Contributions to data replication in large-scale distributed systems

Data replication is a key mechanism for building a reliable and efficient data management system. Indeed, by keeping several replicas for each piece of data, it is possible to improve durability. Furthermore, well-placed copies reduce data access time. However, having multiple copies for a single piece of data creates consistency problems when the data is updated. Over the last years, I made contributions related to these three aspects: data durability, data access performance and data consistency. RelaxDHT and SPLAD enhance data durability by placing data copies smartly. Caju, AREN and POPS reduce access time by improving data locality and by taking popularity into account. To enhance data lookup performance, DONUT creates efficient shortcuts taking data distribution into account. Finally, in the replicated database context, Gargamel parallelizes independent transactions only, improving database performance and avoiding aborting transactions. My research has been carried out in collaboration with height PhD students, four of which have defended. I my future work, I plan to extend these contributions by (i) designing a storage system tailored for MMOGs, which are very demanding, and (ii) designing a data management system that is able to re-distribute data automatically in order to scale the number of servers up and down according to the changing workload, leading to a greener data management.

6 PhD graduated 2012 - 2019