Improving Robustness of Distributed Services (Re: DDoS attacks)

Mailman List Mirror (Read Only)
1

Hello,

I've been involved in running part of another IRC network and
I've been trying to find reasonable ways to immunize networks
to DoS attacks. The reasoning behind this quest is that if
there is no way to deny the service, then the attacker shouldn't
have a reason to even try.

I think this is relevant content for this mailing list because:

* Most of DoS attacks are against IRC networks. Hence, if we can
  get rid of those, the health of the Internet as a whole should
  improve.
* Experience gathered with this approach should be useful to
  developers and administrators of other distributed services
  and protocols.
  (e.g. usenet news, web cache heirarchies, DNS, BGP, ... YMMV)

Protecting the network can be divided into two parts:

Protecting network links and protecting client servers. There
are a lot of well known ways to protect client servers and
someone else can write about that. But apart from myself, I
haven't observed anyone trying to protect network links.
This I find surprising since it is often the links that the
abusers are specifically trying to attack.

The way I feel major inter server links should be protected
is that they should be moved "out of band."

They could for example be moved into permanent virtual circuits
with a few megabits of guaranteed bandwidth per second. Since this
protocol does not use a fully meshed network, but only an
acyclic subset of it, most of the reserved bandwidth would
be available for use by other services. In essence, using this
dedication wouldn't use up any more bandwidth in the backbones
than the current way of running links over the public Internet.
It would just make sure that these critical links couldn't
be caused to flap by denying their bandwidth.

If resources exist and are readily available, the inter server
links could maybe even be assigned a real dedicated circuit.

Wrt. the applicability of this approach to other distributed
services and protocols: It was just yesterday that I was talking
with a friend about BGP transit and peering practises (in general)
and one of the things I mentioned was that I would like to see
the practise of separating the BGP transit session on its own
circuit apart from the transit circuit being deployed more widely.

This would not be such a bad idea at exchanges either. In some
cases it might have very apparent uses for purposes other than
DoS resistance as well:

With the emergence of exchanges for multiple different "Internet
Commodities" (IPv4, IPv4 MBone, IPv6, ...) many of the exchanges
have a separate shared medium (often just a VLAN) for each of
the exchanged protocols.

I don't find debating about the reasons of separating them or not
at this stage of hardware, software and protocol maturity useful.
I'm just stating that this is one of the current practises.

Now, what operators - that are capable of carrying all the
protocols over a single infrastructure - could do at these
exchanges is that they could run peering their sessions for all
the protocols in one and work similar magic with the next-hop
addresses etc. as in the transit session case above. The
exchange point could provide a separate out of band shared
medium over which the operators could run these peering sessions.
This would simultaneously increase network stability by making
the BGP sessions themselves impervious to attacks.

These approaches of course require more complex configuration,
more thought, more clue when debugging, and generally more
everything, but I find that nothing new.

Kindest regards,

2

[...]

I've seen a few suggestions bounced around about way to protect the
inter server links.

Most of the suggestions brought forward revolve around hiding the server
to server links, to make them difficult targets to attack - running
server-to-server links over IPv6 (which is allegedly now as easy to DoS
as IPv4, unfortunately), running them over IPSec w/ private IP addresses
et al. Unfortunately, finding private peering and such for IRC servers
would be a very costly way to go - and not one that's attractive to many
of the educationals/ISPs/private parties that run for the most part,
non-revenue generating IRC services.

There are a few other ways to protect your client servers semi-simply -
move them into a seperate block - which you can easily stop announcing
globally - or even just announce to only those peers with whom you peer
IRC, or those peers whose customers you allow to use your IRC server.
The latter of these would work well for large IRC networks with many
servers, as it controls exactly which servers people can use.

Unfortunately though, I've still not found an elegant solution to
these problems that doesn't also remove the service, or still rely
on shipping the traffic across your borders.

To a certain extent, contacting your transits works to get attacks
stopped. I've certainly had good experiences with at least Cable
and Wireless, Teleglobe, and Genuity. But even then, the time from
attack starting to getting traffic filtered can be in the region of 30
minutes to an hour, and this only stops the traffic coming through your
transits. Peers are another matter entirely, and since even one or two
IXPs can result in the order of 400-500 peers, there's generally no
possible way to call all of them in a short timeframe.

Question for the list:
Does anyone have good or bad experiences with mailing lists containing
all of your transits, and peers NOC addresses, and using these kinds
of lists to contact / request filtering on mass? How do people
on the relevant NOC lists feel about this kind of situation?

Cheers,
Chris.

3

Hello,

I've been involved in running part of another IRC network and
I've been trying to find reasonable ways to immunize networks
to DoS attacks.

[...]

Quote from Chris Roberts:
} I've seen a few suggestions bounced around about way to protect the
} inter server links.

[...]

} There are a few other ways to protect your client servers semi-simply -
} move them into a seperate block - which you can easily stop announcing
} globally - or even just announce to only those peers with whom you peer
} IRC, or those peers whose customers you allow to use your IRC server.
} The latter of these would work well for large IRC networks with many
} servers, as it controls exactly which servers people can use.

While I find the idea of having the IRC server sit in a prefix of its
own with limited visibility very lucrative, it would seem to be very
hard to actually acquire an unaggregated prefix for this purpose.

Would there happen to be someone on the list with an unused B-class
(for example,) that would be willing to give out sub-C's (for example)
for such a purpose? :wink:

} Unfortunately though, I've still not found an elegant solution to
} these problems that doesn't also remove the service, or still rely
} on shipping the traffic across your borders.

Well ... I of course feel the solution I presented would be elegant
wrt that, but I won't deny that provisioning the (virtual or real)
circuits required is very hard, especially if on a zero budjet.

} Question for the list:
} Does anyone have good or bad experiences with mailing lists containing
} all of your transits, and peers NOC addresses, and using these kinds
} of lists to contact / request filtering on mass? How do people
} on the relevant NOC lists feel about this kind of situation?

I haven't done exactly that, but from my experience I would imagine
that using something like that may get your filter installed faster,
but removed slower if at all. It will probably have very little
impact on the number of peers that are willing to install a filter
just for you.

Also: asking for a null host route will more often result in getting the
desired effect than asking for a filter, for various technical reasons.

Kindest regards,