and Scheduling Techniques Agenda for Today Resource management - - PowerPoint PPT Presentation

Day 5 Heuristic Resource Provisioning and Scheduling Techniques

Agenda for Today

• Resource management encompasses all the

characteristics and usage of cloud resources including resource provisioning and resource scheduling

• The resource allocation problem is a recurring issue in

distributed computing.

• It has remained a topic of research in various fields such

as

– job shop, flow shop or open shop scheduling in production environment, printed circuit board assembly scheduling – scheduling of tasks in distributed computing systems such as cluster, grid or cloud.

• We will look at heuristic approaches today..

Cloud Data Center Model

• The Cloud resource provisioning enables virtualized resources

to be allocated to Cloud consumers based on three provisioning plans.

J1 J1 Jn

Cloud consumer Virtualized Resources

request Result

Service Provisioning On-demand Advance Reservation Spot Instances

• Cloud services and basic functions are provided based on

Virtual Resources which are abstracted from Physical Resources.

resource provisioning resource allocation VM scheduling

Resource provisioning

J1 J1 Jn

Cloud consumer Virtualized Resources

request Result

Service Provisioning On-demand Advance Reservation Spot Instances

• Resource provisioning is responsible

– To understand the user needs – To prepare VMs with appropriate resources to match the workloads and QoS requirements – SLA Negotiation – Etc.

resource provisioning resource allocation VM scheduling

Resource Allocation

J1 J1 Jn

Cloud consumer Virtualized Resources

request Result

Service Provisioning On-demand Advance Reservation Spot Instances

• Resource allocation is responsible

– To select an optimal set of physical machines to host the received services (VMs), – To ensure the resource and QoS constraints are met. – To manage changes in resources availability through VMs restore or migration

resource provisioning resource allocation VM scheduling

VM Scheduling

J1 J1 Jn

Cloud consumer Virtualized Resources

request Result

Service Provisioning On-demand Advance Reservation Spot Instances

• VM scheduling is responsible

– A resource can be used either in shared (time sharing) mode or exclusive mode (space sharing). – To ensure that the VMs receive the required services based on the rule (policy) description.

resource provisioning resource allocation VM scheduling

Cloud resource provisioning problem

• The Cloud resource provisioning problem involves

tasks that must be scheduled on cloud resources subject to some constraints to optimize some

• bjective function.
• The growing scale of Cloud computing and the

increasing complexity of users’ requirements introduce additional constraints and make allocation decisions more challenging with difficult tradeoffs between user satisfaction and profit maximization.

Resource Provision and Allocation

• A wide variety of resource provisioning goals exist:

– High resource utilization – Energy efficiency – reliability of services – Low performance interference

• Optimally achieving the above goals in cloud

computing environment has been proved to be an NP-complete problem due to its combinatorial

• ptimization nature.
• There are no algorithms which may produce optimal

solution within polynomial time for such kind of problems.

Optimization algorithms

• Optimization algorithms can be roughly divided into two

categories: exact algorithms and heuristics.

• Exact algorithms are designed in such a way that it is

guaranteed that they will find the optimal solution in a finite amount of time.

– For scheduling optimization problems (e.g. NP-hard or global

• ptimization) this "finite amount of time" may increase

exponentially in respect to the dimensions of the problem.

• Heuristics do not have this guarantee, and therefore

generally return solutions that are worse than optimal.

– Heuristic algorithms usually find "good" solutions in a "reasonable" amount of time.

Heuristic algorithms

• Heuristic algorithms usually are

– adapted to the problem at hand and they try to take full advantage of the particularities of this problem. – However, because they are often too greedy, they usually get trapped in a local optimum and thus fail, in general, to

• btain the global optimum solution.

– Heuristic algorithms are generally make simplifying assumptions to relax constrains.

Relevant Information

• When developing resource provisioning algorithms, it

is imperative to consider information such as

• highly heterogeneous and time-varying workloads.

– daily demand distribution of a typical Internet application – request arrivals and departures statistics

https://gigaom.com/2012/02/11/which-is-less-expensive-amazon-or-self-hosted/, https://github.com/google/cluster-data/blob/master/ClusterData2011_2.md,

stochastic request arrivals and departures

Relevant Information

• Need to consider fundamental challenges that

arise from aspects such as

– the multi-tenant, – resource-abundant, – elastic resource model and – quality of service requirements expressed in terms

• f execution time and cost.
• How about cloud user’s job that often requires

cooperation among multiple VM instances?

Relevant Information

J1 J1 Jn

Cloud consumer

request

Service Provisioning On-demand Advance Reservation Spot Instances

What do we want to minimize:power consumption, SLA violation, etc.? What do we want to maximize: resource usage, revenue, etc.?

What do we want to Minimize: completion time, cost, etc. ?

What are the workload characteristics? What are the VM capability Cloud providers are faced with stochastic request arrivals and departures, which generates highly heterogeneous and time-varying workloads.

Relevant Information

• What type of resources the application needs?

– VM that is abundant in a particular type of resource: bandwidth, CPU, or storage.

• VMs types oblivious Cloud resource

provisioning is most common.

VM type-oblivious Heuristic Algorithms

• This question usually not considered in the heuristic

algorithm as it is usually assumed that

– A single type of VMs is usable by all applications, or – VMs are substitutes in that a high-end VM is equivalent to a number of low-end VMs. For example, a Type II (2 × Core, 2 GB RAM, 40 GB Disk) VM equals two Type I (1 × Core, 1 GB RAM, 20 GB Disk) VMs.

• Such type-oblivious VM heuristic algorithm do not

handle the existing VM heterogeneity in today’s cloud computing platforms.

Resource provisioning classification

• Resource provisioning can be generally classified as

– Static Resource Provisioning

• Given information such as highly heterogeneous and time-varying

workloads , allocation solutions based on static resource

provisioning lead to poor performance and hinder providers from achieving expected profits. – Dynamic Resource Provisioning

• With dynamic provisioning, the provider allocates more

resources as they are needed and removes them when they are not. – Hybrid Resource Provisioning

• Combines other resource provisioning approaches

Plan-oblivious Heuristic Algorithms

• Cloud providers can offer cloud consumers three

different provisioning plans for computing resources, namely reservation and on‐demand plans.

• Cloud resource provisioning that are oblivious to

various plans offered by the service provider is common

– Can we run FCFS on shared systems following the on- demand plan? – How about system with advanced reservations?

Over provisioning problem

• Over-provisioning to handle demand peaks

can result in significant costs and unused capacities

Resource waste

Under provisioning problem

• Planning resources for only usual workloads

request rejection and QoS degradation

applicat ion

QoS violation

Example of Heuristic Algorithms

• Rule-based heuristics

– First Come First Serve (FCFS), – Minimum Completion Time (MCT), – Minimum Execution Time (MET), – Max-min, – Min-min and – Sufferage.

Question

• How do I compare my algorithm with existing

heuristic algorithms for tasks scheduling?

• It is difficult to compare heuristic algorithms

due to the contrasting underlying assumption by each heuristic algorithm.

• The results of the simulation for all the

heuristics are considered under the same assumption.

Performance Implications of VM Co-locations

• Independent VMs can share a

physical machine.

• The extent of interference impact

depends on the type of application

• The analysis of co-located data-

intensive applications [Chiang and Huang] shows that the effects for such applications is more pronounced than others

Task 2 Task 4 VM 1 Task 3 Task 4 VM 2 Hypervisor Virtualization Processor Mem Disk Co-hosted VMs share hardware and software

Data Intensive Application Interference

• Interference slows down the tasks dramatically
• Three approaches for interference prediction model

is proposed.

– Linear Model – Quadratic Model – Exponential Model

• Also, an Interference-Aware Task Scheduling

approach is proposed

TRACON: Interference-Aware Scheduling for Data-Intensive Applications in Virtualized Environments. SC’11 Interference and Locality-Aware Task Scheduling for MapReduce Applications in Virtual Clusters HPDC’ 13

Reinforcement Learning Method

• Other methods that can be classified as Optimizes

long-term reward

• Current decision may have delayed consequences on

both future reward and future state.

• Avoid Local optimum: mathematical optimization
• Q-Learning

System Agent

Adjustm ent of resource

state feedback

• A variety of learning

methods exits including the following:

• Optimal control,

feedback control

• Statistical Modeling

25

Thank you.

Questions, Comments, …?