Cloud-Based Data Processing

Data Centers
Jana Giceva

1
Datacenter Overview

2
Data Centers
 Data center (DC) is a physical facility that enterprises use to house computing and storage
infrastructure in a variety of networked formats.

 Main function is to deliver utilities needed

by the equipment and personnel:
- Power
- Cooling
- Shelter
- Security

 Size of typical data centers:

- 500 – 5000 sqm buildings
- 1 MW to 10-20 MW power (avg 5 MW)

3
Example data centers

4
Datacenters around the globe

https://docs.microsoft.com/en-us/learn/modules/explore-azure-infrastructure/2-azure-datacenter-locations
5
Modern DC for the Cloud architecture
 Geography:
- Two or more regions
- Meets data residency requirements
- Fault-tolerant from complete region failures

 Region:
- Set of datacenters within a metropolitan area
- Network latency perimeter < 2ms

 Availability Zones:
- Unique physical locations within a region
- Each zone made up of one or more DCs
- Independent power, cooling, networking
- Inter-AZ network latency < 2ms
- Fault tolerance from DC failure
Src: Inside Azure Datacenter Architecture with Mark Russinovich 6
Data Centers
 Traditional data centers
- Host a large number of relatively small- or medium-sized applications, each running on a dedicated
hardware infrastructure that is decoupled and protected from other systems in the same facility
- Usually for multiple organizational units or companies

 Modern data centers (a.k.a., Warehouse-scale computers)

- Usually belong to a single company to run a small number of large-scale applications
- Google, Facebook, Microsoft, Amazon, Alibaba, etc.
- Use a relatively homogeneous hardware and system software
- Share a common systems management layer
- Sizes can vary depending on needs

7
Datacenter Architecture

8
Scale-up vs. scale-out
 Scale-up: high cost powerful CPUs, more cores, more memory
 Scale-out: adding more low cost, commodity servers

 Supercomputer vs. data center

 Scale
- Blue waters = 40K 8-core “servers”
- Microsoft Chicago Data centers = 50 containers = 100K 8-core servers
 Network architecture
- Supercomputers: InfiniBand, low-latency, high bandwidth protocols
- Data Centers: (mostly) Ethernet based networks
 Storage
- Supercomputers: separate data farm
- Data Centers: use disk on node + memory cache
9
 Main components of a datacenter

src: The Datacenter as a Computer – Barroso, Clidaras, Holzle 10

Traditional Data Center Architecture
Servers mounted on 19’’
rack cabinets

Racks are placed in single rows forming

corridors between them.

Src: the datacenter as a computer – an introduction to the design of warehouse-scale machines 11

 A Row of Servers in a Google Data Center

Src: the datacenter as a computer – an introduction to the design

of warehouse-scale machines
12
Inside a modern data center
 Today’s DC use shipping containers packed with
1000s servers each.

 For repairs, whole containers are replaced.

13
Costs for operating a data center
 DCs consume 3% of global electricity supply
(416.2 TWh > UK’s 300 TWh)
Monthly cost = $3’530’920
4%
 DCs produce 2% of total greenhouse gas Servers
emissions 13%
Networking
Equipment
 DCs produce as much CO2 as The Netherlands
18% Power Distribution
or Argenti 57% & Cooling

8%
Power
31% power
Other Infrastructure
45,978 servers, 3yr server & 10 yr infrastructure amortization
45,978 servers, 3yr server & 10 yr infrastructure amortization

14
Power Usage Effectiveness (PUE)
 PUE is the ratio of
- The total amount of energy used by a DC facility
- To the energy delivered to the computing equipment

 PUE is the inverse of data center infrastructure efficiency

 Total facility power = covers IT systems (servers, network, storage) +

other equipment (cooling, UPS, switch gear, generators, lights, fans, etc.)

15
Achieving PUE
 Location of the DC – cooling and power load factor

 Raise temperature of aisles

- Usually 18-20 C; Google at 27 C
- Possibly up to 35 C (trade off failures vs. cooling costs)

 Reduce conversion of energy

- E.g., Google motherboards work at 12V
rather than 3.3/5V

 Go to extreme environments
- Arctic circle (Facebook)
- Floating boats (Google)
- Underwater DC (Microsoft)
16
Evolution of data center design
 Case study: Microsoft

https://www.nextplatform.com/2016/09/26/rare-tour-microsofts-hyperscale-datacenters/
17
Evolution of datacenter design
 Gen 6: scalable form factor (2017)
- Reduced infrastructure, scale to demand
- 1.17-1.19 PUE

 Gen 7: Ballard (2018)

- Design execution efficiency
- Flex capacity enabled
- 1.15-1.18 PUE

 Gen 8: Rapid deploy datacenter (2020)

- Modular construction and delivery
- Equipment skidding and preassembly
- Faster speed to market

 Project Natick (future) – 1.07 PUE or less

Src: Inside Azure Datacenter Architecture with Mark Russinovich 18
Datacenter Challenges

19
Challenge 1: Cooling data centers

Cooling plant at a Google DC in Oregon

20
Challenge 2: Energy Proportional Computing
 Average real world DC and servers are
too inefficient. Sub-system power usage in an x86 server as the compute
- waste 2/3+ of their energy load varies from idle to full (reported in 2012).

 Energy consumption is not proportional

to the load
- CPUs are not so bad but the other
components are
- CPU is the dominant energy consumer
in servers – using 2/3
of energy when active/idle.

 Try to optimize workloads

 Virtualization and consolidation. src: “The Datacenter as a Warehouse Computer”

21
Challenge 3: Servers are idle most of the time
 For non-virtualized servers 6-15% utilization

 Server virtualization can boost to

an average 30% utilization

 Need for resource pooling and application

and server consolidation

 Need for resource virtualization

src: Luiz Barroso, Urs Hölzle “The Datacenter as a Computer”

22
Challenge 4: Efficient monitoring
 Even with virtualization and software defined DC,
resource utilization can be poor.

 Need for efficient monitoring (measurement) and cluster

management.

 Goal to meet SLOs and SLIs.

 Job’s tail latency matters!

src: “Heterogeneity and dynamicity of clouds at scale: Google trace analysis” SoCC’12
23
Improving resource utilization
 Hyper-scale system management software
- Treat the datacenter as a warehouse scale computer
- Software defined datacenters
- System software that allows DC operations to manage the entire DC infrastructure
- Compose a system using pooled resources of
compute, network, and storage based on workload requirement

 Dynamic resource allocation

- Virtualization is not enough to improve efficiency
- Need the ability to dynamically allocate CPU resources across servers and racks, allowing admins to quickly
migrate resources to address the shifting demand
- Drive 100-300% better utilization for virtualized WLs, and 200-600% for bare-metal WLs.

24
Software defined DCs

Src: Inside Azure Datacenter Architecture with Mark Russinovich 25

Disaggregation across racks

Src: Inside Azure Datacenter Architecture with Mark Russinovich 26

Software defined Servers

Src: Inside Azure Datacenter Architecture with Mark Russinovich 27

Challenge 5: Managing scale and growth
 In 2016, Gartner estimated that Google has 2.5 million servers
 In 2017, Microsoft Azure was reported to have more than 3 million servers.

200,000+ servers
(estimated)

28
Size and growth of DC (2016-2020)
 The scale and complexity of DC operations grows
constantly.

 By 2020, Cisco estimated that we would have 600 million

GB of new data saved each day
(200 million GB big data)

 So the volume of BigData by 2020 was estimated to be as

much as all of the stored data in 2016.

29
Challenge 6: networking at scale

30
Challenge 6: networking at scale (cont.)
 Building the right abstractions to work for a range of workload at hyper-scale.

Software Defined Networking (SDN)

 Within DC, 32 billion GBs will be transported in 2020

- src: Cisco’s report 2016-2020

 “Machine to machine” traffic is orders of magnitude larger than what goes out on the Internet 31
Cloud Computing Overview

32
Big Data and the need for Cloud

https://www.visualcapitalist.com/every-minute-internet-2020/

33
Cloud and Cloud computing
Datacenter hardware and software that the vendors use to offer the computing resources and services.

 The cloud has a large pool of easily usable

virtualized computing resources, development
platforms, and various services and
applications.

 Cloud computing is the delivery of

computing as a service.

 The shared resources, software,

and data are provided by a provider
as a metered service over a network.

34
Cloud Computing
 Datacenters are vendors that rent servers or other computing resources (e.g., storage)
- Anyone (or company) with a “credit card” can rent
- Cloud resources owned and operated by a third-party (cloud provider).

 Fine-grained pricing model

- Rent resources by the hour or by I/O
- Pay as you go (pay for only what you use)

 Can vary capacity as needed

- No need to build you own IT infrastructure for peak loads

35
Cloud market revenue in billions of dollars

36
Cloud Roles
Application users

SaaS user Cloud Users

 Software / websites that serve real users
- Netflix, Pinterest, Instagram, Spotify, AirBnb, Lyft, Slack, Expendia
Web applications
 Data analytics, machine learning, and other data services
- Databricks, Snowflake, GE Healthcare
SaaS Provider /
 Mobile and IoT backends
Cloud user - Snapchat, Zynga (AWS → zCloud → AWS)
 Datacenter’s own software
Utility computing - Google Drive/One Drive, search, etc.

Cloud Providers
Cloud Provider  Companies with large DCs
- Amazon AWS, Microsoft Azure, Google Cloud Platform, Alibaba Cloud, …

37
Types of Cloud Computing
Public vs. Private
 Public: resources owned and operated by the one organization aka the cloud vendor
 Private: Resources used exclusively by a single business or organization

On-premise vs. Hosted:

 On-premise (on-prem): resources located locally (at a datacenter that the organization operates)
 Hosted: resources hosted and managed by a third-party provider

Private cloud can be both on-prem and hosted (virtual private cloud)

38
Types of Cloud Computing (cont)
Hybrid cloud
 Combines public and private clouds, allows data and applications to be shared between them.
 Better control over sensitive data and functionalities
 Cost effective, scales well and is more flexible
Multi-Cloud
 Use multiple clouds for an application / service
 Avoids data lock-in
 Avoids single point of failure
 But, need to deal with API differences and handle migration across clouds

39
Cloud service models (XaaS)
more Infrastructure as a Service (IaaS)
 Rent IT infrastructure – servers and virtual machines (VMs),
storage, network, firewall, and security

Platform as a Service (PaaS)

 Get on-demand environment for development, testing and management of software
applications: servers, storage, network, OS, databases, etc.
control

Serverless, Function as a Service (FaaS)

 Overlapping with PaaS, serverless focuses on building app functionality without
managing the servers and infrastructure required to do so.
 Cloud vendors provides set-up, capacity planning, and server management.
Software as a Service (SaaS)
 Deliver software applications over the Internet, on demand.
less  Cloud vendor handles software application and underlying infrastructure
40
Infrastructure as a Service
 Immediately available computing infrastructure,
provisioned and managed by a cloud provider.

Managed by user
 Computing resources pooled together to
server multiple users / tenants.

 Computing resources include:

storage, processing, memory, network bandwidth, etc.

cloud provider
src image from Microsoft Azure
41
Platform as a Service
 Complete development and deployment environment.

 Includes system’s software (OS, middleware),

user
platforms, DBMSs, BI services, and libraries to
assist in development and deployment of
cloud-based applications.

 Examples:

Managed by cloud provider

src image from Microsoft Azure 42
Software as a Service

43
Cloud pros and cons
User’s benefits: User’s concerns:
 Elimination of up-front commitment  Dependability on network and internet
 Speed – services are provided on demand connectivity
 Global scale and elasticity  Security and privacy
 Productivity  Cost of migration
 Performance and security  Cost and risk of vendor lock-in
 Customizability
 Ability to pay for use of computing resources
on a short-term basis (as needed)

44
References
In addition to the cross references provided in the slides.

Some material based on:

 Lecture notes from “Scalable Systems for the Cloud” by Prof. Giceva at Imperial
 Lecture notes from “Modern Data Center Systems” by Prof. Zhang at UC San Diego

 Book “The Datacenter as a Computer – An Introduction to the Design of Warehouse-scale Machines” by Luiz Andre
Barroso, Jimmy Clidaras, Urs Holzle
 Talk “Inside Azure Datacenter Architecture” with Mark Russinovich (Azure CTO)
 Paper “Above the Clouds: A Berkeley View of Cloud Computing”
 Web-pages from Amazon AWS, Microsoft Azure and Google CDP

45