+ All Categories
Home > Documents > Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... ·...

Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... ·...

Date post: 25-May-2020
Category:
Upload: others
View: 0 times
Download: 0 times
Share this document with a friend
187
Cloud Economics Ian Kash
Transcript
Page 1: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Cloud EconomicsIan Kash

Page 2: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Outline of the Tutorial9:00-10:30 Economics of a giant pile of compute resources Spot markets and reservations10:30-11:00 Coffee Break 11-12:30 Concrete resources & beyond compute Future of cloud economics

Page 3: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

The Cloud

Page 4: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 5: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

The promise of the cloud

Infinite resources! Pay only for what you need!

Image licensed by Yau Hoong Tang CC BY-NC-ND 2.0

Page 6: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Public Cloud DC

Page 7: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Economics of a Datacenter

Page 8: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Why have a cloud?

Utilization

On prem: 5-10% – IDC / VMWare 2009 12-18% - NRDC 2014 <20 percent – AWS Blog 2015

Page 9: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Reasons for low utilization

"The Economics of the Cloud" Harms and Yamartino 2010

Page 10: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Reasons for low utilization

"The Economics of the Cloud" Harms and Yamartino 2010

Page 11: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Reasons for low utilization

"The Economics of the Cloud" Harms and Yamartino 2010

Page 12: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Reasons for low utilization

"The Economics of the Cloud" Harms and Yamartino 2010

Page 13: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Reasons for low utilization

"The Economics of the Cloud" Harms and Yamartino 2010

Page 14: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

(New) reasons for low utilization

Not yet in steady state

Capacity is discrete

Page 15: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

The cloud solves utilization? Opportunities are real But exploiting them requires solving hard problems Coordination Information Re-engineering Pricing

Better, but still low: AWS claims 65%, but is that billed or real?

Page 16: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Fixed Costs vs Marginal Costs

Page 17: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Cloud Fixed Costs? Servers Infrastructure: Racks, Cabling, Cooling Building Land Software Labor

Page 18: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Cloud Marginal Costs Power Cooling Software Licensing

Page 19: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Economics of Competition Bertrand Competition – decide on price Software Cell Phones Restaurants Airlines?

Cournot Competition – decide on quantity Agriculture Oil Hotel Rooms

Page 20: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Public cloud is profitable!

Page 21: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Price Matching

Page 22: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Other benefits?

“To Move or Not to Move” Tak, Urgaonkar, Sivasubramaniam 2011

Page 23: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Economies of scale

Cheaper power / cooling – locate where it is cheap Buying power – power, hardware, software, capital, … Automation

Page 24: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Renting a VM

Page 25: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 26: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 27: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Hadoop IaaS: Get a bunch of VMs and install Hadoop

PaaS: Amazon EMR

SaaS?: Cloudera

Page 28: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

IaaS vs PaaS vs SaaS

"The Economics of the Cloud" Harms and Yamartino 2010

Page 29: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 30: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Spot Markets

Page 31: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Easy 100% Utilization

Page 32: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Should there be a spot market?

Page 33: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Low prices!

“On-demand, Spot, or Both” Menache, Shamir, Jain 2014

Page 34: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

People are Rational “On-demand, Spot, or Both” Menache, Shamir, Jain

2014 “Bidding Strategies for Spot Instances” Karunakaran

and Sundarraj 2015 “Supercloud” van Renesse, Weatherspoon, Shen,

Song 2018

Page 35: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

A Cautionary Tale…

“Deconstructing Amazon EC2 Spot Instance Pricing” Ben-Yehuda et al. 2012

Page 36: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Another Cautionary Tale?

Page 37: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Spot markets as price discrimination1. Model and equilibrium characterization for

system with PAYG + Spot

2. Analysis of restricted case showing adding Spot hurts revenue

3. Numerical evidence that suggests this is typically true

“Fixed and Market Pricing for Cloud Services.” Abhishek, Kash, and Key 2012

Page 38: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Model

“Fixed and Market Pricing for Cloud Services.” Abhishek, Kash, and Key 2012

Page 39: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Modeling PAYG

“Fixed and Market Pricing for Cloud Services.” Abhishek, Kash, and Key 2012

Page 40: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Modeling Spot

“Fixed and Market Pricing for Cloud Services.” Abhishek, Kash, and Key 2012

Page 41: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

EquilibriumThm: There is a unique equilibrium* where• All jobs truthfully report their type and cost• Each type i has a cost cutoff c̅i s.t. Joins Spot if c < c̅i Joins PAYG or balks otherwise.

*See details in paper

“Fixed and Market Pricing for Cloud Services.” Abhishek, Kash, and Key 2012

Page 42: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Impossibly General?• GI/GI/k• No specified auction design Assume reserve price is 0 Assume priorities are not randomized

“Fixed and Market Pricing for Cloud Services.” Abhishek, Kash, and Key 2012

Page 43: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Impossibly General?• GI/GI/k• No specified auction design Assume reserve price is 0 Assume priorities are not randomized

Insights from auction theory:• Can assume bidders just report c• Waiting time will be decreasing in c• All that matters is the (expected) delay

“Fixed and Market Pricing for Cloud Services.” Abhishek, Kash, and Key 2012

Page 44: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Approach to Theorem

“Fixed and Market Pricing for Cloud Services.” Abhishek, Kash, and Key 2012

Page 45: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Main Revenue TheoremThm: If the revenue maximizing price for PAYG + Spot is low enough that both types participate in PAYG, then:

Revenue(PAYG + Spot) < Revenue(PAYG)

“Fixed and Market Pricing for Cloud Services.” Abhishek, Kash, and Key 2012

Page 46: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Mostly holds in other case too…

“Fixed and Market Pricing for Cloud Services.” Abhishek, Kash, and Key 2012

Page 47: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Related Models• “To Queue or Not to Queue” Hassin and Haviv 2003• “Optimal Price and Delay Differentiation in Queueing

Systems” Maglaras, Yao, Zeevi 2013• “On-demand or Spot? Selling the cloud to risk-averse

customers” Hoy, Immorlica, Lucier 2016• “Pricing and bidding strategies for cloud computing spot

instances” Song and Guerin 2017• “The Spot Market Strikes Back” Dierks and Seuken 2018

Page 48: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

A Cautionary Tale…

Page 49: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

A Cautionary Tale…

“Paris Metro Pricing for the Internet” Odlyzko 1999 Use this style of pricing for network QoS

“Internet Service Classes Under Competition” Gibbens, Mason, Steinberg 2000 Breaks down under competition

Page 51: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Epilogue

Page 52: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Reservations

Page 53: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Long-term reservations

Page 54: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Long-term reservations

Page 55: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Long-term reservations

Page 56: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Part-time Reservations

Page 57: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Length-based Pricing - Model

One server One job arrives per time period Jobs want to use the server for 1+ time periods Shared value per unit time distribution

"Simple Pricing Schemes for the Cloud" Kash, Key, Suksompong 2017

Page 58: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Length-based Pricing - Options

Complex: one price per job length

Simple: on price per unit time

Simpler: that price is chosen from among those used by the complex policy

"Simple Pricing Schemes for the Cloud" Kash, Key, Suksompong 2017

Page 59: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Length-based Pricing - Results

Simpler pricing gets at least 50% of the benefits

This is tight

Simple pricing does too under somewhat less restrictive assumptions but only with optimal pricing

"Simple Pricing Schemes for the Cloud" Kash, Key, Suksompong 2017

Page 60: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Length-based Pricing - Intuitions Longer jobs have higher opportunity cost

With 2 lengths: low price gets at least the revenue from the short jobs and high price from the long ones One of these must be half the revenue

With >2 lengths: more careful about revenue from other lengths

"Simple Pricing Schemes for the Cloud" Kash, Key, Suksompong 2017

Page 61: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Online SchedulingEach job has: An arrival time 𝑎𝑎𝑗𝑗 A duration 𝑙𝑙𝑗𝑗 A deadline 𝑑𝑑𝑗𝑗 A value 𝑣𝑣𝑗𝑗 with density 𝜌𝜌𝑗𝑗 = 𝑣𝑣𝑗𝑗

𝑙𝑙𝑗𝑗

Key assumption: Slack parameter 𝑠𝑠: 𝑑𝑑𝑗𝑗 − 𝑎𝑎𝑗𝑗 ≥ 𝑠𝑠 ⋅ 𝑙𝑙𝑗𝑗

"Efficient Online Scheduling for Deadline-Sensitive Jobs." Lucier et al. 2013

Page 62: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Online Scheduling

"Efficient Online Scheduling for Deadline-Sensitive Jobs." Lucier et al. 2013

Page 63: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Making this Truthful

“Truthful Online Scheduling with Committments." Azar et al. 2015

Page 64: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Stochastic online scheduling

• At each time 𝑡𝑡 ∈ {0, … ,𝑇𝑇}, a job 𝑗𝑗 is realised from the known distribution 𝐷𝐷𝑡𝑡.

• We have to accept or reject the job right away.

Theorem: There is a mechanism for stochastic online scheduling on a single machine with uniform lengths that gives a 4 approximation in expectation

Page 65: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Trick #1: Discretization

• Assume every job has the same length 𝑙𝑙.• We partition the time into time slots of size 2𝑙𝑙.• We consider two partitions: even partition (blue)

and odd partition(red).10 32 54 76 98 1110 1312 1514 16

Page 66: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Trick #1: Discretization

• Claim: Given a subset of scheduled jobs 𝑆𝑆, there is a matching from each job in 𝑆𝑆 to exactly

one partition.

10 32 54 76 98 1110 1312 1514 16

Page 67: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Trick #1: Discretization

• Claim: Given a subset of scheduled jobs 𝑆𝑆, there is a matching from each job in 𝑆𝑆 to exactly

one partition.

10 32 54 76 98 1110 1312 1514 16

Page 68: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Trick #1: Discretization

• Claim: Given a subset of scheduled jobs 𝑆𝑆, there is a matching from each job in 𝑆𝑆 to exactly

one partition.

10 32 54 76 98 1110 1312 1514 16

Page 69: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Trick #1: Discretization

• Choose one of the partitions randomly.• The value we get is exactly half in expectation.

10 32 54 76 98 1110 1312 1514 16

Page 70: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Trick #2: Expected LPs

Page 71: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Trick #3: Prophet Inequalities

Page 72: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Trick #3: Prophet Inequalities

0.5 0.3 0.9

Page 73: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Trick #4: Bellman Equation

AKA Dynamic Programming

Calculate a price for each time slot at each time

Page 74: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Mechanism

Page 75: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Length Heterogeneity

• Assume 𝑙𝑙𝑗𝑗 ∈ 1, … ,𝐿𝐿 .• Consider log 𝐿𝐿 layers / servers.• 𝑘𝑘𝑡𝑡𝑡 layer is responsible for jobs with length 2𝑘𝑘−1 ≤𝑙𝑙 < 2𝑘𝑘.

• In each layer the ratio of the longest job to the shortest job is at most 2.

Page 76: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Value Heterogeneity

• Assume 𝑣𝑣𝑗𝑗 ∈ [1,𝑉𝑉].• Do the same trick.• Consider log 𝑉𝑉 layers / servers.• 𝑘𝑘𝑡𝑡𝑡 layer is responsible for jobs with value 2𝑘𝑘−1 ≤𝑣𝑣 < 2𝑘𝑘.

• In each layer the ratio of the highest valued job to the lowest valued job is at most 2.

Page 77: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Algorithm

[1, 2)

[2, 4)

[4, 8)

[8, 16)

.

.

.

[2𝑙𝑙 , 2𝑙𝑙+1)

1$ 2$ 4$ 8$ 16$ 2 log 𝑉𝑉 $..

Dur

atio

n of

the

job

Price of the machine

Page 78: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Algorithm

[1, 2)

[2, 4)

[4, 8)

[8, 16)

.

.

.

[2𝑙𝑙 , 2𝑙𝑙+1)

1$ 2$ 4$ 8$ 16$ 2𝑣𝑣$..

Dur

atio

n of

the

job

Price of the machine

𝑙𝑙 = 5𝑑𝑑 = 10𝑣𝑣 = 10$

Page 79: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Algorithm

[1, 2)

[2, 4)

[4, 8)

[8, 16)

.

.

.

[2𝑙𝑙 , 2𝑙𝑙+1)

1$ 2$ 4$ 8$ 16$ 2𝑣𝑣$..

Dur

atio

n of

the

job

Price of the machine

𝑙𝑙 = 5𝑑𝑑 = 10𝑣𝑣 = 10$

Page 80: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Algorithm

[1, 2)

[2, 4)

[4, 8)

[8, 16)

.

.

.

[2𝑙𝑙 , 2𝑙𝑙+1)

1$ 2$ 4$ 8$ 16$ 2𝑣𝑣$..

Dur

atio

n of

the

job

Price of the machine

𝑙𝑙 = 5𝑑𝑑 = 10𝑣𝑣 = 10$

Page 81: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Other Related Work “A Truthful Mechanism for Value-Based Scheduling in

Cloud Computing” Jain et al. 2017 “Truth and Regret in Online Scheduling” Chawla et al.

2017 “Stability of Service under Time-of-Use Pricing”

Chawla et al. 2017 “Selling reserved instances in cloud computing”

Wang et al. 2015

Page 82: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

ERA

"ERA: A Framework for Economic Resource Allocation for the Cloud" Babaioff et al. 2017

Page 83: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Airline Pricing?

"Yield Management at American Airlines" Smith, Leimkuhler, Darrow 1992

Page 84: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Coffee!

Page 85: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Cloud EconomicsIan Kash

Page 86: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Part II: Beyond Abstract Compute

Page 87: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Cluster Scheduling

Page 88: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Public Cloud DC

Page 89: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Lots of Options

Page 90: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

VM Series

Page 91: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Generations

Page 92: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Cluster Scheduling Contraints Heterogeneous Clusters What to do about old generations on new CPUs? Underclock? Share cores? Unreliable Performance?

Failure Domains Fragmentation Cores Memory Specialized Hardware

Page 98: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Work on Cluster Scheduling “Resource Central: Understanding and Predicting

Workloads for Improved Resource Management in Large Cloud Platforms” Cortez et al. 2017

“More Than Bin Packing: Dynamic Resource Allocation Strategies in Cloud Data Centers.” Wolke et al. 2015

Page 99: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

When to Introduce Next Generation? Technology improves at a linear rate with time

Users live for 2, have value 𝜃𝜃𝑡𝑡 for time 𝑡𝑡 technology

𝜃𝜃~𝐹𝐹 with monotone hazard rate

New generations cost 𝐶𝐶 to introduce, 𝑐𝑐 to adopt "Optimal Pricing and Introduction Timing of New Virtual Machines" Kash, Key, Zoumpoulis 2018

Page 100: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Myerson Pricing

Revenue of only offering technology 𝑡𝑡:1−𝐹𝐹 𝑝𝑝 𝑝𝑝𝑡𝑡

Charge optimal price 𝑝𝑝∗:𝑝𝑝∗ = (1 − 𝐹𝐹 𝑝𝑝∗ )/𝑝𝑝∗

Do this for every technology

"Optimal Pricing and Introduction Timing of New Virtual Machines" Kash, Key, Zoumpoulis 2018

Page 101: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Myerson Pricing => Periodic Introductions

New customers choose the latest technology

Existing customers may switch, depending on the time since last introduction

If we instead assume periodic introductions, this also shows Myerson is asymptotically optimal

"Optimal Pricing and Introduction Timing of New Virtual Machines" Kash, Key, Zoumpoulis 2018

Page 102: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

With arbitrary introductions

"Optimal Pricing and Introduction Timing of New Virtual Machines" Kash, Key, Zoumpoulis 2018

Page 103: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Storage

Page 104: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Kryder’s Law

Souce: Wikimedia Commons (Public Domain)

Page 105: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

But not throughput…

Souce: R1Soft

Page 106: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

“Max” Contracts

Page 107: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

“Max” Contracts

Page 108: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Pelican Rack

"Pelican: A Building Block for Exascale Cold Data Storage" Balakrishnan et al. 2014

Page 109: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Erasure Coding

Group data blocks into sets of k=15 Add r=3 redundancy blocks Any 15/18 suffice to recover the data

Page 110: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Pelican Rack

"Pelican: A Building Block for Exascale Cold Data Storage" Balakrishnan et al. 2014

Page 111: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Pelican Rack

“Feeding the Pelican” Black et al. 2016

Page 112: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 113: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

"Demonstration of End-to-End Automation of DNA Data Storage" Takahashi et al. 2019

Page 114: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Storage Economics

High throughput and low latency are expensive Initial pricing policies try and capture this Lots of need to improve on both the technology and

pricing sides

Page 115: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Network

Page 116: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

A Cautionary Tale…

https://github.com/stickfigure/blog/wiki/The-Unofficial-Google-App-Engine-Price-Change-FAQ

Page 117: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Performance Isolation is Hard

"Runtime Measurements in the Cloud: Observing, Analyzing, and Reducing Variance" Schad et al. 2010

Page 118: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Performance Isolation is Hard

"Managing Performance Overhead of Virtual Machines in Cloud Computing“Xu et al. 2014

Page 119: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Other Performance Isolation Work “Better Never than Late: Meeting Deadlines in

Datacenter Networks” Wilson et al. 2011 “The Price Is Right: Towards Location-Independent

Costs in Datacenters” Ballani et al. 2011 “Performance Isolation and Fairness for Multi-Tenant

Cloud Storage” Shue, Friedman, and Shaik 2012 “Chatty Tenants and the Cloud Network Sharing

Problem” Ballani, Jang, Karagiannis 2013

Page 120: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

What is fair?

Page 121: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 122: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Homogeneous Divisible Goods

0

0.1

0.2

0.3

0.4

0.5

0.6

Agent 1 Agent 2 Agent 3

Good 1 Good 2

Page 123: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Axiomatic Approach1) Sharing Incentives (SI) – Everyone gets 1/n

2) Envy Freeness (EF) – Everyone prefers his own

3) Strategyproofness (SP) – Truth-telling is optimal

4) Pareto Optimality (PO) – Nothing wasted

Page 124: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Leontief Utilities

Page 125: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Leontief Utilities

Page 126: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Leontief Utilities

Page 127: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

“Everyone gets the same share of his dominant resource”

Page 128: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

0

0.1

0.2

0.3

0.4

0.5

0.6

Agent 1 Agent 2 Agent 3

Good 1 Good 2

Page 129: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

Agent 1 Agent 2 Agent 3

Good 1 Good 2

Page 130: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

Page 131: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

Agent 1 Agent 2 Agent 3

Good 1 Good 2

Page 132: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

Agent 1 Agent 2 Agent 3

Good 1 Good 2

Page 133: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

Agent 1 Agent 2 Agent 3

Good 1 Good 2

Page 134: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

Agent 1 Agent 2 Agent 3

Good 1 Good 2

Page 135: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

Agent 1 Agent 2 Agent 3

Good 1 Good 2

Page 136: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

Page 137: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Dominant Resource Fairness

Theorem: DRF satisfies SI + EF + SP + PO

Page 138: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 139: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 140: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 141: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Page 142: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

One VM in compute server in compute rack

compute storage

Page 143: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

One VM in compute server in compute rack

One VHD in storage server in storage rack

compute storage

Page 144: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Page 145: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Page 146: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Page 147: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Page 148: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Page 149: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

Page 150: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Page 151: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Small customer : one VM accessing storage

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Page 152: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Result: a multi-resource “demand vector”

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Page 153: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Encodes resource id and proportions

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Page 154: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Encodes resource id and proportions

TX1 RX1

VTOR1 VTOR2

SSDb

VTORa VTORb

TXb RXb

VTOR1TX1 VTORb RXb SSDb TXb VTORa VTOR2 RX1

compute storage

Any element could be a bottleneck to performance

Page 155: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

• ~10^5-10^6 agents in a datacenter• Similar number of resources• ~1-10 second control interval• DRF is Quadratic

Page 156: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

DC-DRF: two tactics to improve scalability

1. Algorithmic: extending EDRF Operate to a time deadline chosen by operator (“control interval”) Variable degree of approximation: trading resource utilization for time Treat any resource that is 𝜖𝜖-close to exhausted as exhausted

2. HPC: maximize rate of computation Parallel where possible Optimize for thread and NUMA locality SIMD vector instructions

Page 157: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Utilization relative to baseline

Page 158: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

• Jobs are DAGs

• Room for efficiency at no cost to fairness

• More generally fairness-efficiency tradeoffs

Page 159: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Example• 2 jobs• 1 processor• Each requires 1 unit of processor time

• DRF: • Give each ½ the processor

• Efficient:• One then the other

Page 160: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Data and Machine Learning

Page 161: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Paying for ML Models

Page 162: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Paying for Data?

Page 163: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Potential Issues with ML Market Design

“Model Stealing”

Combining Models Granger causality? Credit assignment Connections to explanability

Page 164: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

The Future

Page 165: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Non-linear pricing?

Page 166: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Linear Pricing

Page 167: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Reasons for non-linear costs

VM type Service Size Availibility needs Duration Scale-outs

Page 168: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Shapley Value / Cost

Consider all possible arrival orders For each order, compute marginal cost Pay average marginal cost

Page 169: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Fair Cost Sharing

"The Shared Assignment Game" Blocq, Bachrach, and Key 2014

Page 170: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

“Max” Contracts

Page 171: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Pretium – WAN Bandwidth Pricing

"Dynamic Pricing and Traffic Engineering for Timely Inter-Datacenter Transfers" Jalaparti et al. 2016

Page 172: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Job-based pricing

"Bridging the Tenant-Provider Gap in Cloud Services" Jalaparti et al. 2012

Page 173: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Job-based pricing

"Exploiting Time-Malleability in Cloud-based Batch Processing Systems" Mai, Kalyvianaki, Costa 2013

Page 174: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Job-based pricing

"Exploiting Time-Malleability in Cloud-based Batch Processing Systems" Mai, Kalyvianaki, Costa 2013

Page 175: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Information Elicitation

Page 176: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

What not to do

Page 177: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Examples of intent elicitationProduct Information

Pay as You Go (Azure, AWS, Google) None

Quota (Azure, AWS, Google) Peak Demand

Fine-grained budgets (AWS) Bound on total usage

Reserved Instances (AWS, Azure) Heavy / light workload

Scheduled Reserved Instances (AWS) Heavy use in a particular pattern

Sport Market (AWS) / Evictable (Azure, Google) High / low value jobs

Tiered Storage pricing (Azure) / Glacier (AWS) Data hot / cold

??? VM short-lived / long-lived

??? Usage steady / bursty

??? Heavy usage at a particular time

Page 178: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Quotas• Provide a “guarantee” to customers• Provide information about peak usage• Allow Azure to do capacity control• Enable customer governance of end users• But always a headache for someone Small quotas require customer management Big quotas are costly for Azure Manual negotiation process

Page 179: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Storage SLAs

Page 180: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 181: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into
Page 182: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Explain Future Access Patterns

Page 183: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Cluster Admission

"On the Cluster Admission Problem for Cloud Computing" Dierks, Kash, Seuken 2019

Page 184: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Cluster Admission

"On the Cluster Admission Problem for Cloud Computing" Dierks, Kash, Seuken 2019

Page 185: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Cluster Admission

"On the Cluster Admission Problem for Cloud Computing" Dierks, Kash, Seuken 2019

Page 186: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Cluster Admission - Pricing Sell Options that permit scale outs

Variance-based pricing

"On the Cluster Admission Problem for Cloud Computing" Dierks, Kash, Seuken 2019

Page 187: Title of Presentation - University of Illinois at Chicagoiankash/slides/Cloud Economics... · 2019-05-16 · Economics of a giant pile of compute resources ... Group data blocks into

Thanks!


Recommended