Kishori Konwar: RECS : Reconfigurable, Erasure coded, Consistent Storage

Friday, November 17, 2017 - 1:00pm to 2:30pm
Kishori Konwar
We present a new dynamic distributed storage system, which offers strong consistency guarantee for read and write operations. Large-scale distributed storage systems are at the heart of many cloud and web-based data-dependent applications. Being dynamic provides the flexibility to scale-up or scale-down the set of storage servers used by various algorithms. It also offers the ability to add, remove or replace any subset of the storage servers without interrupting the service. A strongly consistent storage system guarantees the illusion of a single storage server handling the read and write requests from various clients atomically, thereby, enabling the application developer with simple and intuitive data access semantics. We are interested in storage systems that employ erasure-codes for providing fault-tolerance. Erasure codes help to significantly minimize storage and communication costs of the system when compared to the de facto replication-based algorithms. This is the first ever algorithmic study of incorporating erasure-codes in dynamic storage systems while providing strong consistency guarantees. Our algorithmic framework provides the flexibility to chose the erasure code from among several erasure coding families that are widely considered for modern-day large-scale distributed storage.
In this talk, we present a client-server model for dynamic storage systems which use consensus-based reconfiguration service. We explain the challenges involved in simultaneously supporting strong consistency and erasure codes. We present the algorithmic framework, and discuss its properties and benefits over existing methods.