Towards a Scalable File System on Computer Clusters Using Declustering

Abstract

This study addresses the scalability issues involving file systems as critical components of computer clusters, especially for commercial applications. Given that wide striping is an effective means of achieving scalability as it warrants good load balancing and allows node cooperation, we choose to implement a new data distribution scheme in order to achieve the scalability of computer clusters. We suggest combining both wide striping and replication techniques using a new data distribution technique based on “chained declustering”. Thus, we suggest a complete architecture, using a cluster of clusters, whose performance is not limited by the network and can be adjusted with one-node precision. In addition, update costs are limited as it is not necessary to redistribute data on the existing nodes every time the system is expanded. The simulations indicate that our data distribution technique and our read algorithm balance the load equally amongst all the nodes of the original cluster and the additional ones. Therefore, the scalability of the system is close to the ideal scenario: once the size of the original cluster is well defined, the total number of nodes in the system is no longer limited, and the performance increases linearly.