TOWARD ELASTIC PARTITIONING OF MULTI-TENANT COMPUTING SYSTEMS AT THE EDGE

Fog computing is gaining momentum to extend Cloud resources in close proximity to data sources, end users or both. Among the explored Fog deployment models, the Public Fog offers compute and memory resources for open use to IoT service providers, and is emerging as a fundamental component for an Edge-Fog-Cloud complete compute continuum along which IoT services can be flexibly instantiated. The multi-tenant nature of public Fog nodes represents a major design and management challenge at the intersection of different yet related research disciplines, ranging from dynamic mapping of manycore architectures to resource management for Cloud and Fog resources, and from computing acceleration to software virtualization. The fundamental challenge is to efficiently share the limited pool of Fog resources among multiple consolidated IoT services sharing the same hardware platform. This thesis revolves around the key intuition that multi-tenancy could be reconciled with limited resource capacity through an elastic provisioning of Fog resources. As a result, the thesis proposes a holistic support for elastic Fog computing, following a bottom-up methodology. The support is fundamentally rooted in the capability of the on-chip interconnection network (network-on-chip, NoC) to spatially and temporally isolate communication flows originated by different IoT services. The isolation in space enables to partition the Fog node architecture into spatially-isolated execution environments that provide enhanced security with respect to software-only isolation, and the strictest notion of service composability. At this level, the thesis proposes pLBDR, a lightweight routing mechanism that prevents functional and non-functional interference of intra-partition communication flows with one another. Above all, it combines low complexity with fast dynamic reconfigurability of the partitioning pattern, thus delivering a NoC-supported elastic partitioning in space that is out-of-reach for current NoC technology. In space-multiplexed parallel computing architectures, some communications unavoidably break the spatial locality, especially those associated with memory controller and system configuration traffic. For these flows, this thesis provides efficient time-multiplexing while meeting the distinctive requirements of an elastic Fog environment: low-latency communication scheduling in time, and runtime reconfigurability of the number of time slots. This new set of requirements make the proposed time-multiplexed NoC a unique design point in the open literature. In compliance with its bottom-up approach, the thesis finally tackles the resource management challenge to master the elasticity properties of the underlying compute and memory partitions. In line with mainstream approaches to resource management for manycore systems, the thesis assumes a hierarchical framework where virtual resource reassignments are dynamically changed into the actual reallocation of physical resources. At this level, the shape, size and location of space partitions have to be adjusted in a non-overlapping way to fulfil the variations. The thesis proposes an Integer Linear Programming shape-based model that strives to deliver prioritized latency guarantees to IoT services while perturbing the system state the least possible. The modest running times enable the deployment of the proposed Partition Manager for online use, in combination with a "prior provisioning prompt allocation" scheme for resource utilization in Fog computing. Overall, this thesis is a highly interdisciplinary piece of work that provides an integrated hardware/software support for elastic Fog computing, and paves the way for a dynamically-orchestrated Edge-Fog-Cloud continuum serving as a seamless hosting environment for the next generation of smart IoT services.