Computer Science ETDs

Publication Date

7-1-2016

Abstract

Next generation HPC applications will increasingly time-share system resources with emerging workloads such as in-situ analytics, resilience tasks, runtime adaptation services and power management activities. HPC systems must carefully schedule these co-located codes in order to reduce their impact on application performance. Among the techniques traditionally used to mitigate the performance effects of time- share systems is gang scheduling. This approach, however, leverages global synchronization and time agreement mechanisms that will become hard to support as systems increase in size. Alternative performance interference mitigation approaches must be explored for future HPC systems. This dissertation evaluates the impacts of workload concurrency in future HPC systems. It uses simulation and modeling techniques to study the performance impacts of existing and emerging interference sources on a selection of HPC benchmarks, mini-applications, and applications. It also quantifies the cost and benefits of different approaches to scheduling co-located workloads, studies performance interference mitigation solutions based on gang scheduling, and examines their synchronization requirements. To do so, this dissertation presents and leverages a new Extreme Value Theory- based model to characterize interference sources, and investigate their impact on Bulk Synchronous Parallel (BSP) applications. It demonstrates how this model can be used to analyze the interference attenuation effects of alternative fine-grained OS scheduling approaches based on periodic real time schedulers. This analysis can, in turn, guide the design of those mitigation techniques by providing tools to understand the tradeoffs of selecting scheduling parameters.

Language

English

Keywords

Exascale applications, composed applications, in-situ, resource sharing, performance interference, modeling, extreme value theory

Document Type

Dissertation

Degree Name

Computer Science

Level of Degree

Doctoral

Department Name

Department of Computer Science

First Advisor

Bridges, Patrick G.

First Committee Member (Chair)

Arnold, Dorian

Second Committee Member

Estrada, Trilce

Third Committee Member

Shu, Wei Wennie

Project Sponsors

Exascale Operating and Runtime Systems Program from the United States Department of Energy, Fulbright Colombia, Colciencias, and the Universidad Autonoma de Occidente.

Share

COinS