Day 7 Economic-based Cloud Resource Provisioning Introduction The - - PDF document

2/6/2018 1

Day 7 Economic-based Cloud Resource Provisioning Introduction

• The performance of a distributed system often

depends on the maximum load of any of the machines.

• Modern datacenters employ server virtualization

and consolidation to reduce the cost of operation and to maximize profit.

• From time to time, cloud service providers find

themselves in a situation where performance becomes an issue.

Introduction

• We discuss two possible ways to address this

situation

– Reassignment approach

• The higher the maximum load, the longer the execution time of

the entire system.

• A good load balancing schemes are crucial for efficient

computations on distributed systems.

– Leasing resources from alternative datacenters owned by federated providers.

2/6/2018 2

Cloud Computing

• Cloud computing service providers offer computing

resources as utility

• Resources are charged by its type and duration

J1 J1 Jn

Cloud consumer

request

Cloud service provider

• The cost is computed as the fixed charge multiplied by the

number of resource instances or service units requested by the consumer.

Cloud Service Users

• Cloud users pay providers for cloud resource usage.
• For example, let ℛ = 𝑠

, ⋯ , 𝑠 be the resources used by

Cloud users.

𝑉 = 𝑑 ∙ 𝑢

∈ℛ

– 𝑑 denotes the cost of resource 𝑠 per unit time and 𝑢 denotes the time for which resource 𝑠 is utilized.

• For example, if the cost of using resource r1 is 2000 per time

unit and r2 is 3000 per time unit, where r1 is used for 22 time unit and r2 is used 19 time unit

• 𝑉 = 2000 ∙ 22 + 3000 ∙ 19 = 101,000

Cloud Service Users

• Cloud users expect Quality-of-service guarantees

from the cloud service providers.

• QoS parameters indicate the ability of a service to

meet certain requirements for different aspects of the service

• For example

– Deadline constraint: This represents the time till which the task or the batch of tasks should be finished. – Budget constraint: This represents the restriction on the total cost of executing all tasks. – Cost: resources are provisioned in a cost-efficient way

2/6/2018 3

Cloud User Objectives

• User objective can be stated as maximizing QoS such

as minimize the usage cost while satisfying the performance requirement of the applications

• For example, the objective can be to determine a

configuration where all tasks of a user are executed at a minimal cost.

Minimize ∑ 𝑑 ∙ 𝑢

∈ℛ

• QoS requirements are commonly formalized in the

form of SLAs

Cloud Service Provider

• Cloud providers need to efficiently manage resources to

achieve the performance of their applications and improve the utilization of reserved resources, thereby minimizing the usage cost.

• Some objectives

– Revenue maximization is one of the objectives of cloud computing service provider. – Resource utilization maximization is another objectives.

– This must be done in measured manner is it can lead to system overload.

• Research question 1: An approach that predicts future

resource requirements for workload. These predictive approaches can lead to better resource efficiency.

Cloud Service Provider

• Cloud providers must achieve certain level of QoS
• For instance

– For budget constraint jobs the cost of performing these jobs cannot exceed a certain budget constraint

• Therefore, one of the challenges facing service

provider is how to complete jobs with unpredictable submission time under budget and deadline constrains

• Research problem: Most of existing work do not

consider unpredictable submission time of jobs, as well as budget and deadline constrains simultaneously.

2/6/2018 4

Service Level Agreements

• Service Level Agreements (SLAs) form an important

component of the contractual relationship between a cloud customer and a cloud service provider

J1 J1 Jn

Cloud consumer

request

Cloud service provider Jemal H. Abawajy, Mohd Farhan Md Fudzee, Mohammad Mehedi Hassan, Majed A. AlRubaian: Service level agreement management framework for utility-oriented computing platforms. The Journal of Supercomputing 71(11): 4287-4303 (2015)

• Research question 2: different cloud services and deployment

models will require different approaches to SLAs.

• Research question 3: The global nature of Cloud computing

renders SLAs to cross several jurisdictions. Frameworks for handling data privacy is an important addition to cloud computing.

Penalties for Missing SLA

• Total economical penalties for SLA violations the sum
• f the total proportional penalties costs for

unsatisfied demand of resources.

• Total economical penalties at instant 𝑢 (𝑄);

𝑄 = 𝑆 ∙ ∆ ∙ 𝜚

• – 𝑠: Number of resources assigned to the Cloud user. For example, if the

user is assigned Disk, CPU, RAM memory and network capacity, then 𝑠 = 4. – 𝑆: Revenue for completing the assigned virtual machine; – 𝜚: Penalty factor for resource k, where 𝜚 ≥ 1; – 𝑛 ∶Number of VMs

Cloud Resource Provisioning algorithms

• A variety of cloud resource provisioning has been

developed

– Budget aware Provisioning algorithms

• The objective is to complete jobs such that the cost of performing

these jobs cannot exceed a certain budget constraint

– Cost-aware Provisioning algorithms

• The objective is to satisfy the SLA requirements of the user while

minimizing the total cost.

2/6/2018 5

Cloud Resource Provisioning algorithms

• SLA-aware Provisioning - The objective is to satisfy the SLA

requirements of the user while minimizing the total cost.

• In fact, the SLA requirements consist of Budget and

Deadline.

– As a result, SLA-aware algorithms’ objective can be stated as follows:

Minimize 𝑈𝐷 = 𝑑 ∙ 𝑢

∈ℛ

• Subject to
• 𝑈𝑝𝑢𝑏𝑚 𝐷𝑝𝑡𝑢 ≤ 𝑐𝑣𝑒𝑕𝑓𝑢
• 𝑓𝑦𝑓𝑑𝑣𝑢𝑗𝑝𝑜 𝑢𝑗𝑛𝑓 ≤ 𝑒𝑓𝑏𝑒𝑚𝑗𝑜𝑓

VM allocation optimization

• Cloud service providers dynamically receive

– requests for the placement of cloud services – The requests have different characteristics according to different dynamic parameters.

• It is well known that online decisions made along the
• peration of a dynamic cloud computing infrastructure needs

to be dynamically optimized to avoid negative affects

• A common approach for optimization of the current VM

allocation is through adjusting resource allocation according to demand in order to satisfy SLA.

• It is worth noting that this problem is a multi-objective optimization

problem including guaranteeing service quality and maximizing resource utilization with multiple resource constraints.

Dynamic VM Placement Rebalancing

• The main idea is to proactively monitor the load of

the physical machines and dynamically rebalance the load.

– Some VMs were migrated from the overloaded physical machine to achieve the goals of performing load balancing, improving QoS and degrading the risk of

• verloading the CPU resource.

– The migrated VMs were redeployed in the normal PMs.

• This requires a number of algorithms including

identifying overload host detection

2/6/2018 6

Dynamic VM Placement Rebalancing

• There are several key problems which should be

dealt with:

– when a host is supposed to be heavily loaded, where some VMs from the host should be migrated to another host; – when a host is believed to be moderately loaded or lightly loaded, resulting in a decision to keep all VMs on this host unchanged; – when a host is believed to be little-loaded, where all VMs

• n the host must be migrated to another host;

– selecting a VM or more VMs that should be migrated from the heavily loaded; finding a new host to accommodate VMs migrated from heavily loaded or little-loaded hosts.

Overload host detection

• Overbooked resources may lead to

Quality of Service (QoS) degradation, and consequently Service Level Agreement (SLA) violations with economical penalties.

• This economical penalties should

be minimized for an economical revenue maximization

Threshold-based Heuristics

• A variety of threshold-based heuristics have

been advanced to handle

– Single Threshold (ST) [Buyya, et. al] – Double Thresholds (DT) algorithm [Beloglazov and Buyya] – Three Thresholds algorithm [Abawajy, et al]

2/6/2018 7

Threshold-based Heuristics

• Buyya et al. proposed a method named Single

Threshold (ST) algorithm.

– ST sets a utilization threshold to keep all hosts’ CPU utilization in data centers below this value, consequently control the migration of VM. – The aim is to preserve free resources to prevent SLA violation due to consolidation in cases when the resource demand by VMs increases. – The static threshold method or fixed values of thresholds are not suitable for dynamic workloads, as they do not adapt to the changes in workload

Double Thresholds (DT)

• Double Thresholds (DT) algorithm,

– sets two thresholds (one is low utilization threshold and the other is upper utilization threshold) to control all hosts’ CPU utilization between the two thresholds. – This essentially divides into underloaded, normal and overloaded

• If the CPU utilization of a host exceeds the upper

threshold, some VMs have to be migrated from the host to reduce the utilization in order to prevent SLA violations.

Adaptive three-threshold framework

• We divide the hosts into classes based on the CPU

utilization.

Zhou Zhou a,b, Jemal Abawajy c,*, Morshed Chowdhury d, Zhigang Hue, Keqin Li f, Hongbing Cheng g, Abdulhameed A. Alelaiwi h, Fangmin Li a, Minimizing SLA violation and power consumption in Cloud data centers using adaptive energy- aware algorithms, Future Generation Computing, 201

How can we determine threshold values?

2/6/2018 8

Determining threshold values

• To determine the value of the three thresholds (Ta,

Tb, and Tc ), we use a novel algorithm which refer to as a KMI (K-Means clustering algorithm-Midrange- Interquartile range).

CPU utilization for all hosts

Overload host detection

• Research problem: random workload is common, and

the host overload detection method, which is based

• n a fixed threshold, is incapable of adjusting the

reserved idle resources according to the uncertain workload, which hinders the use of the VM consolidation method to properly allocate resources and causes undesired situations, such as poor service performance and high energy consumption.

Load selection

• Let 𝑠 be the total remaining amount of computation

required for processing task j on computing node i. Then, determining the most remaining load is done by simply evaluating max{Ri,j}, 𝑢 ← max

∈,∈ 𝑠

• Research question: It is to be noted that since cloud resource

configuration is not the same when moving from one cloud to the other, the computation requirement of a load is different

• n different computing nodes. The challenge is to develop a

prediction algorithm that allows a computing node to estimate the computation requirement of a load.

2/6/2018 9

VM Selection

• One of the problems to be addressed the
• ptimization of the current VM allocation problem is

to answer which VM should be moved from the host.

• This step is very important as migrating a wrong VM

can increase SLA violation rate

• Possible approach include

– Selecting a VM that has received the most service – Selecting the host that has not yet received any service – ..

The minimization of migrations policy

• The MM policy selects the minimum number
• f VMs needed to migrate from a host to

lower the CPU utilization below the upper utilization threshold if the upper threshold is violated.

• The selection is based on the fraction of the

CPU utilization allocated to the VM.

The highest potential growth policy

• The policy migrates VMs that have

– the lowest usage of the CPU relatively to the CPU capacity defined by the VM parameters (ie, the fraction of the CPU capacity initially requested for the VM)

2/6/2018 10

The random choice policy

• The Random Choice (RC) policy relies on a

random selection of a number of VMs needed to decrease the CPU utilization by a host below the upper utilization threshold.

VM Migration

• VM migration is an effective method to adjusting resource

allocation according to demand in order to satisfy SLA.

• Migration process is the mechanism by which, virtual

machine’s memory pages and states are transferred to a destination physical machine. Different virualization technique, use different migration mechanism.

Thrashing of VM migration

• Thrashing of VM migration a real issue in live

migration

– It means a VM frequently migrated from a host to another host, thus increasing SLA violations.

• Possible solution

– The number of the time that a VM migrates should be controlled

2/6/2018 11

Workload-aware VM Reallocation

• The ready time of nodes depends on the processing time of

loads, which in turn depends on the capacity of computing nodes and the nature of loads.

• The processing time of a workload may vary on different
• nodes. Workloads may require different amount of computations.

– For instance, one load deals with compression and the other deals with encryption using different key lengths, resulting in different amount of running time.

• Research problem: Existing approaches tend to be workload-

agonistic and will not in most cases produce reallocation favorable to the workload. The challenge is to develop workload-aware reallocation approach and compare it with workload-agonistic approaches

Research Question

• Research Problem: The algorithm we developed and PSO

approach can be integrated in order to decreases both the number of active hosts and VMs migration.

• Research Problem: We need to develop locality-aware

dynamic VM migration.

• Research Problem: Develop VM-self decision protocols where

a VM running on an overloaded physical machine decides autonomously whether or not to migrate to a neighboring resource.

• Research Problem: A learning automata-based load

rebalancing mechanism algorithm for in order to decreases both the number of active hosts and VMs migration.

Cost of VM Migration

• How do we compute the cost of migration?

– Proper VM migration between servers can reduce an SLA violation in data centers. – However, excessive VM migration could bring negative impact on performance of application which runs on the VMs.

• Research question: It is to be noted that since cloud resource

configuration is not the same when moving from one cloud to the other, the computation requirement of a load is different

• n different computing nodes. The challenge is to develop a

prediction algorithm that allows a computing node to estimate the computation requirement of a load.

2/6/2018 12

VM live migration

• Live migration of VMs allows transferring a VM

between physical nodes without suspension and with a short downtime.

• The time required for VM live migration (𝑈)

depends on the total amount of memory used by the VM and the available network bandwidth as follows:

𝑈 = 𝑡𝑢𝑏𝑠𝑢𝑣𝑞 + 𝑛𝑓𝑛𝑝𝑠𝑧 𝑡𝑗𝑨𝑓 𝑐𝑏𝑜𝑒𝑥𝑗𝑒𝑢ℎ

VM live migration cost

• Live migration has a negative impact on performance of

applications running in a VM during a migration.

• The total performance degradation (𝑉) a given 𝑊𝑁

will

experience can be estimated as follows: 𝑉 = 0.1 ∗ 𝑣 𝑢 𝑒𝑢

• 𝑣(𝑢): the CPU utilization by 𝑊𝑁

,

• Each VM migration may cause an SLA violation; therefore, it is

crucial to minimize the number of VM migrations.

• Research problem: The behavior of applications (e.g, number
• f memory pages that the application updates during its

execution) have direct impact on performance degradation and downtime. Research is needed to quantify this impact

SLA Violation

• Service level agreement violation (SLAV) can be computed

as follows:

𝑇𝑀𝐵𝑊 = 𝑃𝑈𝐺 ∙ 𝑄𝐸𝑁

• Both the OTF and PDM metrics independently characterize

the level of SLA violations in the system;

• The overall performance degradation by VMs due to

migrations, performance degradation due to migrations (PDM).

𝑄𝐸𝑁 = 1 𝑛 𝐷 𝐷

• – 𝐷: the estimate of 𝑊𝑁 performance degradation due to migrations;

– 𝐷: is the total CPU capacity 𝑊𝑁 requested during its lifetime. – In this work, Cdj is estimated to be 10% of the CPU utilization in MIPS during al

2/6/2018 13

SLA violation metrics

– The fraction of time during which active PMs have experienced 100% CPU utilization, overload time fraction (OTF)

𝑃𝑈𝐺 = 1 𝑜 𝑈

• 𝑈
• 𝑜 is the number of PMs;
• 𝑈

is the total time during which the PMi has

experienced 100% utilization leading to an SLA violation;

• 𝑈

is the total of the PM that is in the active state

(serving VMs);

Federated Cloud

• Federated cloud brings a different challenges

to the resource provisioning

Leasing Resources

• One way a cloud service provider may support requested

resources that are not able to be provide (e.g., a workload peak) is by leasing resources from alternative datacenters

• wned by federated providers.

Service Provisioning Service Provisioning leasing request

• Should the service providers charge each other lower prices

than offered to customers in the actual cloud market?

2/6/2018 14

Leasing Resources

• The pricing scheme may depend on the prior agreement

between the two cloud service providers .

Service Provisioning Service Provisioning leasing request

• At each discrete time instant 𝑢, the total leasing costs (𝑀𝐷)

that a service provider incurs can be represented as follows:

𝑀𝐷 = 𝑠

∙ 𝑦

• – 𝑠

: Economical revenue for attending Vj in [$] at instant t;

– 𝑛: Number of VMs – 𝑦: Indicates if Vj is allocated for execution locally (𝑦= 0) or externally (𝑦= 1)

Leasing Resources

• One way a cloud service provider may support requested

resources that are not able to be provided (e.g., a workload peak) is by leasing resources from alternative datacenters

• wned by federated providers.

Service Provisioning Service Provisioning leasing request

• Research question: An approach that allow a cloud service

provider to lease resources from another cloud service provider that fulfills:

– It transparently leases low-price resources from alternative providers. – This leasing costs should be minimized in order to maximize economical revenue objective function.

SLA Violation

2/6/2018 15

Minimizing Migration

44

Thank you.

Questions, Comments, …?