Date post: | 20-Aug-2015 |
Category: |
Technology |
Upload: | martin-geddes |
View: | 7,600 times |
Download: | 1 times |
Martin Geddes www.martingeddes.com [email protected]
This presentation was given as a keynote at the Illinois Institute of Technology Real Time Communications
conference in Chicago on 11th September 2012
© 2012 Martin Geddes Consulting Ltd
Cosmologies
This presentation is about belief systems – ‘cosmologies’ – and our most basic assumptions about the
nature of the (networking) universe.
Do our models explain observed reality? If not, the model has scope to be improved.
This is the motion of Mars across the sky, as seen on
successive nights.
Reality vs Model
The ancient Greeks struggled to explain this retrograde motion.
Indeed, the word ‘planet’ means ‘wanderer’.
Geocentric
Their basic assumption was that the Earth is the centre of the universe.
This is an anthropocentric viewpoint.
Everything revolved around the Earth, because that’s where we
are, and what could be more important than us?
Gods!
Geocentric
So the explanation for the motion of planets was to appeal to
unknown and unknowable forces.
Copernicus
Except that didn’t really account for what we observed. There
was a residual motion that wasn’t explained by epicycles. So the basic assumptions had to be
re-visited.
A mere one and a bit millennia later, Copernicus suggests that
maybe we (and the Earth) aren’t the centre of the universe after
all.
#fail
Heliocentric Geocentric
His heresy was to say that the Sun was the centre of the solar
system, and everything went around that.
Which is a shift in the origin of how we reasoned about
planetary motion, and required us to let go of the
assumption the whole universe revolves around
us.
You Know Who
And there were still unexplained residual differences between
observed reality and our models. So relativity fills
that gap.
“Obese” “Lean”
Networking Cosmologies
I am proposing that we need to shift our ‘origin’ and adopt a better cosmology in order to make progress in networking.
The existing dominant
cosmology…
A suggested improved
cosmology…
Agenda
Lean
Manufacturing
Lean Conversations
“Lean Networking” Implications
WHAT &
HOW
WHY (“Obese Pipes”)
Internet, WebRTC,
policy
Future of Distributed
Network Computing
The rest of this presentation
involves comparing a lot of ‘obese’ and
‘lean’ cosmologies.
Theory of
Constraints
This book introduced the basic principles that we are
interested in, and was a ‘cosmological shift’ in
process manufacturing.
Supply of raw
materials
Demand for finished
goods
Bottleneck
Every production system has stages of processing raw inputs into valued
outputs.
There is always a bottleneck somewhere
in the system.
Different Worldviews
Cost Accounting
Utilisation (Cost)
Batch
Bad
Cost per Unit
Throughput Accounting
Flow (Throughput)
Single-piece
Good
Lead Time
Focus on
Work unit
Variation
Key metric
Batches make sure there’s no idle time. Variation just drives
up unit cost.
Variation in quantity and quality of
demand is ordinary.
A resource standing idle is a major waste.
Aha! Fallacy
WRONG!
Goldratt’s great insight was to
identify this fallacy.
Journey to ‘Lean Anything’
Flow efficiency
Re
sou
rce
eff
icie
ncy
LOW HIGH
LOW
HIGH
Are the machines
busy?
Is stuff produced fast
enough?
The natural tendency of managers seeing idle men and machinery is to make
it get busy.
But you can’t get flow efficiency this way. Coping
with variation requires slack, and you don’t have
any.
“A plant in which everyone is working all the time
is very inefficient.”
Eliyahu M. Goldratt
Busy machinery isn’t the same as creating value, which is from products delivered to customers
on time.
Inventory
(WIP)
Keeping machinery busy upstream of a bottleneck just
creates work-in-progress (WIP), which ties up capital,
requires storage, and inhibits flexibility.
Journey to ‘Lean Anything’
Flow efficiency
Re
sou
rce
eff
icie
ncy
LOW HIGH
LOW
HIGH So to get to production nirvana, you need to master flow efficiency (i.e. low lead time) first,
then figure out resource efficiency.
1. Visualise your workflow
2. Limit work-in-progress (WIP)
3. ERR… THAT’S IT!
Key features of methodology
Goldratt gave us a methodology, which I
have grossly oversimplified to these
two steps.
1. Visualise your workflow
2. Limit work-in-progress (WIP)
3. ERR… THAT’S IT!
Key features of methodology
Source: agilemanagement.net
Cumulative Flow Diagram
There are flow visualisation techniques
which emphasise the metrics we care about:
lead time and WIP.
1. Visualise your workflow
2. Limit work-in-progress (WIP)
3. ERR… THAT’S IT!
Key features of methodology
Regulate
Inflow Limit WIP by preventing more inflow into the
system than the bottleneck can process.
LEAN CONVERSATIONS WARM-UP EXAMPLE
Before we start on networking, let’s
practise these techniques. How could we use
lean principles to escape from email
hell?
Bottleneck
24
You really don’t want to spend your life doing
email. And even if you do, there’s a finite amount of time resource to dedicate
to the cause.
Journey to ‘Lean Conversations’
Flow efficiency
Re
sou
rce
eff
icie
ncy
LOW HIGH
LOW
HIGH
So how can we lower lead time and WIP to
create flow efficiency in our email handling?
1. Visualise your workflow
2. Limit work-in-progress (WIP)
3. ERR… THAT’S IT!
Key features of methodology
My inbox
Empty… (except new messages since I got off the plane)
I don’t keep my emails staring at me from my
inbox, as if it was a to-do list anyone can scribble
on.
Kanban visualisation system
I use a product to visualise and prioritise
my work – and at a glace see what needs to be
done next.
1. Visualise your workflow
2. Limit work-in-progress (WIP)
3. ERR… THAT’S IT!
Key features of methodology
Limited WIP
I also force myself to focus on a
maximum of two activities at once.
My attention is the bottleneck.
My personal workflow
I limit WIP to force me to recognise my finite work capacity.
“Jane’s inbox has reached its limit of 30 unread items. To resubmit your message at a later time, click here.”
“You have sent 3 messages to John. He has a limit of 2 messages outstanding per person.
Which message do you want to delete?”
Holiday hell: “You have 319 new messages.” – Just send them all back and ask people to re-submit!
End ‘inboxbloat’ by limiting WIP
Three kinds of network
Network of Promises
Network of Possibilities
Network of
PROBABILITIES
Manage priority, control loops
Manage flow, contention
and trade loss/delay
Manage sessions, admission
and capacity
NON-DETERMINISTIC STOCHASTIC DETERMINISTIC
Internet world
TDM circuit world, IMS world
Bounded flow isolation
Assured loss and delay
Unconditionally stable
Scavenger traffic
UK demo ISP running since 2007
The benefits of connections and connectionless… without needing circuits to reserve capacity or congestion control loops!
My colleagues have built something that appears
impossible in the existing cosmology.
A network element standing idle is a major waste.
Aha! Fallacy
WRONG!
How? Same as Goldratt – they saw the fallacy
common to all networking.
Journey to ‘Lean Networking’
Flow efficiency
Re
sou
rce
eff
icie
ncy
LOW HIGH
LOW
HIGH
Flow efficiency in networks means
containing loss and delay to meet the
bounds required by the application.
Stop making networks
mindlessly busy creating WIP
1. Visualise your workflow
2. Limit work-in-progress (WIP)
3. ERR… THAT’S IT!
Key features of methodology
Improper Probability
Cumulative Distribution Function
See how loss and delay accrue along the path,
using a probability mass.
1. Visualise your workflow
2. Limit work-in-progress (WIP)
3. ERR… THAT’S IT!
Key features of methodology
We use the word ‘idle’ to mean ‘lazy’ as well as to describe a network link that is not transmitting.
This gives us a clue we have made an
inappropriate moral value judgement
“Immense harm is caused by the belief
that work is virtuous.”
Bertrand Russell
What underpins this is a belief that it is morally
wrong for links to be idle.
Limit WIP!
Instead we should build networks so that they
limit packet ingress to the known downstream
bottleneck, and minimise WIP.
Work-
Conserving
Queues
That means the most basic assumption built into
every network router is wrong – or at least
unhelpful and costly.
Work-conserving queues are ones which always transmit if the link is idle and a
packet is waiting.
Networks don’t do work!
They do translocation.
Category Error
The clue seemed to be in the name –
‘net-work’ – but they aren’t machines that
do work.
This is a big deal.
Putting stuff in the wrong place stops you doing something useful, i.e.
moving packets can create negative value
Work metaphor ! !
A lot of the ‘work’ routers do in fact destroys value.
Bandwidth Quality
Attenuation
Let’s give ours a name too.
Let’s give the old cosmology a name – the ‘bandwidth’ model.
An idea like entropy or noise…
…with many mathematical representations
What is ΔQ?
Shannon solved this ‘noise’ stuff for single links; we’re doing it
for complete multiplexed systems.
An ideal network copies data instantly and perfectly
Real networks instead create loss and delay
Zen of Netwrking Alan Turing
assumed this.
And that’s it. All they do is
degrade.
Networks are trading spaces
Zen of Netwrking
This is important.
Pay attention.
They aren’t pipes. They just allocate the
degradation compared to ‘instant and perfect’.
Networks are trading spaces that allocate
{loss-and-delay}
Zen of Netwrking
Which can be traded (e.g. drop
more packets means less delay
for the rest)
Networks are trading spaces that allocate
disappointment
Zen of Netwrking
They allocate this
disappointment (degradation) to different flows.
Networks are trading spaces that allocate
quality attenuation
Zen of Netwrking
Which we give a fancy name
1. ΔQ exists
2. ΔQ is conserved
3. ΔQ has two degrees of freedom (for mutable contention)
The Three Fundamental
Laws of Networking
If you offer a load to a network, loss and delay will happen.
You can’t un-lose or un-delay stuff.
Rather like pressure-volume-temperature you learnt in school for a piston, networks have load-
loss-delay.
These are the ‘thermodynamic’ laws we have to work with and constrain the best
outcome we can deliver.
Bandwidth
Inter-process comms
“Beads on a string”
ΔQ
So we are working with a different mental model that shifts our origin.
Networks aren’t about pushing packets along one behind another. That’s an anthropomorphic model
that wrongly treats packets like physical packages.
Instead, networks just do inter-process communications with
added loss and delay.
Bandwidth
0 ∞ ΔQ
Our origin is ‘zero loss and delay’. That makes the mathematics tractable –
rather like putting the sun at the centre of elliptical orbits and forgetting
epicycles.
In the bandwidth model, a packet is assumed to stay at rest until we do
‘work’ to push it along. The default is it takes an infinite time to get there.
Bandwidth ΔQ
Quality is the presence of something
positive
Quality is the absence of something negative
People talk about ‘quality of service’, but what is this
‘quality’ they talk of?
Our quality is “the absence of excessive disappointment”.
Their quality is “priority over other packets to get work
done”.
Bandwidth
Non-work conserving queues
Work-conserving queues
ΔQ
Doing work that creates negative value is dumb.
Stop doing it!
Two kinds of obesity
Network of Promises
Overprovision! NON-DETERMINISTIC
Overbuild! DETERMINISTIC
Network of Possibilities
The Internet folk want you to wildly overprovision networks
to make quality-demanding applications work. This is
expensive.
The traditional telco folk want to
build lots and lots of parallel
networks to separate out
different quality requirements.
This is expensive. (This doesn’t actually work – but they haven’t
figured it out yet, so let’s keep it a secret between you and me.)
Two kinds of obesity
Network of Possibilities
Overprovision! NON-DETERMINISTIC
Flow efficiency
Re
sou
rce
eff
icie
ncy
LOW HIGH
LOW
HIGH YOU
ARE
HERE
The Internet is stuck here. TCP is designed to saturate networks – i.e. create
resource efficiency. It can’t create flow efficiency.
Offered load
Tran
spo
rte
d lo
ad a
t w
ith
in-b
ou
nd
s q
ual
ity
LOW HIGH
LOW
HIGH
PREDICTABLE CHAOTIC Network
components often behave poorly under
high load
Necessary to manage the load on the network to
remain within the predictable
region of operation
We have lost control over the
customer experience
The Internet is also unstable at
all loads and suffers
congestion collapse effects.
Two kinds of obesity
Network of Promises
Overbuild! DETERMINISTIC
Flow efficiency
Re
sou
rce
eff
icie
ncy
LOW HIGH
LOW
HIGH
YOU
ARE
HERE
Traditional telco folk were
heading the right way with technologies
like ATM.
But they’ve gone totally the wrong way since then with IMS which
creates a horribly bloated and inflexible
network.
Network obesity means
LIFESTYLE CHANGES!
Aspiration Quantify desired outcomes
Expectation Define quality need
Execution Match delivery to requirement
Assurance Closing the loop
Nobody wants to hear the lean
networking doctor tell them that their
morbid obesity demands lifestyle
changes.
e.g. how often is it OK for YouTube to show the “circle of death”? Average time to first
frame? MOS score for VoIP apps?
How do aspirations map to statistical bounds on
loss and delay?
How to make applications fail in the
right order when instantaneous demand
exceeds supply? Did it do what you expected? How do you
know?
INTERNET, WEBRTC & POLICY
SO WHAT?
This is a new technology to put real-time two-way
voice and video into web browsers.
Bufferbloat
The phenomenon that Internet
engineers struggle to explain and resolve
in their current cosmology is called
‘bufferbloat’.
This problem won’t find a resolution
without a change in cosmology; attempts
to fix it will spawn new problems.
The network
Hierarchy of Need
3. Uncontended: Low loss and delay
2. Stable: High stationarity
1. Feasible: High capacity
Note: exact requirements are application-dependent
!
!
The bandwidth school of thought
focuses entirely on capacity. This is not
enough.
Applications need finite bounds on loss and delay to work.
Rapid variation in loss and delay cause
application failure.
We’re not measuring the things that actually matter,
Dropping flow isolation
+ Rising peaks of failure
+ Loss of stationarity
= More ‘extreme weather events’ (and poor for real-time comms)
The Internet’s
‘Global Warming’ problem
The Internet is intrinsically unsuitable for carrying all of society’s real-time
communications needs.
The Internet has an inbuilt ‘statistical noise’ problem…
We’re multiplexing more apps, devices and users
together; and doing more real-time applications as
well as bulk data.
So bufferbloat-like effects cause more problems.
…because TCP breaks basic control theory principles.
Not good.
End-to-end
principle
The guiding ‘end-to-end principle’ on which Internet architecture is based is
a post-rationalisation of bad design decisions from the 1970s.
This has a cost our kids will carry.
Has broken philosophical foundations: Value comes from flows with bounded ΔQ
– Not individual packets
Loss is OK and delay can be good! – It’s not a fault or moral failure to avoid work
System of two degrees of freedom (loss/delay) – Not just one (delay)
Network Neutrality is
‘beads on a string’ model
Network neutrality is intellectual nonsense. But
once most people thought it was obvious the Sun went
round the Earth too.
In man vs mathematics, the mathematics takes a
really hard negotiating stance
Network Neutrality ! !
Network neutrality as currently conceived puts regulators in conflict with the mathematics of statistical multiplexing. The maths will win.
There are rigorous and meaningful ways of defining
non-discrimination.
Just not this way.
Technological Revolutions &
Financial Capital
Carlota Perez
Steam, Coal, Iron, Railways
Electricity, Steel & Heavy
Engineering
Oil, Petrochemicals & Automobiles
IT & Telecoms ? Biotech, Nanotech
1770 2012
Techno-economic revolutions go through this predictable cycle.
The Turning Point
Amazing! It works!
Fit-for-purpose
You are here! c.70 years after transistor
invented.
Networking becomes embedded into
everyday life, invisible in the way electricity is.
We’re done inventing the basic component
technologies – datagrams,
smartphones, etc.
The Turning Point
Networking as alchemy
Networking as chemistry
Networking has been done on a “try it and see” basis, without
rigorous mathematical foundations.
ΔQ model fills the intellectual hole Turing
left behind when he assumed translocation
was instant and perfect.
“Translocatability” now has intellectual
foundations as solid as “computability”.
Obese monoservice networks
Which means we can begin to treat network obesity caused by
excessive consumption of monoservice (single class)
networks.
Build network
Tweak network to get apps working
Get complaints
Build it bigger
Tweak it again
Select new technology
End the Cycle of Failure
Because that really isn’t working any more.
What we need is…
Ubiquitous data transport for a digital society
And will be even less suitable for what we need
in future.
Lean polyservice networks
We need to go from ‘black and white’ to ‘colour’ networks. Each
flow has a different quality (‘colour’) need.
ΔQ RINA
Beyond cloud & internet:
Distributed Computing
ΔQ is only half of the puzzle. The Internet is mis-architected in
other ways. Check out Recursive Internet Architecture for how it
should be done.
www.martingeddes.com
Free newsletter
Congratulations, you have survived this intellectual assault course. For further intellectual assaults, read my newsletter.
www.futureofcomms.com
Future of Voice
& Telco-OTT workshops
London, England October 23rd & 24th
For a more gentle re-shaping of your views on communications,
come to a public workshop.
Twitter: @martingeddes Email: [email protected]
Martin Geddes
Founder, Martin Geddes Consulting Ltd
Get in touch
Additional credits and thanks to: Neil Davies, Peter Thompson, John Day, Fred Goldstein
Pleased to hear from you.