[Paper Review] Predicting Dynamic Replication based on Fuzzy System in Data Grid

Mahnaz Khojand, Mehdi Fatan Serj+2 more|arXiv (Cornell University)|Apr 9, 2018

Distributed and Parallel Computing Systems13 references3 citations

TL;DR

This paper proposes Predictive Fuzzy Replication (PFR), a dynamic replication algorithm for data grids that uses a fuzzy inference system to predict future file access patterns and proactively replicate files in optimal sites based on historical usage, file size, site level, and available storage. PFR improves response time and reduces wasted resources by pre-emptively replicating files and deleting unused replicas, outperforming the Cascading method in efficiency.

ABSTRACT

Data grid replication is an effective method to achieve efficient and fault tolerant data access while reducing access latency and bandwidth consumption in grids. Since we have storage limitation, a replica should be created in the best site. Through evaluation of previously suggested algorithms, we understand that by blind creation of replications on different sites after each demand, we may be able to improve algorithm regarding response time. In practice, however, most of the created replications will never be used and existing resources in Grid will be wasted through the creation of unused replications. In this paper, we propose a new dynamic replication algorithm called Predictive Fuzzy Replication (PFR). PFR not only redefines the Balanced Ant Colony Optimization (BACO) algorithm, which is used for job scheduling in grids, but also uses it for replication in appropriate sites in the data grid. The new algorithm considers the history usage of files, files size, the level of the sites and free available space for replication and tries to predict future needs and pre replicates them in the resources that are more suitable or decides which replica should be deleted if there is not enough space for replicating. This algorithm considers the related files of the replicated file and replicates them considering their own history. PFR acts more efficiently than Cascading method, which is one of the algorithms in optimized use of existing replicas.

Motivation & Objective

• To address the inefficiency of reactive replication in data grids, where unused replicas consume storage and bandwidth.
• To reduce response time and bandwidth consumption by proactively placing replicas based on predicted demand.
• To minimize resource wastage by intelligently deleting unused replicas when storage is constrained.
• To improve upon existing replication strategies like the Cascading method by incorporating predictive intelligence.
• To enhance replica placement decisions using a hybrid approach combining fuzzy logic with Balanced Ant Colony Optimization (BACO).

Proposed method

• PFR employs a fuzzy inference system to evaluate key factors: historical file access frequency, file size, site level (priority), and available free space.
• The system predicts future access demand for a file and its related files, enabling proactive replication.
• Replication decisions are made using a fuzzy logic engine that maps input variables (usage history, size, site level, free space) to a replication score.
• The algorithm integrates BACO to optimize replica placement, balancing load and minimizing access latency.
• It considers dependencies between files, replicating related files based on their own access history and predicted demand.
• When storage is insufficient, PFR uses the fuzzy system to decide which replica to delete based on access likelihood and site priority.

Experimental results

Research questions

• RQ1 How can dynamic replication in data grids be improved to reduce response time and bandwidth consumption?
• RQ2 What factors should be considered to predict future file access patterns and enable proactive replication?
• RQ3 How can fuzzy logic effectively model the trade-offs between storage constraints, file access frequency, and site capabilities?
• RQ4 In what way does integrating BACO with fuzzy logic enhance replica placement decisions compared to existing methods?
• RQ5 How does the proposed PFR algorithm compare in performance to the Cascading replication method in terms of efficiency and resource utilization?

Key findings

• PFR reduces response time by proactively placing replicas in high-demand sites based on predicted access patterns.
• The algorithm significantly reduces the number of unused replicas, minimizing storage waste and bandwidth overhead.
• PFR outperforms the Cascading method in terms of efficiency, particularly in dynamic and high-load grid environments.
• The integration of fuzzy logic with BACO enables more intelligent and adaptive replica placement decisions.
• By considering related files and their access history, PFR improves data availability and access performance for interdependent workloads.
• The system effectively manages storage constraints by intelligently deleting low-potential replicas when space is limited.