Enhancing the Programmability of Cloud Object Storage

Identificador:  TDX:2894

Handle: https://hdl.handle.net/20.500.11797/TDX2894

Autores:  Sampé Domenech, Josep

Resumen:
In a world that is increasingly dependent on technology, digital data is generated in an unprecedented way. This makes companies that require large storage space, such as Netflix or Dropbox, use cloud object storage solutions. This is mainly thanks to their built-in characteristics, such as simplicity, scalability and high-availability. However, cloud object stores face three main challenges: 1) Flexible management of multi-tenant workloads. Commonly, cloud object stores are multi-tenant systems, meaning that all tenants share the same system resources, which could lead to interference problems. Furthermore, it is now complex to manage heterogeneous storage policies in a massive scale. 2) Data self-management. Cloud object stores themselves do not offer much flexibility regarding data self-management by tenants. Typically, they are rigid, which prevent tenants to handle the specific requirements of their objects. 3) Elastic computation close to the data. Placing computations close to the data can be useful to reduce data transfers. But, the challenge here is how to achieve elasticity in those computations without provoking resource contention and interferences in the storage layer. In this thesis, we present three novel research contributions that solve the aforementioned challenges. Firstly, we introduce the first Software-defined Storage (SDS) architecture for cloud object stores that separates the control plane from the data plane, allowing to manage multi-tenant workloads in a flexible and dynamic way. For example, by applying different service levels of bandwidth to different tenants. Secondly, we designed a novel policy abstraction called microcontroller that transforms common objects into smart objects, enabling tenants to programmatically manage their behavior. For example, a content-level access control microcontroller attached to an specific object to filter its content depending on who is accessing it. Finally, we present the first elastic data-driven serverless computing platform that mitigates the resource contention problem of placing computation close to the data.