Distributed Host Capability Advertisements for Cloud Scheduling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing virtual cloud computing infrastructure faces inefficiencies in scheduling virtual resources due to the overhead of maintaining host states in a centralized system, leading to delays and increased computational burden, especially in large data centers with many hosts.
Innovation Solution
A distributed scheduling solution where hosts periodically advertise their capabilities to a scheduler controller, allowing for decentralized computation of host states and simplifying the scheduling process by using a hash table to store and select available resources, reducing the need for a single central controller to maintain all host states.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a centralized system maintains host states in a single controller, then scheduling decisions can be made centrally, but the computational overhead and complexity increase significantly with the number of hosts
Solution Approach 1:
The patent segments the centralized scheduling system into distributed components: each host maintains its own state information and advertises capabilities independently to scheduler controllers. This segmentation reduces the computational burden on any single controller and distributes the complexity across multiple hosts, directly resolving the contradiction between scheduling efficiency and controller complexity
Solution Approach 2:
The patent introduces advertisement messages as intermediaries that carry host capability information from hosts to scheduler controllers. These advertisements serve as a mediator that eliminates the need for controllers to directly query and maintain detailed host states, reducing controller complexity while maintaining scheduling efficiency
2Measurement precision
If a centralized controller queries all hosts for state information, then accurate scheduling decisions can be made, but the time overhead and communication burden increase
Solution Approach 1:
The patent implements preliminary action by having hosts proactively advertise their capabilities and state changes before the scheduler needs this information. Hosts send advertisement messages periodically or upon state changes, eliminating the need for time-consuming queries at scheduling decision points and reducing scheduling delays while maintaining accurate host state information
Solution Approach 2:
The patent establishes a feedback mechanism where hosts continuously advertise their capability changes to scheduler controllers. This feedback loop ensures the scheduler always has access to current host state information without requiring active queries, maintaining measurement precision while minimizing time overhead
3Productivity
If the scheduler maintains detailed host states for all hosts, then resource allocation can be optimized, but the memory and computational resources required increase
Solution Approach 1:
The patent extracts the host state maintenance function from the centralized scheduler and places it at the hosts themselves. Each host maintains its own capability information and advertises it to controllers, eliminating the need for controllers to store detailed host states in memory. This extraction reduces the memory and computational resources required at the controller level while maintaining resource allocation efficiency
Solution Approach 2:
The patent uses advertisement messages as lightweight copies of host state information that are transmitted from hosts to controllers. Instead of maintaining full host state representations in controller memory, the system uses these message-based copies that can be generated on-demand by hosts, significantly reducing the quantity of memory resources required at controllers while preserving resource allocation efficiency
Data Source
AI summary
The present disclosure describes a distributed, advertisement-based, solution for scheduling virtual resources in cloud infrastructures such as the OpenStack. The scheduling algorithm distributes the scheduling requirements and host state feasibility checks to the individual hosts in the datacenter, which can periodically send a summarized advertisement to the scheduler controller listing the number of instances of different type(s) of virtual resources that a particular host can support. The scheduler controller, thus no longer has to compute and maintain individual host states, and the scheduling problem is reduced to selecting the feasible advertisements that satisfy a given request. The solution can be extended to a scenario of multiple scheduler controllers using the same distributed, advertisement-based, approach.