1. Lecture 1. Cloud Economics
COSC6376 Cloud Computing
Lecture 1. Cloud Economics
Instructor: Weidong Shi (Larry), PhD
Computer Science Department
University of Houston

2. Topics
Cloudonomics

3. Web Sites
Class website:

4. What do I need from you Your name Skillset relevant Expectations
Your research or career goals

5. Amazon Education Program

6. Course Grading 30% assignments 70% project
10% assigned reading summaries (due before class)
20% programming assignments
70% project
15% for report one
15% for report two
25% for final report
15% for final presentation

7. NASA

9. Reading Assignment

10. Cloudonomics CLOUD  from an economic viewpoint: Common Infrastructure
pooled standardized resources, statistical multiplexing
Location-independence
ubiquitous availability meeting performance requirements
latency reduction and user experience enhancement
Utility Pricing
usage-sensitive or pay-per-use pricing
benefits environments with variable demand levels
On-demand Resources
scalable, elastic resources provisioned and de-provisioned without delay or costs associated with change

11. The Value of Common Infrastructure
Economies of Scale
Reduced overhead costs
Buyer power through volume purchasing
Statistics of Scale
For infrastructure built to peak requirements: Multiplexing demand  higher utilization
Lower cost per delivered resource than unconsolidated workloads
For infrastructure built to less than peak: Multiplexing demand  reduce the unserved demand
Lower loss of revenue or a Service-Level agreement violation payout.

12. Economics of Cloud Providers
5-7x economies of scale [Hamilton 2008]
Extra benefits
Amazon: utilize off-peak capacity
Microsoft: sell .NET tools
Google: reuse existing infrastructure
Resource
Cost in
Medium DC
Very Large DC
Ratio
Network
$95 / Mbps / month
$13 / Mbps / month
7.1x
Storage
$2.20 / GB / month
$0.40 / GB / month
5.7x
Administration
≈140 servers/admin
>1000 servers/admin

13. Economics of Cloud Providers
Cost of power. Currently representing 15%-20%, TCO. Power Usage Effectiveness tends to be significantly lower in large facilities than in smaller ones
Power Utilization Effectiveness equals total power delivered into a datacenter divided by critical power ( the power needed to actually run the servers).
Infrastructure labor costs. A single system administrator manages thousands of servers in larger data centers.
Buying power. Operators of large data centers can get discounts on hardware purchases of up to 30 percent over smaller buyers.

14. Economies of Scale in the Cloud
A 100,000-server datacenter has an 80% lower total cost of ownership (TCO) compared to a 1,000-server datacenter.

15. Economics of Cloud Users
Risk of over-provisioning: underutilization

16. Economics of Cloud Users
Risk of over-provisioning: underutilization
Demand
Capacity
Time
Resources
Unused resources
Static data center

17. Data center in the cloud
Cloud Economics
Pay by use instead of provisioning for peak
Demand
Capacity
Time
Resources
Demand
Capacity
Time
Resources
Unused resources
Static data center
Data center in the cloud

18. Economics of Cloud Users
Heavy penalty for under-provisioning
Resources
Demand
Capacity
Time (days) 1 2 3
Resources
Demand
Capacity
Time (days) 1 2 3
Lost revenue
Resources
Demand
Capacity
Time (days) 1 2 3
Lost users

19. Economics of Cloud Users
Cloud computing has the ability to add or remove resources at a fine grain and with a lead time of minutes rather than weeks. Thus it matches the workload much more closely.
e.g. Target uses AWS for the Target.com website. While other retailers had severe performance problems on “Black Friday(Nov 28)”, Target was only slower by about 50%.

20. Twitter Usage by Hour

21. Twitter Usage by Day of Week

22. Weekday Hourly Traffic

23. Weekly Traffic

24. Work Output for a Production Data Center Over a Three-week Period
Work Output for a Production Data Center Over a Three-week Period. Data was collected over a three-week period from November 29, 2009 to December 19, 2009 and includes approximately 3900 servers.

25. Workload Characterization and Prediction

26. Workload Randomness (Exchange Server)
Source: Microsoft.
User access patterns contain a certain degree of randomness. For example, people check their at different times.

27. Cloud Computing Economics
The risk of mis-estimating workload is shifted from the service operator to the cloud vendor.
Services like Google Appengine automatically scales in response to load increases and decreases.
That is why the price is expensive.

28. A Useful Measure of “Smoothness”
The coefficient of variation CV 
≠ the variance σ2 nor the correlation coefficient 
Ratio of the standard deviation σ to the absolute value of the mean |μ|
“smoother” curves:
large mean for a given standard deviation
or smaller standard deviation for a given mean
Importance of smoothness: 
a facility with fixed assets servicing highly variable demand will achieve lower utilization than a similar one servicing relatively smooth demand. 
Multiplexing demand from multiple sources may reduce the coefficient of variation CV

29. Coefficient of Variation CV
X1, Xn ,…, Xn independent random variables for demand
Identical standard variation σ and mean μ
Aggregated demand:
Mean  sum of means: n. μ
Variance  sum of variances: n.σ2
Coefficient of variance  𝑛 .𝜎 𝑛.𝜇 = 𝜎 𝑛 .𝜇 = 1 𝑛 𝐶𝑣
Adding n independent demands reduces the Cv by 1/√n
Penalty of insufficient/excess resources grows smaller
Aggregating 100 workloads bring the penalty to 10%

30. Running the same workload for multiple time zones on the same servers or by running workloads with complementary time-of-day patterns (for example, consumer services and enterprise services) on the same servers.

31. Dependent Workloads? Negative correlation demands
X and 1-X  Sum is random variable 1
Appropriate selection of customer segments
Perfectly correlated demands
Aggregated demand: n.X, variance of sum: n2𝜎2(X)
Mean: n.μ, standard deviation: n.𝜎(X)
Coefficient of Variance remains constant
Simultaneous peaks

32. Value of Location Independence
We used to go to the computers , but applications, services and contents now come to us!
Through networks: Wired, wireless, satellite, etc.
But what about latency?
Human response latency: 10s to 100s miliseconds
Latency is correlated with:
Distance (Strongly)
Routing algorithms of routers and switches (second order effects)
Speed of light in fiber: only 124 miles per milisecond
If the Google word suggestion took 2 seconds 
VOIP with latency of 200ms or more 
Supporting a global user base requires a dispersed service architecture

33. Value of Utility Pricing
Cloud services don’t need to be cheaper to be economical!
Consider a car
Buy or lease for $10 per day
Rent a car for $45 a day
If you need a car for 2 days in a trip, buying would be much more costly than renting
It depends on the demand

35. Workload Variability
Retail firms see a spike during the holiday shopping season while U.S. tax firms will see a peak before April 15.
If capacity is built for the expected peak (plus a margin of error). Most of this capacity will sit idle the rest of the time.

36. Utility Pricing in Detail
D(t): demand for resources, 0<t<T
P = max( D(t) ) : Peak Demand
A = Avg( D(t) ) : Average Demand
B = Baseline (owned) unit cost ; BT = Total Baseline Cost
C = Cloud unit cost; CT = Total Cloud Cost
U = C / B : Utility Premium
For the rental car example, U=4.5
CT = 0 𝑇 𝑈×𝐵×𝐷 𝑡 𝑑𝑡=𝐴×𝑈×𝐵×𝑇
BT = 𝑃×𝐵×𝑇
Because the Baseline should handle peak demand
When is cloud cheaper than owning?
𝑪 𝑻 < 𝑩 𝑻 →𝐴×𝑈×𝐵×𝑇<𝑃×𝐵×𝑇→𝑼< 𝑷 𝑨
When utility premium is less than ratio of Peak demand to Average demand

37. Utility Pricing in Real World
In practice demands are often highly spiky
News stories
Marketing promotions
Product launches
Internet flash floods (slashdot effect)
Tax season
Christmas shopping
etc.

38. Utility Pricing in Real World
Often a hybrid model is the best
You own a car for daily commute, and rent a car when traveling or when you need a van to move
Key factor is again the ratio of peak to average demand
But we should also consider other costs
Network cost (both fixed costs and usage costs)
Interoperability overhead
Consider Reliability, accessibility

39. Example
200 Amazon EC2 instances

40. Value of on-Demand Services
Simple problem: when owning your resources, you will pay a penalty whenever your resources do not match the instantaneous demand
Either pay for unused resources, or suffer the penalty of missing service delivery
D(t) : Instantaneous Demand at time t
R(t) : Resources at time t
Penalty cost ∝ |𝐷 𝑡 −𝑅 𝑡 |𝑑𝑡
If demand is flat, penalty = 0
If demand is linear, periodic provisioning is acceptable

41. TechCrunch Effect
It is difficult to predict future need for computing resources (particularly true for tech startups).
It takes long lead-time to bring new capacity online.
It is common that approval for IT investments is required well in advance of actually knowing the actual demand for infrastructure.

42. E.g. Penalty Costs for Exponential Demand
If demand is exponential (D(t)=et), any fixed provisioning interval (tp) according to the current demands will fall exponentially behind
R(t) = 𝑒 𝑡− 𝑡 𝑝
D(t) – R(t) = 𝑒 𝑡 − 𝑒 𝑡− 𝑡 𝑝 = 𝑒 𝑡 1− 𝑒 𝑡 𝑝 = 𝑘 1 𝑒 𝑡
Penalty cost ∝ c.k1et

43. Youtube

44. Twitter Users

45. Data Center as a Computer
Consider a data center as a giant computer
Accelerators for the data center computer

46. Cloud Accelerators

47. Pricing for GPU Node Transcode 1,000 hours video data using cloud
Assume
Transcoding with dedicated GPU is 10x faster than CPU
CPU rental fee is $0.1 per hour
GPU node is $0.5 per hour
From customer perspective
1,000 hours x $0.1 per hour > 1,000 hours x 10% x $0.5 per hour

48. Behavioral Cloudonomics
Humans do not always make purely rational and quantitative decisions
Cons:
“loss aversion:” people generally get less satisfaction from gaining a dollar than they feel pain from losing one
customers may recognize the financial advantage of pay-per-use, but avoid it due to a “flat-rate” bias
E.g. fear of an unexpected large monthly cell phone bill  a flat-rate plan > measured service
Pros:
special attraction of “free” 
The lack of upfront investment in using the cloud becomes extremely attractive

49. Variable Pricing in Electricity
To further smooth demand, sophisticated pricing can be employed.
Similar to the electricity market, customers can be incented to shift their demand from high utilization periods to low utilization periods.

50. Classification of Clouds Based on Economic Models
Weinman mentions a “classification” of clouds based on economic models
Compute Clouds
Hotel Clouds
Rental Car Clouds
Restaurant Clouds
Etc.