COE 558Lecture 02Reference

Reference sheet

E2C continuum and cloud computing compressed onto one page: the definitions, formulas and numbers to have in your head before a quiz or exam.

The E2C continuum and the fog

The continuum is a range of latencies: moving away from the client adds latency and adds capacity. The cloud stays centralized for economies of scale, so mini data centers (the fog) go out instead. Part 01: Latency continuum and the fog

RTT_{\min} = \dfrac{2d}{v}, \qquad v \approx 2 \times 10^{8}\ \text{m/s}

Physics floor over fiber: 4,000 km gives at least 40 ms

Layer	Span	Services	Priority	RTT floor	Example
Edge	LAN	Compute, storage	Autonomy	~1 µs at 100 m	Obstacle detection
Fog	WAN	Compute, storage, networking	Coordination	~0.5 ms at 50 km	Traffic signals, collision coordination
Cloud	Internet	Full data center	Global connectivity	~40 ms at 4,000 km	Fleet-wide model training

Layer map (slide 9) with RTT floors

Resource	Medium	Very large	Ratio
Network	$95 per Mbit/s per month	$13	7.1
Storage	$2.20 per GB per month	$0.40	5.7
Administration	~140 servers per admin	>1,000	7.1

Economies of scale, 2006 (Armbrust et al.): about 1,000 versus 50,000 servers

Aspect	Fog	Edge
Structure	Hierarchical, multi-layer	A few peripheral devices
Services	Compute, networking, storage, control, acceleration	Specific applications in a fixed logic location
Location	Between end devices and the cloud	The end-device layer itself

Fog versus edge (NIST SP 500-325)

Roles of edge, fog and cloud

Edge serves the client, fog serves many edges, cloud serves everyone. Part 02: Edge, fog and cloud layers

Layer	Role	Scope	Data horizon	Example
Edge	Immediate processing near the client	One client or site	Milliseconds (act and discard)	Wake-word detection
Fog	Intermediate layer for a region	Many edges in a district	Seconds to days	Traffic management
Cloud	Centralized, large-scale processing	Everyone, globally	Months to years	Training an AI model

Roles of the three layers

Fog benefits: less edge-to-cloud traffic, geo-restricted data for compliance, shorter edge-to-edge paths.
Edge roles: WAN boundary, offload target, fast small-data processing, gateway that hides the fog (like a reverse proxy).
Cloud: big compute (HPC), big storage (big data), big networking (multi-cloud).
IoT tree: devices sense, gateways process locally, fog coordinates regionally, cloud analyses globally.
Fog filtering example: 200 × 4 Mbps = 800 Mbps raw becomes 3.2 Mbps of events, about 250× less.

d = v \cdot t

Distance travelled while waiting for a decision

Vehicle platoon

Setup: v = 30 m/s (108 km/h), gap about 10 m
Cloud, 60 ms: d = 30 × 0.06 = 1.8 m
RSU (fog), 5 ms: d = 30 × 0.005 = 0.15 m
Car or V2V, 1 ms: d = 30 × 0.001 = 0.03 m
Mapping: Cars are edge nodes, RSUs and the RSUC are fog, the cloud coordinates globally.

Question	This course	NIST SP 500-325
What edge names	Layer on the WAN boundary	End devices and their users
Where a gateway sits	Edge layer	Among the fog nodes
Use it for	Exam answers in this course	Papers citing NIST SP 500-325

Where the edge boundary sits

Legacy IT and virtualization

A VM is an efficient, isolated duplicate of a real machine. Part 03: Legacy IT and virtualization basics

Legacy IT stack

Bottom to top: Data center, networking, storage, server, virtualization, operating system, databases, security, applications
Legacy IT: The customer manages all nine layers.
McCarthy (1961): Computing sold like a utility: shared capacity, pay per use, network access.

Popek and Goldberg's three properties

Equivalence: Programs see an environment essentially identical to the real machine.
Efficiency: A statistically dominant subset of instructions runs directly on the real CPU. Rules out emulators and simulators.
Resource control: The VMM keeps complete control of real resources; a guest cannot reach what it was not given.

\text{Sensitive} \subseteq \text{Privileged}

Theorem 1: a sufficient condition for a trap-based VMM

Aspect	Virtualization	Emulation	Simulation
Core idea	Logical copies of physical resources	Mimics a hardware/software interface	Models a system's behaviour
Runs natively	Most guest instructions	Nothing, all translated	The model only
Same ISA	Yes	No	Not applicable
Speed	Close to native	Much slower	Whatever the model needs
Example	VMware, VirtualBox	BlueStacks, console emulator	Flight simulator

Virtualization, emulation and simulation

	Full (unmodified guest)	Para (modified guest)
Type 1, bare metal	VMware ESXi, Xen HVM	Xen PV
Type 2, hosted	VirtualBox, VMware Workstation	Paravirtual drivers in hosted products

Two independent axes: placement (rows) by guest modification (columns)

Hypervisors and containers

Cloud providers use Type 1: less overhead is more sellable capacity, and a thinner layer between tenants is stronger isolation. Part 04: Hypervisors and containers

Property	Type 1 (bare metal)	Type 2 (hosted)
Runs on	The hardware, as the lowest layer	A process of a host OS
Overhead (slide)	≈5%	≈10%
Resource management	Direct	Indirect, through the host OS
Administration	Needs more skill	Easy, like a desktop app
Isolation	Stronger	Weaker: host OS compromise exposes guests
Examples	ESXi, Hyper-V, KVM (debated), AWS Nitro	VirtualBox, VMware Workstation, Parallels

Type 1 versus Type 2

Property	VM	Container
Kernel	One guest kernel per VM	Shared host kernel
Start-up	Seconds to minutes	Milliseconds to about a second
Isolation boundary	Hypervisor and virtual hardware	System call interface of a shared kernel
OS family	Any the virtual hardware supports	Same as the host kernel only
Image size	Gigabytes	Megabytes

Virtual machine versus container

Aspect	Namespaces	Cgroups
Question	What can a process see?	How much can it use?
Mechanism	Own view of a kernel resource	Quotas, weights, accounting, freezing
Types or functions	PID, network, mount, UTS, IPC, user, cgroup, time	Limiting, prioritization, accounting, control
Example files or flags	CLONE_NEWPID, CLONE_NEWNET	cpu.max, cpu.weight, memory.max, cgroup.freeze

Namespaces versus cgroups

\begin{gathered} \text{CPUs} = \frac{\text{quota}}{\text{period}} = \frac{150000}{100000} = 1.5 \\ s_i = \frac{w_i}{\sum_j w_j} \end{gathered}

cpu.max is a hard ceiling; cpu.weight is a share that matters only under contention

On-premise cost and TCO

Five dual-socket Xeon E5-2640 v2 servers, open-source software, a 3-year horizon. Part 05: On-premise TCO case study

Aspect	CapEx	OpEx
When paid	Upfront and lumpy	Recurring and smooth
Accounting	Capitalised, then depreciated	Expensed in the period
Examples	Servers, SAN, switches, racks	Power, cooling, rent, salaries, maintenance
Cloud analogue	None, the provider owns the hardware	Pay as you go

CapEx versus OpEx

\text{CapEx} = \sum_i q_i \, p_i

Quantity times unit price over every asset bought at t = 0

\begin{gathered} E_{\text{yr}} = \frac{N \cdot P \cdot 8760}{1000}\ \text{kWh} \\ \text{cost}_{\text{yr}} = E_{\text{yr}} \times \text{price} \end{gathered}

Yearly energy and its cost

\text{TCO}_{3\,\text{yr}} = \text{CapEx} + 3 \times \text{OpEx}_{\text{yr}}

Three-year total cost of ownership

Case study numbers (slides 33 and 34)

Servers: 5 × 3,500 = 17,500 EUR
SAN: 35,000 EUR (≈50.9% of CapEx, 39.0% of TCO)
Switches: 4 × 3,677.50 = 14,710 EUR
Facilities, cooling kit: 897 + 717 EUR
CapEx: 68,824 EUR
Power: 5 × 308 W, 13,490 kWh at 0.22 EUR/kWh: 2,962 EUR/yr (slide)
Cooling: 5 × 385 W: 3,702 EUR/yr (slide)
Rent: 5 m² × 5 EUR × 12 = 300 EUR/yr
OpEx: 6,964 EUR/yr, 20,891 EUR over three years
TCO, 3 years: 89,715 EUR, about 498 EUR per server per month (76.7% CapEx)

\text{PUE} = \frac{E_{\text{facility}}}{E_{\text{IT}}} \geq \frac{308 + 385}{308} = 2.25

Implied PUE, against an industry average near 1.56 (2024)

Cloud cost and the break-even

A cloud bill is pure OpEx: unit price times metered quantity. Part 06: Cost case study and NIST models

\begin{gathered} C_{\text{month}} = \sum_{r} p_r \cdot q_r \\ C_{\text{year}} = 12\, C_{\text{month}} \end{gathered}

Cloud cost per resource line

Period	Traditional IT	AWS
Start	68,824	0
Each year	6,964	19,364 (5 VMs)
After 3 years	89,715	58,093
After 5 years	103,644	96,820
After 6 years	110,608	116,184

Traditional IT versus AWS, cumulative EUR

\text{savings} = \frac{89715-58093}{89715} \approx 35.2\%

Three-year saving of AWS

\begin{aligned} t^{*} &= \frac{\text{CapEx}}{C_{\text{cloud}}-\text{OpEx}_{\text{on}}} \\ &= \frac{68824}{19364-6964} \approx 5.55\ \text{years} \end{aligned}

Break-even year; servers are usually refreshed before it

NIST SP 800-145: 5-3-4

Five essential characteristics, three service models, four deployment models. Part 06: Cost case study and NIST models

Characteristic	Meaning	Example
On-demand self-service	Provision unilaterally, no human at the provider	Launch a VM from the console at 2 a.m.
Broad network access	Standard mechanisms, heterogeneous clients	Same storage from phone, laptop or script
Resource pooling	Multi-tenant, location independent	You pick a region, never a rack
Rapid elasticity	Scale out and in with demand, appears unlimited	Auto scaling group at peak
Measured service	Metered, reported to provider and consumer	Line items of the slide 35 bill

Five essential characteristics

Model	Users	Location
Public	The general public	On the provider's premises
Private	One organization	On or off premises
Community	Organizations with shared concerns	On or off premises
Hybrid	Composition of distinct clouds bound by portability technology	Spans its components

Four deployment models: who may use it

Cloud bursting is the canonical hybrid example.
Keep some traditional IT for latency, data residency, regulation and vendor lock-in.

Service models: IaaS, PaaS, SaaS

Machine means IaaS, runtime means PaaS, finished app means SaaS. Part 07: Service models

\begin{gathered} \text{customer-managed layers} \\ 9 \rightarrow 4 \rightarrow 1 \rightarrow 0 \end{gathered}

On-prem, IaaS, PaaS, SaaS as the slides draw it

Layer	On-prem	IaaS	PaaS	SaaS
Applications	Customer	Customer	Customer	Provider
Security	Customer	Customer	Provider	Provider
Databases	Customer	Customer	Provider	Provider
Operating systems	Customer	Customer	Provider	Provider
Virtualization	Customer	Provider	Provider	Provider
Server, storage, networking, data center	Customer	Provider	Provider	Provider

Who manages each layer (slide diagrams)

Type	Addressing	Access	AWS	Typical use
Block	Fixed-size block on a volume	Mounted as a raw disk by one server	EBS	Boot disk, database files
File	Path in a directory tree	Shared over the network	EFS	Shared home directories
Object	Unique ID in a flat namespace	HTTP API (PUT, GET)	S3	Media, backups, data lakes

Block, file and object storage

Public cloud hyperscalers

Hyperscale is economics: roughly 1/5 to 1/7 of medium-sized unit prices. Part 08: Public cloud hyperscalers

Market and history

Market, Q2 2026: $143.4B, +43% year on year
Shares: Amazon 28%, Microsoft 20%, Google 15%, top three 63%
AWS: S3 (March 2006), EC2 (August 2006), from Amazon's internal platform
Google Cloud: App Engine PaaS (April 2008), Compute Engine GA (December 2013)
Azure: Announced October 2008, GA February 2010, renamed Microsoft Azure in 2014
Spectrum: EC2 gives control (kernel up), App Engine gives convenience (constrained app shape), Azure 2009 sat between.

Block	Need	AWS	Google Cloud	Azure
Compute	Virtual machines	EC2	Compute Engine	Virtual Machines
Compute	Kubernetes	EKS (slide: ECS)	GKE	AKS
Storage	Object	S3	Cloud Storage	Blob Storage
Storage	Block	EBS	Persistent Disk	Managed Disks
Database	NoSQL	DynamoDB	Bigtable	Cosmos DB
Database	Relational	RDS	Cloud SQL	SQL Database
Networking	Private network	VPC	VPC	Virtual Network
Networking	DNS	Route 53	Cloud DNS	Azure DNS
Networking	CDN	CloudFront	Cloud CDN	Azure Front Door
Deployment and Management	Infrastructure as code	CloudFormation	Infrastructure Manager	Resource Manager
Deployment and Management	Identity and access	IAM	Cloud IAM	Entra ID with Azure RBAC
Application Services	Messaging	SQS, SNS	Pub/Sub	Queue Storage, Service Bus

Service equivalence by block (comparable, not identical)

Private cloud platforms

A private IaaS platform turns owned servers into self-service, pooled, metered resources for one organization. Part 09: Private cloud platforms

Service	Manages	AWS equivalent
Nova	Compute instances	EC2
Swift	Objects over HTTP	S3
Cinder	Block volumes	EBS
Neutron	Virtual networks	VPC
Keystone	Identity, tokens, catalog	IAM
Glance	Boot images	AMI catalog
Horizon	Web dashboard	Management Console
Placement	Host inventory and usage	No direct equivalent

Core OpenStack services

Aspect	OpenStack	OpenNebula
Origin	Rackspace and NASA	Complutense University of Madrid
First release	2010	2008
Steward	OpenInfra, under the Linux Foundation	OpenNebula Systems
Architecture	Many REST services, AMQP, SQL	Single front-end (oned) plus drivers
Main API	REST, authenticated by Keystone	XML-RPC
Hypervisors today	Mostly KVM	KVM and LXC
License	Apache 2.0	Apache 2.0

OpenStack versus OpenNebula

SLAs and availability

Each extra nine cuts the downtime budget tenfold. Part 10: SLAs and availability

\begin{gathered} A=\frac{\text{Uptime}}{\text{Uptime}+\text{Downtime}} \\ D=(1-A)\times T \end{gathered}

Availability and allowed downtime over period T

A_{est}=\frac{\text{MTBF}}{\text{MTBF}+\text{MTTR}}

Estimate from failure frequency (MTBF) and mean time to recover (MTTR)

Availability	Per year	Per month	Per day
99%	3.65 d	7.2 h	14.4 min
99.9%	8.76 h	43.2 min	1.44 min
99.95%	4.38 h	21.6 min	43.2 s
99.99%	52.56 min	4.32 min	8.64 s
99.999%	5.26 min	25.9 s	0.86 s

Allowed downtime (365-day year, 30-day month)

SLI, SLO, SLA

SLI: The measurement, for example the share of valid requests that succeeded.
SLO: An internal target for an SLI. Missing it triggers engineering work, not payments.
SLA: A contract naming SLOs and the consequence of missing them.
EC2 example: 99.99% monthly per region (99.5% single instance). Credits 10%, 30% below 99.0%, 100% below 95.0%. Credits are the sole remedy.

\begin{gathered} A_{\text{series}}=\prod_{i=1}^{n} A_i \le \min_i A_i \\ A_{\text{parallel}}=1-\prod_{i=1}^{n}(1-A_i) \end{gathered}

Series multiplies; independent replicas fail only together

Topology	Formula	Result
Series (hard dependencies)	A = ∏ Aᵢ ≤ min Aᵢ	99.8001%
Parallel (redundant replicas)	A = 1 − ∏ (1 − Aᵢ)	99.9999%

Two components at 99.9%

Slide errata

Answer with the corrected fact, and mention the slide's version when a question depends on it.

What the slides get wrong

Slide 10: "Mini Data Cener" should read Mini Data Center. Part 02
Slide 20: Full and para (guest modified?) are a separate axis from Type 1 and Type 2 (where it runs). VM/370 is 1972, not 1970/71. Part 03
Slide 22: vSphere is the suite; the Type 1 hypervisor is ESXi. KVM's label is debated. Part 04
Slide 24: Type 2 has less isolation than Type 1. 5% and 10% are workload-dependent approximations. Part 04
Slide 25: Virtual hardware is shared only when the container host is a VM. The SCI is the kernel entry point; the kernel manages hardware. Part 04
Slides 26, 27: Linux has eight namespace types. One namespace is not a container: a container is several namespaces plus cgroups plus a root filesystem. Part 04
Slide 32: Routine maintenance is OpEx (IAS 16 paragraph 12), not CapEx. Part 05
Slides 33, 34: European number format (17.500 € is 17,500 EUR). KVM is a keyboard-video-mouse switch. Rent is per m² per month. Part 05
Slide 35: Monthly rows sum to 318, not 323; yearly rows to 3,816, not 3,878. Part 06
Slide 36: The AWS column is five VMs (19,364 ≈ 5 × 3,873), never stated. Part 06
Slides 42, 44: "Comany" should read Company. Linking private clouds to the community cloud is a hybrid composition. Part 06
Slides 47, 48: The customer always keeps data, identities, access management and endpoints, even in PaaS and SaaS. Part 07
Slides 53, 54, 56: Gmail (2004) predates App Engine. Container Engine is GKE, Deployment Manager gives way to Infrastructure Manager, "Cosmo DB" is Cosmos DB, IAM maps to Entra ID plus RBAC. Part 08
Slides 60, 62: Titled "Public" but OpenStack and OpenNebula are private platforms. OpenNebula dropped Firecracker, LXD and vCenter drivers. Part 09
Slide 66: The second box should read Service 2, at 99.9%. Part 10
Slide 67: The calculator screenshot is internally inconsistent; 0.11% of a 365-day year is about 9 h 38 min. Part 10