COE 558Lecture 01Reference

Reference sheet

Introduction compressed onto one page: the definitions, formulas and numbers to have in your head before a quiz or exam.

TCP/IP stack

Five layers, bottom-up PL, DLL, NL, TL, AL. Each payload is the PDU of the layer above. Part 03: TCP/IP stack

Layer	PDU	Payload	Protocols	Connection	Address
Application	Message	Application bytes	HTTP, SMTP, IMAP	Process to process, rides on E2E	None
Transport	Segment (TCP) or datagram (UDP)	Message	TCP, UDP	E2E, virtual, one per path	Port number
Network	Packet (datagram)	Segment or UDP datagram	IP, routing protocols	H2H, virtual, one per path	IP address
Data link	Frame	Packet	Ethernet, 802.11, PPP	P2P, physical, one per hop	MAC address
Physical	Bits	Frame, as signals	Signal specs of Ethernet and Wi-Fi	The medium itself	None

Layer, PDU, payload, protocols, connection scope and address

Sockets, TCP and UDP

$\text{socket} = (\text{IP}, \text{port})$ names one endpoint. A connection is the pair of sockets, and a flow is the 5-tuple (src IP, dst IP, protocol, src port, dst port). Ports: 0 to 1023 system, 1024 to 49151 user, 49152 to 65535 dynamic. Part 05: Sockets and web apps

Property	TCP	UDP
POSIX type	SOCK_STREAM	SOCK_DGRAM
Setup	Three-way handshake before data	None, each datagram stands alone
Delivery	Loss detected and retransmitted	No delivery or duplicate guarantee
Ordering	In-order byte stream	May arrive out of order
Boundaries	Continuous stream, no boundaries	One datagram is one message
Header	20 bytes minimum plus connection state	8 bytes, no state
Fits	Web pages, file transfer, APIs	Live voice and video, DNS queries, QUIC

TCP socket versus UDP socket

Decentralized versus distributed

Two independent dimensions, not opposites. Bitcoin is both. Part 04: Network applications

Aspect	Decentralized	Distributed
What is split	The application, into separate client and server programs	The computation and state, across client, server or many machines
Stresses	Structure: programs talking through a protocol and sockets	Resources: where compute and state live
Lecture example	BitTorrent, Bitcoin	Google Cloud
Design-space axis	Vertical: which layer	Horizontal: which compute or storage resource

The lecture's definitions

Edge-to-cloud layers

Going outward, proximity falls and resources rise. Chain order: device, mist, near edge, far edge, fog, regional cloud, cloud. Part 08: Edge, fog and cloud

Layer	Proximity	Resources	Role	Examples
Mist	Closest	Very limited	Immediate, local decisions	Sensor chip, Raspberry Pi Pico
Near edge	Very close	Limited to moderate	Local processing	Home router, Raspberry Pi 4
Far edge	Close	Moderate to high	Low-latency applications	Jetson Orin Nano, AWS Wavelength Zone
Fog or regional	Intermediate	High	Aggregation and coordination	Metro micro data center, Azure Stack Edge
Cloud	Distant	Very high	Large-scale computation and storage	AWS Region and its availability zones

E2C layers

Latency formulas

One-way lower bound from geodesic distance at c ≈ 300,000 km/s: about 3.3 ms per 1,000 km. Fiber runs near 200,000 km/s, so about 1.5× the bound. Part 07: Latency

\text{Latency}_{LB} = \dfrac{d_{\text{geodesic}}}{c},\qquad \text{RTT}_{LB} = \dfrac{2\,d_{\text{geodesic}}}{c}

Lower bound, one way and round trip

d_{\text{prop}} = \dfrac{d}{s}\qquad d_{\text{trans}} = \dfrac{L}{R}

d: link length, s: signal speed, L: bits, R: bits per second

d_{\text{nodal}} = d_{\text{proc}} + d_{\text{queue}} + d_{\text{trans}} + d_{\text{prop}}

Per node, summed over every hop

Response time and budget. Part 09: Mapping tasks, Part 10: Latency budget

RT = L_1 + PT + L_2 \le \tau

Out, processing, back, within the budget τ

RT = L_{c \to e} + PT_1 + L_{e \to f} + PT_2 + L_{f \to e} + PT_3 + L_{e \to c}

Mapping 4: every hop plus every task

The four delays

Delay	Where	Driven by	Formula or factor
Processing	Inside each router, switch or host	Header and error checks, forwarding lookup	Device speed, usually microseconds
Queueing	Router and server buffers	Congestion ahead of your packet	Traffic intensity, zero to very large
Transmission	Network interface pushing bits out	Packet size and bandwidth	L / R
Propagation	The physical channel	Link length and medium	d / s

Where each delay happens

Region	Geodesic	Bound one way / RTT	Measured RTT	Ratio
Frankfurt (eu-central-1)	≈4,390 km	14.6 ms / 29.3 ms	100 ms	≈3.4×
Tokyo (ap-northeast-1)	≈8,300 km	27.7 ms / 55.3 ms	163 ms	≈2.9×
Oregon (us-west-2)	≈11,900 km	39.7 ms / 79.4 ms	243 ms	≈3.1×

From Dammam: measured RTT is about 3× the round-trip bound

The four mappings

Pipeline T1 resize, T2 detect, T3 draw. Part 09: Mapping tasks

Mapping	Where T1, T2, T3 run	Strengths	Weaknesses	Example
1	All in the cloud	Consistency, simple client, easy scaling	Large L1 and L2, unfit for AR or gaming	Photo tagging in a distant data center
2	All on the device	No network, privacy, works offline	Battery, limited hardware, long PT	Smartwatch activity recognition
3	All at the edge	Low latency, data stays local	Limited reach, limited scalability	Retail store analytics
4	T1 edge, T2 fog, T3 edge	Shared model, more compute, model updates	Extra inter-layer hops and transfer cost	Smart-city traffic cameras

Strengths and weaknesses

New York client, California cloud

Worked numbers

Terms: L1 = 70 ms, PT = 30 ms, L2 = 70 ms
Response time: 70 + 30 + 70 = 170 ms
Computation share: 30 / 170 ≈ 17.6%, network 140 / 170 ≈ 82%
Server 2x faster: 70 + 15 + 70 = 155 ms, speedup 170 / 155 ≈ 1.097x (about 9%)
Infinitely fast server: 70 + 0 + 70 = 140 ms, ceiling 170 / 140 ≈ 1.21x
Amdahl check: f = 30/170, k = 2: 1 / (0.8235 + 0.0882) ≈ 1.097

S = \frac{1}{(1 - f) + f/k}

Amdahl: the unimproved network caps the speedup

Design exercise method

Same four steps for smart glasses, warehouse AR and VR. Part 10: Latency budget

Four steps

1. List the tasks: Break the app into its pipeline: capture, compress, detect or recognize, render, store, retrain.
2. Set τ per task: Cite published budgets: 20 ms motion-to-photon for XR (RFC 9699), 370 to 1000 ms face recognition, 150 ms one-way voice (G.114). Storage and training have no deadline.
3. Place each task: Rendering and tracking on the device (mist), heavy inference offloaded to a nearby edge, storage and training in the cloud.
4. Check RT and trade-offs: Write RT = out + PT + back with numbers against τ per task, then weigh battery, heat, privacy, cost and scale.

Worked check at τ = 100 ms: fog 20 + 70 + 20 = 110 ms fails, edge 5 + 70 + 5 = 80 ms passes.
Check budgets at a percentile, not the mean. A faster CPU only shrinks PT.

Slide errata

Answer with the corrected fact, and name the slide version if a question depends on it. Part 01: Course map

What the slides get wrong

Slide 5: Deployment models (public, private, hybrid, multi-cloud) differ from NIST: private, community, public, hybrid.
Slide 18: 100 W x 10 h = 1,000 Wh = 1 kWh, an energy, not 1,000 watts.
Slide 25: Email is an application (SMTP, IMAP), and MAC addresses are an addressing scheme, not protocols.
Slide 26: Point-to-Point means one hop between adjacent interfaces, not PPP or peer-to-peer.
Slide 28: IP plus port names one socket. A connection needs both endpoints plus protocol (the 5-tuple).
Slide 37: The socket is the endpoint. TCP establishes the connection; UDP never does.
Slide 42: TCP data can be lost in the network; TCP detects and retransmits. More than two socket types exist.
Slide 45: "Font-end" means Front-end. server.route is the hapi framework on Node.js.
Slide 48: "Layer 1, 2, 3" are path tiers. Switching is layer 2, routing layer 3.
Slide 49: The 1 / 5 / 15 ns figures are a ranking only; main memory is about 90 ns.
Slide 54: Use geodesic, not planar, distance. Oregon is about 11,900 km, so 39.7 ms, not 37 ms.
Slides 60, 71: The pictured board is a Raspberry Pi 4 single-board computer, not a microcontroller board.
Slide 72: Compare RT, not latency, with τ.
Slide 76: Captions skip the device layer, which must run tracking for a 20 ms XR budget.
Slide 77: RT = f(Latency, PT) means the sum of every hop plus every task's PT.