Cloud Computing

CP 424

Lecture 3
Cloud Infrastructure

1
Overview
1. Historical Perspective of Data Centers
2. Datacenter Components: IT Equipment and Facilities
3. Design Considerations: Requirements, Power, Efficiency, &
Redundancy, Power Calculations, PUE and Challenges in
Cloud Data Centers
4. Cloud Management and Cloud Software Deployment
Considerations

2
1. Historical Perspective of Data
Centers
• A data center is often built with a large number of servers
through a huge interconnection network.
• The data centers are built economics of scale— meaning lower
unit cost for larger data centers.
• A small data center could have 1,000 servers and the larger the
data center, the lower the operational cost.
• The approximate monthly cost to operate a huge 400-server
data center is estimated by network cost $13/Mbps; storage
cost $0.4/GB; and administration costs.

3
Fig.1: A huge data center that is 11 times the size of a football field, housing 400,000 to 1 million servers. 4
2. Datacenter Components: IT
Equipment and Facilities
• The enterprise data center networks connected clients with
servers and storage using traditional LAN technology.
• The data center normally buys networking devices from a single
Original Equipment Manufacturer (OEM) to ease the
management and support burden because of large set of data
center network equipment.
• The networking equipment and software from a single vendor
with a feature set originally tailored for the enterprise LAN are
very expensive.

5
• The servers within the rack are interconnected using a top of
rack (ToR) switches containing a large number of 1Gb Ethernet
(1GbE) ports that may have been originally designed as an
Ethernet LAN workgroup switch.
• The ToR switches are connected to each other through an
Ethernet aggregation switch, similar to a LAN aggregation
switch designed to interconnect multiple LAN workgroup
switches.
• These ToR switch uplink ports to the aggregation switches
traditionally use multiple 1GbE ports in a link aggregation group
or 10GbE ports.

6
• Aggregation switches extend core switch/router to interconnect
multiple LANs on large corporate campuses.
• The core switch/routers are very expensive systems that have a
large number of layer 3 features.
• They are designed to support many of the TCP/IP standards
such as OSPF, MPLS, IP-in-IP, BGP, VPWS, VPLS and various
other IP tunneling protocols.
• This is why companies like Google decided to develop their own
data center networking equipment at a much lower cost.

7
Fig. 2: Traditional Enterprise Data Center 8
• The servers within the rack connect to a ToR switch using a star
topology as shown in Figure 1.
• Every server has a dedicated link to the ToR switch and the ToR
switch may forward data to other servers in the rack, or out of
the rack through high bandwidth uplink ports.
• Due to factors such as multiple VMs sharing the same network
connection, 10Gb Ethernet is widely used as the link between
the server shelf and the ToR switch in cloud data center
networks.

9
• A typical rack may contain on the order of 40 or more servers,
so many ToR switches contain up to 48 10GbE ports and four
40GbE uplink ports which are connected to aggregation
switches.
• Although this is a 3:1 mismatch in server bandwidth (480Gbps)
versus uplink bandwidth (160Gbps), data center traffic is very
bursty in nature and it is rare for all the servers to be using their
full allocation of 10Gbps bandwidth simultaneously.

10
2.1 Data center traffic patterns
• In data center, one request from user may also trigger pulling
information from a variety of servers and running specialized
programs based on user input.
• For example, a Google map request could spawn some of the
following transactions:
Determine the device type the client is using (laptop, handheld,
etc.)
Send information to a local search engine to match the map
address.
Based on the match, pull up the appropriate map data and send it
to the client based on device type.

11
 Based on the match and the user settings, pull up relative
nearby attractions to display on the map and send it to the
client.
Retrieve relative information about the client based on
recent web transaction.
Use this information to identify shopping history and send
targeted advertising to be displayed on the client’s web page.
• As you can see, a single web click can spawn multiple or
sometimes even hundreds of server-to-server transactions
within the data center.
12
• Two types of traffics are generated in a data center:
North-south traffic is generally between server and client,
East-west traffic is generated between servers and/or VMs.
East-west traffic is also increasing due to improvements in
business agility and continuity. Examples include VM
migration and storage replication.
• Traditional 3-teir data center networks do not do a good job
supporting east-west traffic due to factors such as poor
bandwidth distribution, high latency, and high latency variation
as indicated in Figure 3.
13
Fig. 3: Traffic flow in a data center 14
2.2 ToR switch features
• ToR switches are used within each server rack in large cloud
data centers.
• They have links to each server in the rack using a star fabric
architecture and high-bandwidth uplinks out of the rack.
• A typical rack can hold over 40 servers, which are connected to
the ToR switch using either 1Gb or 10Gb Ethernet links.
• Typical cloud data center ToR port configurations are 48 10GbE
ports using SFP+, plus four 40GbE QSFP+ ports for uplink
connectivity.

15
• Low latency
• High bandwidth
• Tunneling features
• Large table sizes
• Server virtualization
support
• Network convergence
• Open software Fig.4 :A simplified top view of a 1U ToR
switch

16
3 Data Center Design Consideration
Construction requirement
• Power is very expensive for the operations of data centers.
• According to the EPA’s report to congress on server and data
center energy, servers consume 80% of the total IT load and
40% of total data center power consumption.
• The rest of power is consumed by other devices like
transformers, distribution wiring, air conditioners, pumps, and
lighting.
• The power consumption of cooling equipment is important but it
is proportional to the IT power consumption.

17
• The data-center room has raised floors for hiding cables, power
lines, and cooling supplies.
• The cooling system is somewhat simpler than the power
system.
• The raised floor has a steel grid resting on stanchions about 2–
4 ft above the concrete floor.
• The under-floor area is often used to route power cables to
racks, but its primary use is to distribute cool air to the server
rack.

18
• The CRAC (computer room air conditioning) unit pressurizes
the raised floor plenum by blowing cold air into the plenum.
• The cold air escapes from the plenum through perforated tiles
that are placed in front of server racks.
• Racks are arranged in long aisles that alternate between cold
aisles and hot aisles to avoid mixing hot and cold air.
• The hot air produced by the servers circulates back to the
intakes of the CRAC units that cool it and then exhaust the cool
air into the raised floor plenum again.
• Typically, the incoming coolant is at 12–14°C and the warm
coolant returns to a chiller

19
Fig.5: The cooling system in a raised-floor data center with hot-cold air circulation supporting water heat exchange20facilities
3.1 Data-Center Interconnection Networks
• A critical core design of a data center is the interconnection
network among all servers in the data-center cluster.
• This network design must meet five special requirements: low
latency, high bandwidth, low cost, message-passing interface
(MPI) communication support, and fault tolerance.
• The design of an inter-server network must satisfy both point-to-
point and collective communication patterns among all server
nodes.

21
3.2 Specific internetwork design
considerations
• Application Traffic Support: The network topology should
support all MPI communication patterns.
• Network Expandability: With thousands or even hundreds of
thousands of server nodes, the cluster network interconnection
should be allowed to expand once more servers are added to
the data center
• Fault Tolerance and Graceful Degradation: The interconnection
network should provide some mechanism to tolerate link or
switch failures. In addition, multiple paths should be established
between any two server nodes in a data center
22
3.3 Switch-centric Data-Center Design
• There are two approaches to building data-center-scale
networks:
i. One is switch-centric
ii. Server-centric.
• In a switch-centric network, the switches are used to connect
the server nodes.
• The switch-centric design does not affect the server side. No
modifications to the servers are needed.
• The server-centric design does modify the operating system
running on the servers.

23
3.4 Power Measurement and Modeling in
Cloud
• Power consumption indicates the rate at which a machine can
perform its work.
• It is obtained by multiplying voltage and current while electrical
energy is the amount of power used over a period of time.
• The standard metric unit of power is the watt (W) and the
energy unit is watt-hour (Wh).

𝑃 = 𝐼𝑉 … … … … … … … … … 1
𝐸 = 𝑃𝑇 … … … … … … … … … .2
where, P is power consumption, I is current, V is voltage, E is energy
and T is time interval.

24
• Power direct measurement in Cloud can be achieved in data
centers that embed monitoring capabilities and probes such as
smart power distribution units (PDUs).
• For instance, power measurement introduced by Green Open
Cloud (GOC) framework relies on energy sensors to monitor the
electricity consumed by Cloud resources.
• The framework collects statistics of the power usage in real time
and embeds electrical sensors that provide dynamic
measurements of energy consumption and an energy data
collector.
25
• However, it is difficult to measure power consumption in virtual
machines (VMs).
• Models are normally used to estimate power consumptions in
VMs.
• The CPU of a server consumes the most important amount of
power.
• CPU-based linear models represent a lightweight and a simple
way to estimate servers’ power usage.

26
• Simple utilization-based power
models for servers are
proposed.
• They assume that CPU is the
only factor in their power
models and present an
approximation for total power
against CPU utilization (U).
𝑃 = 𝑃𝑖𝑑𝑙𝑒 + 𝑈(𝑃𝑝𝑒𝑎𝑘 𝑃𝑖𝑑𝑙𝑒 )

Figure 6: Server power model based on CPU utilization.

27
3.5 Power Modeling for VM Migration
• Virtual machine live migration (VMLM) consists of moving VM
between physical hosts without service interruption.
• This mechanism allows VM consolidation to achieve better
energy efficiency; however, it also brings additional power
consumption and its cost in terms of energy is not negligible.
• Energy costs of migration have not been considered when
migrating VMs.
• To investigate the energy cost of VMM and to model it, the
energy overhead of live migration depends essentially on the
amount of memory used by the VM and on the available
network bandwidth.
28
• The model is derived through linear regression and the
relationship between the energy cost of migration, the network
bandwidth and the VM size is expressed in equation 3.

𝐸𝑚𝑖𝑔 = 𝐴 + 𝐵𝑠 + 𝐶𝑏 … … … … … … … … 3
where s represents VMs size, b represents the network
bandwidth and A, B and C represent constant values.

29
3.6 Energy Efficiency Metrics
• In addition to power models, improving energy efficiency in
cloud data centers requires metrics that reflect data centers and
servers’ efficiency.
• and provide the necessary information for high level
management and scheduling decisions.
• Some metrics of energy efficiency have been proposed for data
centers.
• The Green Grid defined two data centers efficiency metrics:
power usage effectiveness (PUE) and data center efficiency
(DCE).
30
 𝑇𝑜𝑡𝑎𝑙𝐹𝑎𝑐𝑖𝑙𝑖𝑡𝑦𝑃𝑜𝑤𝑒𝑟
𝑃𝑈𝐸 =
𝐼𝑇𝐸𝑞𝑢𝑖𝑝𝑚𝑒𝑛𝑡𝑃𝑜𝑤𝑒𝑟

• DCE is the indicator ratio of IT data center energy efficiency and

is defined as the reciprocal of PUE.

1 𝐼𝑇𝐸𝑞𝑢𝑖𝑝𝑚𝑒𝑛𝑡𝑃𝑜𝑤𝑒𝑟
𝐷𝐶𝐸 = =
𝑃𝑈𝐸 𝑇𝑜𝑡𝑎𝑙𝐹𝑎𝑐𝑖𝑙𝑖𝑡𝑦𝑃𝑜𝑤𝑒𝑟

31
4. Cloud Management and Cloud Software
Deployment Considerations
• The cloud management consists of softwares that are used in
managing the cloud.
• The softwares can be a cloud operating system (OS), a
software that acts as an interface between the data center
(actual resources) and the user, or a management software that
allows managing resources.
• These softwares usually allow resource management
(scheduling, provisioning, etc.), optimization (server
consolidation, storage workload consolidation), and internal
cloud governance
32
• The SaaS delivery model is different from the traditional license-
based traditional software:
i. SaaS provides web access to commercial software on pay-as-you-
use basis.
ii. SaaS applications are developed, deployed, and managed from a
central location by the service provider.
iii. It allows the same instance of the SaaS application to be shared by
multiple customers or tenants.
iv. Updates are performed by the service provider, not by the user.
v. SaaS applications allow service integration with other third-party
services through application programming interfaces (APIs) provided33
• Some of the notable benefits of SaaS are mentioned in the
following:
i. Pay per use
ii. Zero infrastructure
iii. Automated updates
iv. Composite services
v. Dynamic scaling
vi. Green IT solutions

34
• The following are some applications where SaaS may not be
the best option:-
i. Some real-time applications where fast processing of data is needed

ii. Applications where the organization’s data are more confidential and
the organization does not want to host their data externally

iii. Applications where existing on-premise applications fulfill the

organization’s needs

35
• The following are examples where SaaS is the best option:-
i. For applications where the end user is looking for on-demand
software rather than full-term/licensing-based software
ii. For a start-up company that cannot invest more money on
buying licensed software
iii. For applications that need the accessibility from handheld
devices or thin clients
iv. For applications with unpredictable and dynamic load

36