Presentation
GrIOt: Graph-Based Modeling of HPC Application I/O Call Stacks for Predictive Prefetch
DescriptionModern HPC storage systems use tiers of heterogeneous storage technologies to compromise between capacity, performance, and cost. Prefetching is a technique used in these systems to move the right data at the right time from a slower to a high-performance tier in order to improve performance with a limited cost. Prefetching requires knowledge of the application I/O patterns, which can be extracted through I/O tracing tools or functions call stacks. State-of-the-art solutions based on the latter focus on applications with regular I/O profiles because of scalability issues. In this paper, we present an approach based on I/O call stacks that models I/O patterns for both regular and irregular applications by using directed graphs. We present two different models for prefetching with different trade-offs between complexity and accuracy of the prefetch predictions. Our models were able to predict I/Os with an accuracy of up to 98%, while keeping a lower overhead.
