Date of Award

2018

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Madhusudhan Govindaraju

Abstract

This dissertation presents the techniques for adaptation of MapReduce frameworks to incorporate heterogeneity-aware scheduling algorithms, an inspection of cluster configurations and how they impact these scheduling algorithms, an analysis regarding how the cluster configuration and the heterogeneity-aware scheduling can work together to minimize turnaround time and/or power consumption of the cluster when executing MapReduce applications, and how these lessons can be applied more broadly to Big Data infrastructure outside of MapReduce that supports multiple Big Data frameworks simultaneously.

Heterogeneity exists in various capacities in any given cluster, from static (Physical and Platform) heterogeneity to dynamic heterogeneity (Transient Data, Transient Applications, and Irregular Hardware Behavior). Within the cluster there are historically several types of mitigation strategies for each of these types of heterogeneity, and each has their pros and cons. We discuss these mitigation strategies and the types of heterogeneity each of these strategies is able to address, and the history of the related work in the field.

After this, we consider taking host-level metrics and using them to schedule tasks in real time, with a desire to address cluster-wide energy usage. To do this, we consider estimators for power consumption that are available on-chip, namely temperature. We establish a correlation between CPU temperature and power consumption, then derive a scheduling algorithm that eliminates nodes that are consuming too much power from the pool of schedule-able resources. In order to do this we focus on the ability of MapReduce frameworks, constructed as we have constructed the frameworks described in this thesis, to delay binding of tasks to specific workers. We analyze the impacts this has on turnaround time of a MapReduce application, with analysis around setting this threshold properly to reduce impact on turnaround time while shifting power consumption around in the cluster, away from nodes that are over-consuming.

We also address concerns with respect to upgrading a cluster in stages, introducing more Physical Heterogeneity at various levels and the types of adjustments that need to be made to MapReduce configurations in order to combat the increased Heterogeneity. In particular, we look at the concerns for MapReduce platform mis-configuration and its impacts on turnaround time, analyzing the ways in which these types of errors can be mitigated between incremental platform upgrades. In an effort to address this, we introduce a Dynamic Heterogeneity Awareness (DHA) module to our MapReduce framework in order to address these upgrades, and allow better spreading of tasks by the framework, in order to further improve turnaround time and resource utilization.

Finally we consider the implications for framework and application co-tenancy, and we describe the state of art in these areas. We focus on describing what co-tenancy is, why it's important, and how the state of the art can be expanded to in order to leverage findings from this thesis to make these co-tenant clusters increase application and framework performance as well as improving these clusters with considerations for energy efficiency.

Share

COinS