Reward-Based Admission Controller for Cloud Resource Allocation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current cloud management systems (CMSs) fail to effectively manage 'hidden resources' such as stack limits and process identifiers in Linux containers, leading to resource exhaustion and cascading failures due to unmonitored resource limits, which existing solutions like OpenStack, Kubernetes, and Docker Swarm do not address.
Innovation Solution
A reward-based admission controller that learns from historical data to evaluate the similarity of new job requests to past successful or failed jobs, using a valuation function to determine an expected reward and decide whether to accept or reject requests, thereby preventing failures and cascading issues by considering both known and hidden resource limits.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional admission control is used to manage cloud resources, then resource allocation is straightforward and fast, but hidden resources like stack limits and process identifiers are not monitored, leading to resource exhaustion and cascading failures
Solution Approach 1:
The system performs preliminary clustering of job characteristics and establishes reward thresholds before actual resource allocation. By pre-processing job characteristics and creating clusters based on historical data, the system prepares admission criteria in advance, enabling reliable resource management without requiring complex real-time monitoring of all hidden resources
Solution Approach 2:
The system implements feedback mechanisms by evaluating rewards based on job completion status and using this information to update admission decisions. The reward-based approach continuously learns from past job outcomes, adjusting acceptance criteria to prevent resource exhaustion while maintaining system stability without complex direct monitoring
2Productivity
If all resource requests are accepted without evaluation, then resource utilization is maximized, but failures and cascading issues occur due to unmonitored hidden resource limits
Solution Approach 1:
The system clusters job characteristics and establishes reward thresholds before resource allocation. By pre-processing and categorizing jobs based on historical data, the system can quickly evaluate new requests against established patterns, accepting jobs that match successful patterns while rejecting those likely to cause failures, thus maintaining high utilization without causing cascading failures
Solution Approach 2:
The system changes the evaluation parameter from direct resource monitoring to reward-based scoring. Instead of monitoring each hidden resource limit individually, the system transforms the evaluation into a reward score based on job characteristic clustering and historical performance, enabling efficient productivity maintenance with improved reliability
3Measurement precision
If reward-based evaluation with historical data analysis is implemented, then resource allocation accuracy and failure prevention improve, but computational overhead and processing time increase
Solution Approach 1:
The system performs clustering and threshold establishment in advance based on historical data. By pre-processing job characteristics and creating clusters before actual resource allocation requests, the system reduces real-time computational requirements, achieving high measurement precision without excessive processing delays
Solution Approach 2:
The system applies reward-based evaluation selectively rather than uniformly to all requests. By using pre-established clusters and thresholds, the system performs partial evaluation focusing on key distinguishing features of job characteristics, achieving sufficient accuracy without complete analysis of all parameters, thus reducing processing time
Data Source
AI summary
Technology for selecting job characteristics to determine the similarity among jobs in terms of performance. Technology based on similarity among jobs calculated by selected characteristics for determining jobs that are likely to lead to successful performance of a requested new job by a cloud. Also, technology based on similarity among jobs calculated by selected characteristics for determining jobs that are likely to lead to failure when performing a requested new job by the cloud. When the new job request is accepted, because its characteristics of the new job matches job characteristics characterized by success and/or fails to match job characteristics characterized by failure, then the new job is said to lead to a “reward” or an “expected reward” because the new job will be rewarded by being allowed to use, by an admission controller of a cloud management system, use of cloud computing resources of the cloud.