Network Architecture (R02) #4 21/10/2013Names, Places and Identity

Jon Crowcroft, http://www.cl.cam.ac.uk/~jac22

http://www.cl.cam.ac.uk/teaching/1213/R02

IP addr v. Id+Loc

IP Addr == Interface + Route Hints. TCP state = 5 tuple

Src+Dst port Src+Dst Addr, IP Proto Can’t change during session

If we move, have to get new addr to be reachable

Need to advertise (DNS) for new people Need to tell old people to reconnect

Or tunnel, or rewrite to maintain TCP Mobile IP has agents to do tunnels Mobile IPv6 can cut the triangle case out

Why not just leave as mobile ip

Don’t like triangles for ipv4 Don’t like tunnel overhead So what about new addr trick

Map/encap service or IPv6 trick (8+8, for example)

Re-write v. map/encap

As all said, re-write has potential security problems, but low overhead/scales in router terms

But map/encap has deployment simplicity, but o/h problems - both for encap and for binding service

-ve security for re-write?

Not clear there really is a security problem

Re-writer == NAT, we trust NATs now! E2E TCP/IP 5 tuple assumes

IPv4 I/f+route is some sort of secure thing Never true!

Correct model is TCP state should be bound to EID, and not care about last hop of path/route at all!

Syn-cookie/nonce to secure state Or TSL/SSL or other

-ve overheads for map/encap

As currently formulated… Fast moving device would cause a lot of

re-binding But why not try to localize this?

Movement geographically often doesn’t change provider or even topology much

Separate geo/topo/provider cases and deal with seperately?

Alternative 1 - just ignore

Today, clients move; servers fixed Move- get new IP via DHCP Break TCP connection HTTP recover

Cross layer optimise recovery RTP/UDP don’t care…

Or use Multipath TCP and just add subpath transparently (make before break, though)

What about both ends move?

In a way, unusual! But if routers are also part of movement,

then very “ad hoc” world - so Make hosts routers Believe their route updates… Use App level recovery, or MPTCP make

before break What about new clients of re-moved

servers?

Alt 2 - change TCP

TCP shares state with routers today in Compressed header case

So why not cache this info When you move, send a “SYN” packet from

new addr with compressed state reset to other end (if it hasn’t moved) And copy to router where we _were_(*)

If it has moved, then the router there Which should have state(*) to forward it

Could generalise for all bi-dir protocols (most transport protocols have roughly symmetric packet counts)

DNS

DNS update with TTL 0 is not that big a deal! Even the whole DNS Update rate on one

large site isn’t that big a deal www.tjd.phlegethon.org/words/

thesis.pdf Experimental results (see

Naming for Networking byAtkinson&Bhatti

http://www.cs.st-andrews.ac.uk/~saleem/publications.html http://portal.acm.org/citation.cfm?id=1298105

DNS Update rate

Locality? In london, 10M people move over 1

hour in commute 10^7/60*60 <10000 updates per

second This is trivial to run a transaction

(secure DynDNS) for on a single machine…

New topic: Scaling == Complexity?

When we ask if an architecture, system or protocol scales, what do we mean?

Computer Science defines complexity In terms of incremental cost of

algorithm in terms of input scale - e.g. Dijkstra is O(n^2) cpu in number of routers Link state is O(E) msgs in number of edges A FIB might be O(ln(n)) memory re: routers

Other types of complexity?

Yes - emergent properties Synchronisation effects

Routng update-resonance Phase shifts

Most long flow or most short (tcp congestion control regimes)

Different operating regimes Most web data cacheable, verus most dynamic

Interactions - Scanning worm versus routing updates Epidemic, Pandemic, no spread Susceptibility, Infectious, Recover, Mortality?

Other?

Complex versus Complicated?

Some stuff is complicated E.g. network configuration (CLI/IOS)

Important, but not really amenable to much CS

But could undermine safety C.f. BGP misconfigs locally disrupt

global system. Other eg.??

Old Name Spaces 1

Hosts.txt Update → send new entry to FTP.ISI.ARPA Distribution → FTP to * (Broadcast) How is this different from *.com? The Main Name Service

Note risks of root → Verisign accidental deletion of entire database

Risk worse if TTL for Rrs set to 0 !

Old Name Spaces 2

X.500 Directory Tree Structure (like DNS) Distributed and Delegated Content = (Attrib, Value)* Can wildcard lookup attributes ==

SEARCH! Update has secure per record locks

Old Name Spaces 3

DNS Hackery... Rotaries

What is returned on each call varies Load balancer Hide mobility or failures No interface for policy (backdoor

hack) Requires low(0?) to stop cacheing

interfereing with LB

Old Name Spaces 4

DynDNS Allow rapid update Controlled more at leaves of system Used by bad guys (e.g. botnets IRC

controllers find bots via Dyn DNS) Secure DNS

Update (not just edit named boot file and kill -restart:) Better be secure

Needs key distribution architecture Better not have cyclic dependency

between key and dyndns/update :-(

INS + CCN

Binding Design, Code, Compile, Link, Load/Run,

Call What, Where and When...

Semantics of thing returned Dynamicity?

Use? Esp. Update rate/churn in map

INS looks like evolution CCN looks like revolution

CCN does a lot of other things too...

INS limits

Self org – why not use CAN, Chord etc?

Looks like directory too Search efficiency? Churn etc/ambiguity See X.500/LDAP experiences...

See also Appletalk :-)

CCN – What are people lookin for?

Many URLs (dynamic web page) name content

So make this first class thing Most content is already replicated

“finding” replica is a DNS hack e.g. LB or Akamaized URLs

=> So make this hack part of architecture Fix limits

(content not just@source or fully replicated (akamai/youtube limited model)

See also Multicast and Pub/Sub But promote “cache” and “router” to be same

thing

CCN problems?

Integrity (pollution proof) HH Need to have hash/MIC on set of blocks And not break if “bad” block inserted

Privacy of Subscription HC Need to match interest to datum Equiv. To tag matching in encrypted domain Is x=y, when X and Y both crypted

Computationally expensive...

Some new results on homomorphic cryptography promising