User Replica Placement Views in Distributed Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The computational complexity of determining user data replica placement in cloud computing systems grows exponentially with the number of user data replicas, making it inefficient and time-consuming to allocate resources effectively.
Innovation Solution
The method involves model order reduction by transforming the high-dimensional problem of user data replica placement into a lower-dimensional one, using linear programming to simplify the distribution and allocate resources more efficiently, while maintaining inter-replica constraints such as shared-fate risk minimization, co-regional placement, and failover volume restrictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional methods are used to determine user data replica placement, then placement accuracy can be maintained, but computational complexity grows exponentially with the number of replicas
Solution Approach 1:
The patent segments the high-dimensional replica placement problem into multiple lower-dimensional subproblems by introducing hierarchical levels of abstraction. Instead of optimizing all replica placements simultaneously across all clusters, the system divides the problem into user-level, replica-level, and cluster-level decisions, solving each segment separately to reduce overall computational complexity while maintaining placement accuracy
Solution Approach 2:
The patent transforms the exponentially complex high-dimensional optimization problem into a series of lower-dimensional problems by adding temporal and hierarchical dimensions. By separating the optimization into different time scales (coarse-grained vs. fine-grained) and organizational levels (user-level vs. replica-level), the system reduces the dimensionality of each subproblem while preserving the essential constraints and objectives
2Reliability
If detailed replica placement optimization is performed, then service quality can be improved, but resource allocation time increases significantly
Solution Approach 1:
The patent implements periodic action by alternating between coarse-grained optimization (performed less frequently) and fine-grained optimization (performed more frequently). The coarse-grained phase establishes baseline replica placements at user-level, while the fine-grained phase periodically refines these placements at replica-level. This periodic alternation maintains service quality through regular optimization while reducing average allocation time by not performing full detailed optimization continuously
Solution Approach 2:
The patent applies preliminary action by first performing coarse-grained optimization to establish initial replica placements before conducting fine-grained optimization. The coarse-grained phase pre-processes the problem by determining user-level replica distributions, which serves as a foundation for the subsequent fine-grained optimization. This preliminary action reduces the search space for the detailed optimization phase, thereby reducing overall computation time while maintaining service quality
3Reliability
If the number of user data replicas per user is increased, then system availability can be improved, but computational complexity increases exponentially
Solution Approach 1:
The patent segments the replica placement problem by introducing hierarchical levels that scale linearly with the number of replicas rather than exponentially. By organizing the optimization hierarchy from user-level to replica-level to cluster-level, each level processes a manageable subset of decisions independent of the total replica count, allowing the system to handle increased replica numbers for improved availability without exponential complexity growth
Solution Approach 2:
The patent transforms the exponential complexity problem into linear scalability by adding hierarchical and temporal dimensions to the optimization process. The multi-level hierarchy decomposes the problem space, and the alternating optimization phases separate time scales, converting what would be an exponentially complex single-dimensional problem into a series of linearly scalable lower-dimensional subproblems that can handle increased replica counts efficiently
Data Source
AI summary
This description describes technologies relating to resource allocation in distributed systems. In general, one aspect of the subject matter described in this specification can be embodied in methods that identify an original distribution of replicas for a group of users, the original distribution including at least one view corresponding to an assignment of all a user's replicas a combination of clusters; identify a desired distribution of replicas of the group of users, the desired distribution corresponding to an assignment of one user replica to one cluster; determine detailed user replica views for the group of users based on the original distribution and the desired distribution; and allocate resources according to the detailed user replica views for the group.