Einladung zum Statusvortrag von Herrn M.Sc. Sebastian Oeste

Abstract:

Scientific workloads increasingly move vast volumes of data. In modern high-performance computing (HPC), the Input/Output (I/O) subsystem is built as a stack of layers connected by interfaces. Each interface exposes an API and implicit semantic properties such as atomicity, visibility and ordering that determine the meaning of I/O operations. Although operations often look similar across layers, their guarantees differ, creating a semantic gap that obscures application intent as requests traverse the stack. This gap complicates correctness reasoning and limits cross-layer optimization.

This work makes three contributions. First, it provides a general definition of I/O semantics and a unified classification along six categories: concurrent access, persistency, consistency, spatiality, temporality, and mutability. This taxonomy provides a common language to reason about, and compare I/O behaviors across layers. Second, the introduction of IOVerbs, an API that allows applications to declare the semantics of their data accesses and propagate them through the lower layers of the stack, enabling targeted reshaping, buffering, and data placement without sacrificing correctness. Third, an analysis technique that derives the required I/O semantics from application event traces, identifying the minimal semantics necessary for correctness while maximizing optimization freedom.