Disputation von Jakob Lüttgau am 19.10.2021, 16:00 Uhr
19. Oktober 2021, von Reinhard Zierke
Einladung zur hochschulöffentlichen Disputation im Rahmen des Promotionsverfahrens von
Herrn Jakob Lüttgau
Titel der Dissertation:
Decision-Support for Workflow-Aware High-Performance Storage Systems
Abstract:
As a result of technological advancements and data center economics, rapid changing systems and increasing complexity render manual system-specific tuning of application and workflow input/output (I/O) unfeasible. The next generation of computer systems enable and require adaptive heuristics for optimal system exploitation. A key challenge to empower adaptive heuristics is to provide decision components with the necessary information (decision aids) to base decisions on. In many cases, these decisions will have to negotiate the user's intent, the temporal unraveling of I/O activity, and an operational perspective. This notion of intent is embedded in the workflows of users and scientists. The thesis conducts an extensive analysis of scientific workflows as well as the storage data path to inform an architecture proposal adding workflow-adaptivity to existing and future storage systems. At the core of this architecture are decision points and information sources scattered across high-performance computing systems. For a particular workflow, the assumption is, that various decision aids can be compiled from information sources and workflow artifacts that can confidently make claims about the I/O characteristics of a workflow. These decision aids in turn, can be used to optimize utilization and performance. To make adoption of this approach feasible, the architecture introduces a novel API and a new service. The Decision API allows applications and middleware the declaration of decision points and thus delegate a decision to exchangeable components. The Decision Aid Propagation Service (DAPS) simplifies the efficient collection and distribution of decision aids which may be generated in-or out-of-band. The results demonstrate how workflow-specific optimization can be achieved with the proposed architecture. The results also demonstrate how breaking up an intractable end-to-end goal into smaller independent steps gives rise to a flexible architecture to discover and utilize insights about application and workflow behavior.
Datum und Uhrzeit: Dienstag, 19. Oktober 2021 um 16:00 Uhr
Ort: per Videokonferenz in Zoom
Betreuer: Prof. Dr. Thomas Ludwig
Prof. Dr. Matthias Rarey
Vorsitzender des Fachpromotionsausschusses Informatik