Erasure Coding for Distributed Storage Systems
Roberta Barbi (Université de Neuchâtel)
A huge quantity of data is generated and stored every day. Preeminent examples are the 300 million photos that get daily uploaded to Facebook or the 156 million emails sent every minute, just think that almost 90% of the data in the world was generated in the last 2 years.
Such data do not settle in the user's machine: they are hosted in data centers which must be able to serve them upon request. Hence, a paramount property for distributed storage systems is reliability, which can be insured carefully introducing redundancy in the system to be able to tolerate adverse situations. To this aim, erasure coding techniques optimizing the storage overhead - fault tolerance tradeoff have been used in production systems for many years.
Only recently, the winds of change have begun to blow, e.g. a local reconstruction code has been introduced in Microsoft Azure in 2012, creating wide space to a number of erasure coding techniques studied during the last 10/20 years. Indeed, more metrics other than fault tolerance and storage overhead are now considered, e.g., repair bandwidth, disk I/O, repair locality, together with encoding and decoding performance. As it is not possible to optimize all of them at the same time, the optimal code depends on the particular application.
In this talk, we focus on the understanding of the different metrics to evaluate erasure codes and how they affect the use-case of storage systems. Then, we study how to introduce data entanglement on top of the coding technique to enhance desirable properties of the resulting system.
Roberta Barbi got a Bachelor Degree in Applied Mathematics from the University of Verona (Italy) in 2013 and a Master Degree in Mathematics with a minor in coding theory and cryptography from the University of Trento (Italy) in December 2015. She started the PhD at the University of Neuchâtel in January 2016 and she expects to defend by February 2019.
Back to EBSIS Events section.