Virtual Machine Migration

Virtual Machine Migration

• Migration moves VM instances across various physical hosts and ensures minimal downtime for running services.
• Services remain oblivious to migration while maintaining network connections, and VMs are treated as black boxes, keeping their internal state unchanged.

Types of VM Migration

Online (Live) Migration

• VM remains operational during the migration.
• Common reasons include host maintenance, balancing workloads, optimizing resource utilization, and it requires sufficient network bandwidth.
• Use cases include vacating a host, targeting a particular VM, balancing workloads, and optimizing resource utilization.

Offline Migration

• VM is shut down before migration.
• Requires fewer resources like CPU and memory.
• Necessary when live migration isn't supported, such as with different processor types.
• Use cases include less resource usage, different processor types, unsupported online migration, or stopped VMs requiring configuration information transfer.

Design Challenges

• Minimizing service downtime.
• Reducing migration duration.
• Avoiding disruptions to active applications.

Live Migration Techniques

Pre-copy Migration

• Memory pages are copied while the VM is running.
• A final stop-and-copy phase ensures data consistency. Steps:
  • Memory pages are copied to the destination while the VM continues running.
  • The process repeats for modified pages.
  • VM is briefly paused while the last memory pages are transferred.
• Pros: Reduces application performance impact.
• Cons: Requires multiple iterations, increases migration time.

Post-copy (Lazy) Migration

• VM execution starts at the destination before all memory is copied.
• Pages are fetched on demand from the source using on-demand memory fetching. Steps:
  • The VM is stopped, and non-memory states like CPU and disk are moved.
  • Execution starts on the new machine.
  • On-demand memory fetching: If a page is needed but missing, it is copied from the source.
  • Background copying ensures all memory pages are eventually transferred.
• Pros: Faster migration.
• Cons: High risk of failure if network issues occur.

Hybrid Migration

• Aims to balance speed, reliability, and resource efficiency.
• Combines pre-copy and post-copy techniques.

Challenges in VM Migration

Traffic Contention

• Migration traffic competes with normal network traffic.
• Can cause congestion affecting application performance.

Residual Dependencies

• Ensures no lingering dependencies on the original host after migration.
• Original host should be able to power down safely.

Solutions for Traffic Contention

• Limit Migration Traffic: Restrict the amount of data transferred per iteration.
• Control Transfer Rate: Balance bandwidth usage between migration and applications.
• Traffic-Sensitive Migration: Selects migration methods based on current network load.

Future Prospects

• Improve traffic-sensitive migration for resource allocation.
• Enhance automation for destination host selection.
• Reduce migration overhead while maintaining service quality.
• The time required to transfer a VM depends on factors like RAM, working set size, and the write rate.

Introduction to Scheduling in Cloud and Fog Computing

• Example scheduling goals should tasks not complete successfully include timeout and running out of memory.
• Scheduling on a real-time system must be scalable.
• An effective scheduler can reduce operational costs, reduce queue waiting time and increase resource utilization.

Cloud and Fog Computing Overview

Cloud Computing Overview

• Provides scalable, on-demand computing power through virtualized resources, operates on a pay-as-you-go model, adhering to Service Level Agreements (SLAs). Cloud service models can include:
  • Software as a Service (SaaS): Fully managed applications.
  • Platform as a Service (PaaS): A development platform for applications.
  • Infrastructure as a Service (IaaS): Virtualized computing resources.

Fog Computing Overview

• Decentralized computing model where computation occurs closer to the data source, reduces latency compared to cloud computing, but introduces complex scheduling challenges.

Scheduling Goals

• Manage performance and QoS, optimize CPU and memory utilization, maximize resource utilization while minimizing task execution time, improve fairness in task allocation, increase the number of successfully completed tasks, enable real-time task scheduling, achieve high system throughput, and improve load balancing.

Different Types of Task Scheduling

• Immediate Scheduling: Tasks are assigned to Virtual Machines (VMs) when they arrive.
• Batch Scheduling: Tasks are grouped into batches before being processed.
• Static Scheduling: Based on prior knowledge of the system's global state, does not consider current VM states, typically uses Round Robin (RR) or random scheduling algorithms.
• Dynamic Scheduling: Considers real-time VM states for decision-making, adjusts task assignments based on current system load.
• Preemptive Scheduling: Tasks can be interrupted and reassigned to another machine mid-execution.
• Non-Preemptive Scheduling: Once a task starts, it cannot be reallocated until it completes.

Job Scheduling

• Job Shop Scheduling Problem (JSSP) is a key issue in cloud computing and manufacturing.
• The 10x10 JSSP (10 jobs, 10 machines) assigns jobs to machines optimally.
• It also represents job allocation over time showing execution of jobs on different machines, machine utilization, and scheduling gaps/overlaps.
• The goal is to minimize makespan (total execution time) while maximizing resource efficiency.

System Architecture and Task Scheduling Approach

Fog Computing System Architecture

• Fog Broker: Central control unit managing tasks between cloud nodes (CN) and fog nodes (FN).
• Cloud Layer: Cloud nodes responsible for heavy computation and long-term storage.
• Fog Layer: Handles local, low-latency processing for real-time applications.
• IoT Devices: Wearable devices, smart homes, and healthcare devices requiring computation.
• Task Scheduler & Resource Monitoring: Allocates tasks based on available resources and priority.
• Request Receiver: Collects and processes incoming tasks.

Task Scheduling Approach

• Fog Broker: Receives tasks from users and assigns them to either fog nodes (FN) or cloud nodes (CN).
• Task Distribution: Tasks are labeled as T1, T2, ..., Tn.
  • Cloud Nodes (CN1, CN2): Handle complex, high-latency tasks.
  • Fog Nodes (FN1, FN2, FN3): Process tasks with lower latency and real-time requirements.
• Goal: Efficient scheduling by minimizing computation, communication, and violation costs while maintaining QoS.

Task Scheduling Challenges and Execution Timing

• The scheduler is responsible for assigning tasks to available machines.
• Machines can be in two states
  • Busy: Actively processing tasks.
  • Free: Available for task allocation.
• Efficient scheduling ensures free machines are utilized while minimizing idle time.
• Queue Waiting Time: The time a task spends waiting before assignment to a machine.
• Execution Time: The time taken by a machine to complete a task.
• Relaxed Execution Time: A flexible execution period, allowing minor delays while ensuring task completion within acceptable limits.
• Goal: Reduce queue waiting time while maximizing execution efficiency.

Min-CCV Algorithm

• Aims to ensure efficiency by minimizing three key costs: computation, communication, and violation.

Process Overview

• Tasks are assigned to the node with the lowest total costs.
• Each node's available time is updated after task allocation, and ensures balance between computational efficiency, communication overhead, and deadline adherence.

Min-V Algorithm

• Min-V focuses primarily on minimizing deadline violations, tasks are sorted based on the earliest deadline first.
• The algorithm attempts to assign tasks to nodes while ensuring deadlines are met.
• Tasks are categorized into satisfied and unsatisfied.
• If a task meets its deadline on at least one node, it is scheduled efficiently.

Comparison: Min-CCV vs Min-V

Performance Evaluation

• Varying fog nodes while keeping cloud nodes constant.
• Varying cloud nodes while keeping tasks and fog nodes constant.

Conclusion

• Min-CCV is for cost-efficient scheduling that balances multiple factors.
• Min-V is suited for latency-sensitive applications where deadlines are the highest priority.

Cloud Resource Management and Scheduling

• Resource Management: Deciding computing resources.
• Scheduling: Execution of algorithms. Importance lies on:
  • Functionality
  • Performance
  • Costs

Cloud Resource Management (CRM) Policies

• Admission Control: Prevents system overload by restricting incoming workloads.
• Capacity Allocation: Assigns resources for service activations.
• Load Balancing: Ensures even distribution of workload across servers.
• Energy Optimization: Minimizes power consumption.
• Quality of Service (QoS): Maintains performance guarantees per SLA.

Control Theory in Cloud Resource Management

• Main Components: Workload and CRM policies, system performance, resource adjustments, and allocation of resources to applications.

Cloud Controller Structure

• Uses real-time system feedback to predict disturbances and optimize inputs.
• Balances performance and energy efficiency.
• Predictive Filter: Analyzes data to predict resource demand.
• Optimal Controller: Determines action for resource allocation. - Uses predictive data and system feedback to compute the most efficient allocation strategy over a defined time horizon.
• Queuing Dynamics: Represents the relationship between incoming tasks, resource availability, and system performance.
• Disturbance Input: External factors that may impact system performance.
• State Feedback: Real-time monitoring data about current system state. - Enables the controller to make adjustments based on conditions.
• External Traffic: Incoming requests from users or applications.

Key Design Decisions in Cloud Resource Management

• App controllers determine if more resources are needed, and allocation happens through cloud controllers.

Considerations include:

• Fine-grained vs. coarse-grained control - Fine = more parameters
• Static vs. dynamic thresholds.
  • If you get where resources/scaling to more VMs isnt possible- dynamically raise the threshold

Proportional Thresholding Algorithm

• Calculate high and low thresholds based on CPU utilization by taking the integral
  • More data points = better idea of setting the th reshold value
  • request additional VMs if utilization exceeds the high threshold
• Release VMs if utilization drops below the low threshold.
  • Dynamic thresholds outperform static thresholds.
  • Two thresholds are better than one.

Cloud Scheduling Algorithms

• Resource Sharing Levels: Servers shared among VMs, VMs hosting multiple applications, applications with multiple threads.
• Objectives:
  • Batch systems: Maximize throughput, minimize turnaround time.
  • Real-time systems: Meet deadlines, maintain predictability.
• Common Algorithms:
  • Round-robin.
  • First-Come-First-Serve (FCFS).
  • Shortest-Job-First (SJF).
  • Priority-based scheduling.
• Scheduling with Deadlines:
  • Hard deadlines means strict time limits, missing them results in penalties, soft deadlines means flexible limits with no penalties.

Security in DCNs and Cloud Computing

Business Drivers That Influence Security

• Key Business Concerns: Correct & reliable operation, service-level agreements, IT asset value, brand/reputation, legal & regulatory compliance, financial loss & liability, critical infrastructure, safety and survival, insight must balance tangible & intangible assets.

IT Drivers That Influence Security

• Internal threats, external threats, and emphasis on overall impact requiring layered security.

General Network Security Issues

• Common Threats: Data-stealing malware, web threats, malware variants; Considerations: Exploiting vulnerabilities, adaptive/proactive defense.

Virtualization and Cloud Computing Security Issues

• Infrastructure Challenges include: Communication blind spots, Inter-VM attacks, mixed trust level VMs, instant on gaps, resource contention; Implication: Specialized tools are required.

Communication Blind Spots

• This is where security appliances only monitor traffic passing through external network interfaces (traffic between VMs in the same host can bypass these).
• Solutions: Deploy a scanning VM and/or implement self-defending VMs.

Inter-VM Attacks and Hypervisor Compromises

• Inter-VM Attacks: May allow attacker to compromise VMs (use firewalls and IDSs).
• Hypervisor Vulnerabilities: Hyperjacking, API exploitation (strengthen hypervisor security/layered defenses).

Mixed Trust Level VMs

• This is where mission-critical VMs are placed with less-critical VMs on the same physical host.
• High Risks: Risks are lowered by enhancing security, segregating VMs and limiting attack surface.

Instant-On Gaps

• VMs become vulnerable before security policies apply and out-of-date templates are activated.
• Key Point: Continuous security updates are critical to prevent vulnerabilities.

Resource Contention

• Antivirus scans can degrade performance; use randomized/grouped/agent-less scheduling.

Additional Cloud Risks

• Cloning/Rapid Resource Pooling: Has the potential to replicate vulnerabilities.
• Motility/Data Remnants: Data frequently moves/leaves remnants.
• Elastic Perimeter: Increases the attack surface.
• Unencrypted Data: Can expose sensitive data.
• Shared Multi-Tenant Environments: Risk associated with sharing infrastructure.
• Challenges ensure consistent control/access in the cloud
• Overarching theme is to introduce unique challenges that require vigilance

Cloning and Rapid Resource Pooling

• Example: An AWS member uploaded a pre-built machine that still contained an SSH key.
• Lesson: Implement rigorous security/sanitization.
• All VMs should be properly vetted and sanitized.

Motility of Data and Data Remnants

• Aims to know where data resides and should remove data via “sanitization".
• Make sure no data or remnants are left behind

Unencrypted Data

• Implement policy-based key management, and implement more stringent validation.
• In a simple definition, one could not control who could access your information or data.
• It's often a single key.

Shared Multi-Tenant Environments of the Public Cloud

• Encrypt data to ensure it remains unreadable if accessed and prevent unauthorized interactions. It also needs robust isolation controls.

Control and Availability

• Consider weighing the benefits/risks of outages/downtime. - High availability data is crucial-private cloud/ on premise may be preferred

Attackers in the Cloud

• Attackers exploit multi-tenant environments, launching inter-VM attacks or deploying their VMs for compromise.
• The main take is to understand attackers in the cloud and design countermeasures

Common Cloud Security Attacks

• Data Loss and Breaches: caused by software bugs and human error.
• Hypervisor attacks include: VM escape/hopping (protect hypervisor layer for lateral movement).
• Denial of service (DoS) attacks: overwhelms network resources (requires multi-layered defense).

Hypervisor Attack

• Security comes from strengthening the overall workload

DoS Attack Methods

• Overwhelm resources, with methods including flooding HTTP requests, connections, and DNS amplification
• A multi-layered defense strategy is required-encryption/ access control, access control via strict policies.

Cloud Load Balancing

• Achieves high performance by distributing workloads across computing resources.

Goals of Load Balancing

• Load balancing helps achieve optimum resource utilization, throughput, response time, and overload prevention.

Two Categories of Load Balancing

• Node/Server Load Balancing that distributes computing workloads among servers.
• Link/Traffic balancing manages data flow across network paths to prevent congestion.

Server Load Balancing

• Server Load Balancing helps ensure that user-requests reach the appropriate resource.
• Achieved thought the support of: Server availability, session persistence, and protocol and the utilization of natural techniques

Why Use Load Balance Servers

• Scalability – Increases capacity by distributing workloads across multiple servers.
• Maintenance flexibility – Servers can be removed from rotation for updates.
• Resource sharing – Multiple application instances can be hosted efficiently.

Load Balancing Algorithms

Static Load Balancing

• Static means that the distribution happens before anything starts.

Dynamic Load Balancing

• Dynamic distribution: Adjusts real-time (centralized via the master and decentralized via autonomous processors).

Static Load Balancing Techniques

• Round Robin assigns tasks sequentially, randomized assignment distributes tasks randomly, and partition-based divides workload.

Common Server Load Balancing Algorithms

• Least connections is when servers directs new tasks to the least-busy server, and weighted least connections, considering server capacity.
• Round Robin routes requests through all servers, while weighted Round Robin adjusts workload based on server performance.

Centralized Load Balancing Design

• A queue is maintain by the master to assign the task to the various slaves. The slave processes then request termination, process, and assign resources given the instruction.

Dynamic Load Balancing - Decentralized

• Nodes manage workloads independently, addressing workload imbalances in real-time (underloaded request work).

Dynamic load is determined by

• Nodes when they are overloaded seeks to distribute various tasks.
• And selects those nodes to interact and threshold.
• Threshold is what prevents the unnecessary coms

Load Balancing Model

• there is no central node where workloads are exchanged and nodes coordinate via peer-to-peer (P2P).

Processes and Selection

• Receivers are those that require the help that has tasks they must request.
• Senders push tasks to less-busy nodes (most effective in light-load scenarios, though thrashing can occur when load levels are too high).

Strategy Includes

• High synchronization overhead/ balancing achieved using global selection or easier to manage with imbalance, using the local selection.
• Neighbor selection/ round robin distributes tasks or adaptive contracting for assign them

Load detection occurs from the use of

-processor idleness in the way that it prevent those low load conditions

• preventing over load/ highload detection

Google Cloud Load Balancing

• Purpose: Distributes traffic across multiple backend instances to optimize performance, reliability, and availability.
• Software-defined and fully distributed to reduce hardware-related issues. Supports traffic types for HTTP protocols, TCP and more, while handling traffic that provides low latency.

Cloud Load Balancing Overview

• Scalable, managed service that scales to accommodate both internal and external traffic. This is done by traffic balancing based on demand and it quickly scales. - This is all done through: Load Balancers in google Cloud!

Google Cloud Load Balancing types

• Application load balancer that handles requests for HTTP(s)
• Network Load balancer uses TCP, UDP, and SSL protocols at the transport layer.
• Internal load balancer traffic to reduce issues in a private virtual private cloud.

Various types for the configuration on load distribution are

• Cross-Region Load Balancer: Handles global traffic distribution for disaster recovery.
• Passthrough Load Balancer: Designed for very low-latency applications where it flows directly to backend.

Load Balancer Design

• The following features/concepts are at hand; global load balancing, while regional is at hand for residence/compliance

Load Balancing Benefits

• Scaling (handles requests), health checks (Automatically detects / remove failing ones), global supported , traffic routing, and integration.

Various load balancer configuration points are at hand that impact the selection of use

• Application: user Network Internal

• Various parameters for latency and configuration must occur for geographical issues!

• Backend load distribution/ auto-scaling

General Steps to Configure

• Setting up all the backend services/ instance, the create the actual config, then the defining of logging.