traffic down to what they
can mutually handle.

I don't see why it should cause /problems/, though - a slow-down I can
understand, but why problems/

Depends how many packets in your connection are lost. Your or the remote
end's TCP stack may decide its too many and give up.

Righto.

Note also though that Internet routing takes no account of pipe
bandwidth capacity, only on paths.

I /think/ I can work out what that means.

If a circuit goes down then traffic
is re-routed, but if that causes some pipe to be badly overloaded you
will get packets just dropped. It's up to the end-to-end connection to
recover from that.

Well, yes - but that's how it's designed to work, so surely the only
downside from the user's point of view when that happens is that it's
slower?

Up to a point as I say. But it can be *very* much slower.

Ah.

It may be that the backup path for some failure
condition hasn't been configured properly or tested - or that more than
one failure has taken place.

Ah - so it's normal for error recovery to not be implemented to spec,
then?

The end-to-end recovery is handled by your local TCP stack and the
remote one; an IP backbone router only delivers packets on a
best-efforts basis.

Okay.

I believe that (could be wrong) if packets are
dropped because of congestion, that is done silently.

Okay - but the packet loss is detetected because the packets don't
arrive at their destination. So how is this an issue of any sort?

If it's because
the destination became unreachable then I think an ICMP "unreachable"
packet gets sent back to your IP stack which can then inform the user.

Righto.

You can't take account of all failure modes in the wide-area Internet.

But surely dropped packets get detected and re-sent using the standard
mechanism?

Yes, but you can have intermittent failures that cause "route flapping".
Lets say that you have your own /24 network (thats 256 addresses) that
you connect to some ISP via a router (not a home ADSL one, I should add,
but a small IP router from say Cisco costing a grand or so).

What difference does the `nature of the router' make?

Now, your router and the ISP's router will send each other keepalive
packets so each knows that the link is up. Also, the ISP's router will
be advertising to the rest of the Internet, the address of your network,
so all ISPs know how to send packets to your network. So, look in any
ISP's router, and you will see in the routing table an entry for your
network, one of the 250,000 or so routes that big ISP routers have [1].

Righto. I think...

Now suppose you power down your router. After some timeout period (say,
3 missed keepalive packets), the neighbouring router will conclude your
network is offline and will send a "withdraw route" message to the rest
of the Internet, and an "Add route" when you power it on again.

Uhuh.

But
these routing announcements take some while to reach everyone. This
means that the Internet's knowledge of itself is being constantly
updated (mostly automatically), but you can see that it can never have
complete and up to date knowledge of its actual state.

<puzzled> The internet knows nothing. It contains information, some of
it more accurate than others.

Information propagates at finite speed, therefore given a non-zero size
to a system like the internet, it's obviously impossible for any given
location to hold a complete and up to date state record of the entire
shebang.

From what I've read, it's physically impossible to have a 100% up to
date complete and accurate knowledge in any given place of the state of
any physical system - the only way to achieve anything like that (IIRC)
is to duplicate the system, run a parallel model and that model is the
knowledge, sortathing. But when I say `duplicate the system' I mean
down to the quarks and run a model which is an *EXACT* copy. That is, a
duplicate version of `reality'.

(I get Physics World every month - quite often, it gets rather esoteric)

So if your link is flaky and keeps going up and down, your ISP will be
sending these route announcements about your network constantly, so if
traffic is routed towards your network it could easily be blackholed
somewhere. It's unlikely to be possible to recover from that.

Eh? So you're suggesting that that could cause a /permanent/ cut-off
from the internet? Surely shome mishtake?

A typical backbone router might receive several thousand such
announcements per sec.

Uhuh.

To reduce the load due to these "route flaps",
typically such routers are configured to "damp" this traffic. If a given
route flaps more than x times, it's withdrawn altogether until it's been
up for some period of time.

I'm not sure what that means - could you elucidate?

So the router doing the damping won't pass
the "flaps" further along. Tuning these parameters requires the
expertise of an experienced wide area network engineer. Not a job for a
monkey and not something I would try, either.

Umm. Not sure what this means due to confusion above.

And if there's some sort of serious routing problem, then I can expect
to see my link to `Part A65 of the internet'just not working at all?

Yes.

Righto.

The whole point of a lot of IT kit (especially in the 1980s and 1990s)
is/was work creation for more people, rather than automation to save
work.

Well, to be fair, some of the work requires judgement,

So?

but I feel a lot
more of it could be automated.

It could be made more accessible. One of the problems we've got is
fucking awful documentation. US computer firms are very bad for
supplying useful documenation. I don't know why, exactly, but I have my
suspicions.

Still, I'm surprised the Internet works
as well as it does.

A lot of people are trying very hard to keep it working. And a lot of
them are the sort of people who are willing to get out of bed at 3am and
start work right now if need be.

All ready for IPv6, are we? :-)

What would I know?

[1] In practice, the address space for your network would typically have
been allocated from a larger block by your ISP. Your ISP would advertise
the aggregate address block to the whole Internet, and that's what would
be in other ISPs' routers' routing tables.

Okay.

Ta, most interesting,
Rowland.

--
Remove the animal for email address: rowland.mcdonnell@xxxxxxxxxxxxxxx
Sorry - the spam got to me
http://www.mag-uk.org http://www.bmf.co.uk
UK biker? Join MAG and the BMF and stop the Eurocrats banning biking
.

Quantcast