An AI eternity service (instead of "sovereign" AI)

Being digitally colonized has been a serious threat to national sovereignty, but also to individual freedoms from survellance and censorship for decades. This applied to the Internet, the WWW, the Cloud and now AI. 

Whether the digital Emperors are based in the USA or China, they are there.

To avoid these sorts of risk for content, Ross Anderson proposed the eternity service which finesses the problems in (typical for Ross) an ingeneous way by structuring the infrastructure as a mix of sharing, striping, and sharding and builds in the threat of mutually assured destruction - if you are a low level engineer/computer scientist, the idea is like CDMA or Network Coding or what some colleagues re-purposed to be spread spectrum computing.

A simpler idea is more coarse grained - organisations that provide critical infrastructure (railways, power grids, water&sewerage, Internet etc) can source technology from (say) three different providers. The London Internet Neutral Exchange (LINX), which extends this to cooperative ownership as well. So undermining of one supplier's gear has a limit to damage on the service - indeed, many services operate with a headroom for coping with simple natural disasters in any case (internet and power grids also to allow for wide variance in demand/supply) so this is a natural way to do things.

Another digital version is the Certificate Transparency, which also creates a merkle tree space for (horrible word) coopetition (cooperation amongst competitors), enforced by the tamper evident (or to some extent, socio-economically tamper proof) service space, in a way a single application version of eternity.

This would apply to sourcing data, training, and models themselves + inferencing after the training.

So how about, using the state of the art multi-AI protocols to connect agents, we construct a multi-national AI substrate that serves no-one in particular, but everyone in general. Any attack (removal of an agency, pollution of data) would damage the attacker as much as everyone else. It is in everyones interest to keep the system running and running honestly.

So how to combine neural networks? (something that would also be useful during training or inference so as to share GPU or other accelerator h/w)? You'd need some sort of way to interpret multiple interleaved graphs with multiple (XORd or turbocoded) weights. This is research. Margin's too small to put it here =>

RSPAI?

 

Towards an RSPI for AI (UK) - alpha name RSPAI (short name "pai") 

The rspi was at one point going to be the BBC Nano (or model n) but ended up as rspi because ... my reaction to the 100-200$ cost of the one laptop per child project, and to the limitations thereof...was to "blow a raspberry"...

An AI, like a computer, is a general purpose machine-a purely physical tool analogy is a swiss army knife, but software on computers is tools and die - tool and die makers design, and build or repair tools. they have a bench with lathes and saws and drills and hammers and screwdrivers and boxes full of bits...

A s/w toolbench is something like unix, with subsytems (network stack, file systems, i/o in general (serial, display etc) plus SDK for development (vi, cc, add etc) and then some handy pre-made tools (regex/grep, sort, awk, sed, etc) with source and documents available so people could use them as design patterns (templates etc)

the RSPI prospered as it had 30 year pre-history  (BBC Micro, Acorn, ARM, Broadcom system on chip) and ran  Linux (descendent of bell labs et al) and had,  on the system on chip a GPU (so it could use openGL rendering / gaming s/w) and  wifi, and gpio (so it could connect to sensors and actuators and do robotics or similar

An RSPAI would also need to scale up - by being modular, and networks  (c.f. wifi and gpio above - nowadays MCP or AI2AI or similar) and  federated learning tools/platforms- the equivalent of s/w development environment but with examples (e.g. start with some huggingface pieces) and flower.ai

A network, small and large...with a way to train (NN to FL)- in fact networking at on chip, in software, and between chips/systems...

It also needs some data (kind of AI equivalent to sensor input) - this doesn't have to be _on_ the RSPAI - it needs to be where you can get it (equivalent was that 

the Pi part of the Rasperry Pi name came from the original idea that like the BBC micro that booted into basic, the RSPI would boot into python. we discarded that early on and said kids must learn the Command Line (shell) but one simple example lesson (how to write Snakes in Python) would start with how to download and install python, and then... ... ...

It could run on an RSPI easily (esp. given cheapo GPU) but what is the core tool bench for the Ai part of this RSPAI? is it just huggingface with pytorch tensorflow and all that gubbins, or is there a core that could be still general, but simpler to start off with and afford people with easy lessons - and here is a candidate lean language model that runs on a laptop, as a starter, fresh out of the Turing Institute!

(one class we ran with the RSPI had 11yr old schoolgirls withotu any teacher go from nothing to writing snakes in 1 hour from scratch) - what is the equiv to that, that lets you still also go all the way to writing a version of asteroids and control a bipedal lego robot?  or genai including stable difussion video liks this:- https://youtu.be/kG8fmSW_5wM?si=hsbBdeCUtdshNyBM

A small neural net, a regression/stats library, some causal inference graph stuff? what what what?

And who is it for? policy makers, wannabe AGI gods, defense contractors, health and environment researchers, Jane Public who who who?

Answers on a postcard please...

Principles of Communications Week 9 to Dec 3rd/L16

Today we finished with systems design patterns, and a course summary.

For revision, maybe take a look at example supervision questions and past exam papers

Principles of Communications Week 8 to Nov 28th/L15

This week we covered the multiple time scales of traffic engineering and signaling, from packet time/scheduling to end-to-end/RTT latency admission control/call setup or else congestion control/queue management, to network wide provisioning and the interaction with pricing/utility, and consequences for possible accountability of end user/identity/payment systems...

Then we looked at a toolkit of system design patterns, with particular attention to communications systems, but with broad applicability in many domains/sectors/walks of life...

Principles of Communications Week 7 to Nov 21st/L13

This week we're covering firstly data centers (through the lens of q-jump and its clever "fan-in" trick), and secondly, optimisation of routes and rates. Both of these topics relate to practice, and theory of quite a few machine learning systems/platforms.

The mix of application platforms we see in the data center lecture are classic mapreduce (as represented by Apache Hadoop) and stream processing (as represented by Microsoft's Naiad)[2], as well as memcached (an in memory file store cache, widely used to reduce storage access latency and increase file throughput - e.g. for aforesaid platforms), as well as PTP (a more precise clock sysnch system than the widely used Network Time Protocol[1]). For an example of use of map reduce, consider Google's original pagerank in hadoop. For  the interested reader, this paper on timely data flow is perhaps useful background.

Optimisation is a very large topic in itself,  and underpins many of the ideas in machine learning when it comes to training - ideas like stochastic gradient descent (SGD) are seen in how assign traffic flows to routes, here. In contrast, the decentralised, implicit optimisation that a collection of TCP or "TCP friendly" flows use is more akin to federated learning, which is another whole topic in itself.

Why a log function? maybe see bernouilli on risk

Why proportional fairness? from social choice theory!

Are people prepared to pay more for more bandwidth? One famous Index Experiment says yes.

0. See Computer Systems Modeling for why the delay grows quickly as load approaches capacity.

1. see IB Distributed Systems for clock synch

2. see last year's Cloud Computing (II) module for a bit more about data centers&platforms.

Principles of Communications Week 6 to Nov 14th/L11

 This week we covered congestion control (with a brief mention of latest ideas of CUBIC and BBR) and scheduling (work conserving, max min fairness, and round robin with various tweaks, as well as queue management, with another appearance of "randomness" as a useful trick).

My analogy for the overall system that is TCP + IP is that it is like a bicycle, with a chain connecting two sets of gears - you can change gear at each end to change rate/speed, but the wheels slip (forwarding) and the chain is made of soggy bread (end to end rate).

IP forwarding+scheduling is like the bit where the rubber hits the road, and the foot on the pedal is like the sender, while the force backwards from friction, skidding and effort is the flow & congestion control.

Now loosely couple the chains of 10 billion bikes and riders and roads, and run them anything from 1 to a billion RPM...

and make the road surface out of jelly. and take a chain link out, or put on in every now and then (re-routing)...

and many of the bikes are penny farthings but some are jetskis (i.e. not everyone is running the same algorithm, either at IP level or end-to-end/TCP level).

Principles of Communications Week 5 to Nov 7th/L9

We covered mobile and telephony routing - key interesting idea is Dynamic Alternative Routing, a.k.a. sticky random routing, where the Gibbens paper covers the Max Flow analysis, plus this paper gives the statistical background on the triangle problem.

We then covered (essentially, revision of last years networks) flow and congestion control. There's an important link to control theory here which is covered in the Computer Systems Modeling module next term.

Principles of Communications Week 4 to Oct 31st/L7

This week, we'll finish up BGP, and cover multicast routing.

General lessons from BGP - information hiding can be harmful to decentralised algorithms. but information hiding may be a necessary dimension to some distributed systems due to business cases (commercial inconfidence/competitive data). 

Are there other ways to retain decentralised or federated operations but to retain also confidentiality? Perhaps using secure multiparty computations (not covered in this course!). There are a number of other federated services emerging in the world (applications like Mastodon and Matrix, and also, federated Machine Learning systems like flower) so this probably needs a good solution.

Multicast has a great future behind it - some neat thinking, but largely replaced by application layer content distribution networks (e.g. netflix, youtube, apple/microsoft software update distribution), and the move away from simultaneous mass consumption of video/audio. For more info on limitations of multicast, this paper is a good (quick) read. One of the key objections to multicast is that, like broadcast, it has great power (reach) but also great potential for harm - as a tool for launching denial of service attacks on many people at once, it could be rather handy, for bad people....source specific multicast was one way to limit this risk (so that ISPs could police those sources) - indeed, it would be fun if one could only have source-specific unicast so one would only have to receive packets from designated senders!

Finally, multicast IP has the same packet lost characteristics as normal IP, so you need to put an end-to-end reliable protocol on top of it for applications like software distribution (though perhaps not for video/audio etc). - this isn't as simple as using TCP (or QUIC) as the acknowledgement mecahsnisms would not work well for large groups (each data packet to a group of N recipients would cause a flood of N acks which woiuld likely overwhealm the sender or the senders network access). So you need new multicast transport protocols that manage reliability differently. This is another challenge for using IP multicast - although in some limited deployments (e.g. data centers) it might be a good fit for some applications.

In the analog world (e.g. radio, or audio) broadcast is obviously both good and bad - spark gap generators jam all the radios within considerable range, and white noise generators mess with everyone's hearing nearby...

Are application layer overlays a solution to other "in network" alleged enhancements? Possibly - for example, Cloudflare obviate the need tor always on IPv4 or IPv6 addresses - this also wasn't covered in this course, but might be of interest. Thir work is described here and other papers of theirs may be of current interest.

Principles of Communications Week 3 to Oct 24th/L5

This week, we'll start on BGP/Policy Routing, getting up to traffic engineering, including the amusingly named "hot and cold potato". Next week we will wrap up BGP covering semantics & performance.

A great overview of BGP if you want an alternate source.

Principles of Communications Week 2, to Oct 17th, L3

This week we started on more fundamentals of routing, which can be

centralised - we gave an example, using SDN controllers to add FIBs into a routing domain - in "fibbing" this is done in a hybrid distributed/centralised way, so that the default is that a distributed routing algorithm (typically link state routing) runs, but is then enhanced by central control. At one extreme, everything could be done through central control and dirct control of routers via OpenFlow, and robustness achieved by replicated the controllers.

decentralised - one decentralise approach is source routing, where routers are completely dumn with regards paths, but just forward packets based in fields in the packet. One example  of this is the use of loose source routes in iPv6 in Segment Routing.

distributed - classic Link State Routing is distributed, and does not distinguish any special node.  Of course, link metrics might be configured centrally to influence path selection (see traffic enginering)

federated. - the internet is a network of networks, where each autonomous system runs its own routing protocol (any of the paradigms here work), but to connect networks together and find paths,  BGP is a classic example of fedeated routing. 

We'll start on BGP details next week.

The Control Plane verus the Data Plane

Routing is part of the control plane, whereas forwarding is part of the data plane - control planes are where the management sits (think like light switches or thermostats) whereas data planes is where the users business lies (power, pressure, data...).

Multi-protocol label switching and Segment Routing are a mixture of control and data. MPLS requires other protocols help to install forwarding state (labels), and segment rouring requires other protocols to tell end systems what segement addresses to put in packets. Those other protocols might be pure management protocols (centralised) or just piggyback on distributed routing (add information to link state advertisements), or use some new protocol to obtain map/internet graph data. They then also need additional action in the data plane (to put label on user's data packet's MPLS header, or to put a segment source route list in a user's data packet's IP header).

So MPLS is not a routing protocol, but rather addresses on the forwarding problem, and interacts with other protocols to install labels - this can be done via label distribution protocols (LDP) or signaling protocols like RSVP (which amount to source routing in some senses). In Segment Routing (SR), which is an IP layer analog of MPLS, we can also choose to distibute segment labels via additional link state update fields, or we can use lists of intermediate router addresses in IPv6 to achieve source routing. 

Note MPLS operates "below" IP in the protocol stack, and hence can support multiple network layers for forwarding (hence its name), while SR operates "within" the IP layer.

The main way we illustrate that neither MPLS nor SR are routnig protocols themselves, is that they ave ad hoc methods to deal with switch/router outages and recovery from outages (or a priori protection).

Principles of Communications Week 1, to Oct 10th, L1

Today sees start of 2023/2024 year and for Part II taking Principles of Communications I'll be noting progress and also adding occasional related reading/ and corrections on this blog.

If you want to revise anything to warm up for the course, I suggest last year's Computer Networks course should be a quick re-read! 

This week we'll just make a start on routing.

Glossary of Terms: from ISOC. Some acronyms come from the 7 layer model of the communications stack including terms like PHY (short for physical, so not really an acronym).

For an influential source on how a lot of the way we really "traverse" the Internet today, Cloudflare's head of research writes s very good blog 

For a very nice talk about how Internet Access technology works, this description of the evolution of last mile access is very up to date and watchable.

Principles of Communications Final Week to Nov 28th/L16

 Wrap up - you should now be able to see the supervisor pages which summarise the course and have links to questions and releavnt/in scope past exam questions.

Principles of Communications Week 8 to Nov 23rd/L15

 This week is two Big Pictures 

On Nov 21, we do:

1/ Traffic Management over multiple time scales - linking together several ideas between routing/congestion control, open and closed loop flow control, admission, pricing, and, finally, a bit about signaling.

feedback/admission work at any time scale

short> reduce rate/ lose packets

medium> try again later

long>change to a better provider

So note peak rate pricing is just a slow version of congestion control + shadow prices.

Note using a single network may amortize costs well, but a single technology has single failure mode - the model of copper/fiber/radio has a lot of fault tolerance, which is high value for critical infrastructure.

Note that the problem with capacity planning and traffic matrix/source behaviour in the general internet is that a) it isnt a regular topology, b) each ISP is doing it in competition but also in cooperation with other ISPs! However, this paper showed how well the ISPs coped when everyone switched to working/learning from him the internet in the age of covid

On Nov 23rd we'll cover:

2/ System Design Rules - these are rules of thumb that apply in many systems, processsor design, operating systems, networks, etc etc

More can and will be said...

Principles of Communications Week 7 to Nov 17th/L13

 This week we're covering firstly data centers (through the lens of q-jump and its clever "fan-in" trick), and secondly, optimisation of routes and rates.

The mix of application platforms we see in the data center lecture are classic mapreduce (as represented by Apache Hadoop) and stream processing (as represented by Microsoft's Naiad)[2], as well as memcached (an in memory file store cache, widely used to reduce storage access latency and increase file throughput - e.g. for aforesaid platforms), as well as PTP (a more precise clock sysnch system than the widely used Network Time Protocol[1]). For an example of use of map reduce, consider Google's original pagerank in hadoop. For  the interested reader, this paper on timely data flow is perhaps useful background.

Optimisation is a very large topic in itself,  and underpins many of the ideas in machine learning when it comes to training - ideas like stochastic gradient descent (SGD) are seen in how assign traffic flows to routes, here. In contrast, the decentralised, implicit optimisation that a collection of TCP or "TCP friendly" flows use is more akin to federated learning, which is another whole topic in itself.

Why a log function? maybe see bernouilli on risk

Why proportional fairness? from social choice theory!

Are people prepared to pay more for more bandwidth? One famous Index Experiment says yes.

1. see IB Distributed Systems for clock synch

2. see last year's Cloud Computing (II) module for a bit more about data centers&platforms.

Principles of Communications Week 6 to Nov 10th/L11

 This week, we cover two closely related topics:-

1. Flow&Congestion Control

There are three regimes  for where there may need to be a matching process between sender and recipient rates.

a) sender - the sender may not always have data to send - think about sourcing data from a camera or microphone with a given coding, you get some peak rate anyhow.  or you just might have a short data transfer that never "picks up speed"

b) network - here some type of feedback is needed to tell senders about network conditions - this could be a call setup (open loop control) - not something we do in today's internet, or it could be closed loop feedback via dupacks, packet loss or ECN, and a rate adjustment in the sender (congestion window or explicit rate control.

c) receiver - the receiver can slow down the sender by advertising a small receive windor or just delaying acks or both.

Much of today's network uses these three approachs via TCP or QUIC. There's a nice study based on wide area measrements that shows how congestion control (case b above) is quite rare and more often we have a or c...see this paper:-

TCP equation&reality

2. Scheduling&Queue Management

Schedulers in switches and routers try to minimise the "damage" caused by one flow to other packet flows through a network device that has more packets arriving than it can send just now. Buffering packets turns into jitter, delay, which for audio or video turns into playout buffer delay, or packet loss, which for TCP data traffic is a key congestion signal, and for audio/video is quality degradation.

Schedulers try to allocate bottleneck capacity fairly e.g. via round robin, but how long do queues really need to be? as the network scales up, perhaps the impact of queueing decreases and fairly simple schedulers may suffice - see this paper/talk  for more ideas in this space. Of course, actual overload leading to packet loss is still bad, but the actual delay and jitter seen by a large aggregation of flows through a switch/router may not be the significant worry...

law of large numbers and what happens in queues then really

Principles of Communications Week 5 to Nov 3rd/L9

This week were covering Mobile Networking(*rant*) (briefly) and then telephony networks(^motives for innovation!) - architecture, but in particular, random routing as it is a neat illustration of the use of non-deterministic approaches to distributed systems which yields efficient results See also Valiant Load Balancing (for example this work at Microsoft, also used in Operating Systems (NGINX, for example. 

 For a general discussion of random choices and their application, see this study (very much deep background reading only!) 

 *rant* The TCP checksum is designed to protect end-to-end packets against corruption - it is fairly low cost to. compute (ones-compliment sum of all the 16 bit fields in the TCP header and application data) compared to more complex integrity checks used in lower level protocols (link level checks are usually designed to be implemented in hardware). 

However, it features one incredibly poor additional design choice, which is that the check also includes adding in fields from the "pseudo header" which includes parts of the IP header (source & destination addresses, protocol field and IP segment length). (see for details) This has two problems: firstly it messes with the fate sharing design of the Internet so that end system comms is now entangled with network layer. Secondly, it stops a TCP connection surviving IP address re-assignment, which is essentially what happens if you move network (e.g. switch from using your wifi to using your cellular data link). 

The argument made by the TCP designers was that this prevents some security threat from someone messing with your IP address. It is not clear that that threat actually really represents a real problem since nowadays, TCP implements multipath with multiple IP addresses and switches or combines them safely. Also other protocols (e.g. TLS, above TCP, or else QUIC, as a replacement for TCP+TLS) make use of a secure association between end systems, which is far more secure and resilient to mobility anyhow. So another mistake in the Internet design.

^Motivation for Innovation! The first automatic exchange was designed by an undertaker, Almon Strwoger who believed human operators were diverting business calls to his competitors - real story is a bit less interesting, but still important - precursor to the cross-bar switch!

Started on Flow Control - for closed loop/feedback control systems, this paper by James CLerk Maxwell, back in 1868 covered some nice maths behind how to desgin controllers!

Principles of Communications Week 4 to Oct 26th/L7

 This week, we'll finish up BGP, and cover multicast routing.

General lessons from BGP - information hiding can be harmful to decentralised algorithms. but information hiding may be a necessary dimension to some distributed systems due to business cases (commercial inconfidence/competitive data). 

Are there other ways to retain decentralised or federated operations but to retain also confidentiality? Perhaps using secure multiparty computations (not covered in this course!). There are a number of other federated services emerging in the world (applications like Mastodon and Matrix, and also, federated Machine Learning systems like flower) so this probably needs a good solution.

Multicast has a great future behind it - some neat thinking, but largely replaced by application layer content distribution networks (e.g. netflix, youtube, apple/microsoft software update distribution), and the move away from simultaneous mass consumption of video/audio. For more info on limitations of multicast, this paper is a good (quick) read.

Are application layer overlays a solution to other "in network" alleged enhancements? Possibly - for example, Cloudflare obviate the need tor always on IPv4 or IPv6 addresses - this also wasn't covered in this course, but might be of interest. Thir work is described here and other papers of theirs may be of current interest.

Tup A previously unavailable route is announced as avail- able. This represents a route repair.

Tdown A previously available route is withdrawn. This represents a route failure.

Tshort An active route with a long ASPath is implicitly re-

placed
with a new route p ossessing a shorter ASPath. This represents b oth a route repair and failover.

Tlong An active route with a short ASPath is implicitly replaced with a new route p ossessing a longer ASPath. This represents b oth a route failure and failover. 

Couple of questions on Graphs in last section on BGP 

1. A Few Bad Apples - graph doesn't asymptote, despite being cummulative as it prefixes can be announced more than once (i.e. > 100%!)

2 How long does BGP take to adapt to changes - the legend refers to

Tup A previously unavailable route is announced as available. This represents a route repair.

Tdown A previously available route is withdrawn. This represents a route failure.

Tshort An active route with a long ASPath is implicitly placed

with a new route p ossessing a shorter ASPath. This represents b oth a route repair and failover.

Tlong An active route with a short ASPath is implicitly replaced with a new route p ossessing a longer ASPath. This represents b oth a route failure and failover.

Tup A previously unavailable route is announced as avail- able. This represents a route repair.

Tdown A previously available route is withdrawn. This represents a route failure.

Tshort An active route with a long ASPath is implicitly re-

placed
with a new route p ossessing a shorter ASPath. This represents b oth a route repair and failover.

Tlong An active route with a short ASPath is implicitly replaced with a new route p ossessing a longer ASPath. This represents b oth a route failure and failover.

Principles of Communications Week 3 to Oct 19th/L5

Wrapping up with centralised routing, MPLS and Segment Routing - if you are interested in latter topic, see this survey of SR

This week, we'll start on BGP/Policy Routing, getting up to traffic engineering, including the amusingly named "hot and cold potato". Next week we will wrap up BGP covering semantics & performance.

A great overview of BGP if you want an alternate source.

Some General High level questions about material together with rough dates when each topic will appear in lectures.

Principles of Communications Week 2 13/10/2023

 This week we covered "centralised" routing - two aspects:

Fibbing - hybrid of SDN/central controller and distributed computation using link-state

(note on terminology. )

If you are interested in how to replicate state machines, have a look at this work on distributed consensus . Noting that these replication schemes themselves are very subtle and difficult to get right.

MPLS - and segment routing - which is somewhat stateful forwarding, which therefore requires some input from some controller or management plane to do anything other than default paths (if MPLS or SR just use the IGP to setup labels/segements, you just get whatever the IGP does, so kind of making the use of MLPS or SR rather pointless).

Some useful backup material on segmenet routing, operationally from Juniper Networks

Principles of Communications Week 1 5/10/2023

 

Today sees start of 2023/2024 year and for Part II taking Principles of Communications I'll be noting progress and also adding occasional related reading/ and corrections on this blog.

If you want to revise anything to warm up for the course, I suggest last year's Computer Networks course should be a quick re-read! A fun review of 40+ years of the Internet

This week we'll just make a start on routing.

For fun, you might find this discussion on why LLMs aren't much use for networking, mostly interesting - video recording of panel session

Reference requested for Glossary of Terms: from ISOC

Some acronyms come from the 7 layer model of the communications stack including terms like PHY (short for physical, so not really an acronym).