Cloud-Edge Computing: Resource Allocation

• Dr. Rajiv Misra is a Professor in the Dept. of Computer Science & Engg. at the Indian Institute of Technology Patna.
• Collaborative cloud-edge approaches can yield better performance and efficiency compared to solely cloud or edge approaches.
• Resource allocation strategies can be customized for specific use cases and evolve with user demand and network conditions.

The Collaborative Cloud-Edge Environment

• The "user-edge-cloud" model distributes computing resources across user devices, edge nodes, and cloud servers.
• Optimizing system performance and ensuring efficient resource use are key goals of resource allocation.
• Collaborative cloud-edge approaches can be more effective compared to approaches that focus solely on cloud or edge resources.
• Cloud services include private and public clouds.
• Private clouds dedicate resources to a single org to enable greater control and security.
• Public clouds are shared by multiple orgs for greater control and scalability.
• Edge nodes consist of local computing resources closer to the end user than cloud nodes enabling reduced latency.
• Edge nodes offer low-latency, high-bandwidth services and offload some processing from the main cloud.
• Cloud resource allocation can be based on user demand, network conditions, and available resources.
• Machine learning algorithms optimize resource allocation over time in collaborative cloud-edge approaches.
• Load balancing, task offloading, and caching are common resource allocation techniques for cloud and edge resources.
• In multi-edge-node scenarios, cloud and edge nodes must coordinate to allocate resources effectively.
• Communication protocols and data sharing are enablers for effective coordination in cloud-edge environments.

Public vs. Private Cloud Environments

• In public clouds, the cloud provider offers different pricing models for cloud services based on demand.
• Pricing models also have different cost structures affecting resource allocation strategies.
• Amazon, Microsoft, and Alicloud offer three different pricing models, and each with distinct cost structures.
• Edge nodes should select the appropriate pricing model and user allocation to rented or own VMs.
• In private clouds, edge nodes have their own virtual machines (VMs) to process user demands.
• VMs can be rented from the cloud node to scale up if the edge node's capacity is exceeded.
• The cost of the private cloud fluctuates dynamically with its physical computing cost, so resource allocation should reflect this.
• Computing costs can be calculated and utilized after allocation to process new computing tasks in the next time slot.

User Settings

• Time is discretized into T time slots in this context.
• For each time slot t, user demand is defined as Dt, comprised of dt and lt.
• dt is the number of VMs requested and lt is the computing time needed.

Computing Resources and Cost of Edge Nodes

• Variable E represents the total computing resources owned by the edge node.
• Variable et is the number of remaining VMs of the edge node in time slot t, as resources are allocated to the end users.
• The variable d^e_t expresses the number of VMs the edge node provided.
• The number of VMs provided by the cloud node is expressed as df_t.
• Edge nodes without available resources will hand over the computing tasks to the cloud service for processing rather than queuing.
• Number of VMs the edge node provide from the equation d^e_t = dt - e_t, e_t ≥ 0, or 0 otherwise.
• An allocation record, h_t = (d^e_t, l_t), is generated for each processed demand when resource allocation is conducted successfully.
• If allocation is successful, an allocation list is created denoted as h = <h1, h2,...hm>.
• At the end of each time slot the edge node traverses the list and reduces remaining computing time of each record by one.
• Completed records corresponding to records that reach 0 remaining computing time duration release allocated VMs and are removed.
• The number of VMs waiting to be released at the end of time spot t is denoted as _n_t.
• n_t = sum from 1 to m ( d^e_i )
• The constraint also expresses l_i=0 where h_i ϵ H.
• The remaining VMs at time slot t+1 are calculated from the number of VMs at the beginning of time slot t:
  • et+1 = et - d^e_t + n_t
• The cost of the edge node in time slot t as the sum of standby cost (et * pe) plus computing cost ((E – et)*pf).
  • C^e_t = et * pe + (E - _e_t)*pf

Cost of Collaborative Cloud-Side Computing

• The cost of collaborative cloud-edge in a private environment:
  • c^pri_t = df_t * pc + C^e_t
• df_t represents number of VMs cloud provides, and pc is the unit cost of VMs in private cloud.
• The collaborative cloud-edge cost in a public environment:
  • c^pub_t =X₁p_odd^f_t _+_X₂p_upfrontd^f*_t + X₃p_red^f_t + X₄p_td^f_t + C^e_t
• Where X_i ϵ {0,1}, to represent if server is being used
• p_od, represents cost of on demand instances
• p_upfront + p_re cost of reserved instances plus cost of reserved instance + customization price of reserved instances.
• p_t cost of Spot Instances.

Goal

• Time is split into T time slots, and the user submits the demand at the beginning of each slot.
• The edge node allocates on resource allocation strategy.
• The type of cloud service in public cloud depends the cost, corresponding with other factors.
• The cost of current time slot t, is labelled as Ct, is calculated with the allocation, also the price.
• Algorithm goal is to minimize the long-term cost.

Resource Allocation Algorithms: Markov Decision Process

• Making resource allocation a sequential decision-making problem can be implemented through a Markov decision process.
• Markov Process decision process is represented as (S, A, P, r, γ).
  • "S" is the finite set of states,
  • "A" the finite set of actions,
  • "P" the state transition
  • "r" the immediate award
  • "γ" the discount factor.
• The tuple St = (et , nt-1 , Dt , pt ) ∈ S describes the state of each edge node.
• Component et shows the remaining VMs of the edge node at t, also nt shows number of VMs returned
• The reward is given by Dt and a unit VMs cost.
• Action (k_e , x_k) where k=1 is a private cloud model and function reward is a negative value = -C_private cloud.

Parameterized Action Markov Decision Process (PAMDP)

• Edge nodes first determine the pricing mode, then resource segmentation.
• Resource allocation described by parametric action.
• PAMDP has to be implemented to have sequential parameterized actions.
• A PAMDP is a tuple (S, A, P, r, γ).
• Different is finite set of parameterized actions.
• The state is s = (e_t , n_t-1 Pt, εt )

• The reward function can be described and parameterized the edge node to run.

• Cost efficient resource allocation with private cloud) Algorithm: - In this the Actor implements, Actor main, parameters also implemeents the reward function that is obtained. - Then user task is performed and actions according status from the private cloud.

Resource Allocation Based on Deep Deterministic Policy Gradient

• A DDPG algorithm represents the actor rating
• Actor generates actions, while the critic evaluates the action and generates.
• The Actor is used according function equation to follow:
• α = πρ(S) + N.
• For loss function:
  • 𝛻𝐽(𝜔) = 𝑚(𝑦𝑗 − 𝑄(𝜙(𝑠),𝑎𝑗,𝜔)²

Resource Allocation Algorithms

• DDPG structure is shown in figure.
• The users algorithm contains user requests, demands.
• The edge node, and the state of the cloud are obtained iterative.
• The number of demands and notes are used from private sector.
• New methods, interaction environment is stored in experiencal
• CERAI sample, Loss Functions, ETC.
• A training will be done according ensure continuation of resource.

Cost Efficient Resource Allocation Algorithm

• Cost effective resource allocation is done using following steps
  • Inititalize Actor and critic network.
  • Select random, exploration parameters.

Resource Allocation Based on P-DQN

• Action can ∈ A where the parameter space
• Function that is used: Q
• DQN uses the Q function.
• P and DQN also estimates value where to use.

AWS Reference Model

• AWS services are organized into a 5-layered reference model.
• AWS is globally present with the datacentres present in different available regions.
• Infrastructure is present with edge locations. Regions and availability zones are a key part of the underlying global infrastructure.
• The regions themselves are made up of availability zones, each designed as independent regions.
• Support is a reliable way to explain AWS Cloud

AWS Global Datacenter

• Some key facts about AWS global infrastructure:
  • 20 regions with 5 more coming soon
  • 61 availability zone
  • 158 edge locations
  • 11 regional caches
  • 130 services

Regions

• Region as geographic area that offers variety of AWS functions.
• Customers can select specific regions for services.
• Eleven regions exist globally.

AWS Region

• Out of 13, 1 of the major cloud services, is Redundant with transit paths
• Also, the Transit Centers Connect Through Private Links between Regions.
• It has private links that exist direct through the internet also with an area between
• AZ area usually always has a peak.

Example AWS Data Center

• Single DC typically over 50,000 , Larger DCs undesirable.
• Up to 102 Tbps single DC aws Custom network equipment

AWS Security

• AWS has shared security between customers
• Customers have applications between zones -also aws has Global Infrastructure, networking, etc

Virtual Private Cloud (VPC)

• VPC = logical Section from the Aws cloud, virtual network
• When launching to get Customer, All have vpcs

Networking

• Dedicated connections to vpc through secure connection

Region and Avalailability Zone

Why Availability Zones

AWS Account, Userrs and Service Scope

AWS Compute and Anylitics, Storage and Database Services, Network and Management Services, Application and Development Services

Development and Test Enviroments

Big Data

• Fully managed and simple
• Easily launch customize etc or Amazon

High Performance Computing HPC

• Amazon can have SSD platform and Network Performance

Storage Backup etc

Disaster Recovery

• Disaster backup system is with AWS
• On premises infrastructure, AWS provides multi-site solutions.

AppStream

• An app based stream with High quality user streaming

Facebook based services

• Facebook is implemented based in application to run

CERAU

• is based in P-DQN
• Also contains information or resource demands
• Has exploration parameters that implement.