Advisory Committee Chair
Anthony Skjellum
Advisory Committee Members
Alan Sprague
Matthew L Curry
Purushotham Bangalore
Robert B Ross
Document Type
Dissertation
Date of Award
2013
Degree Name by School
Doctor of Philosophy (PhD) College of Arts and Sciences
Abstract
In order to satisfy the storage demands of Exascale computing by 2018 (a stated goal of the US government and the HPC research community), a number of Exascale storage systems have begun to be designed and explored to facilitate application checkpointing and parallel file storage. Nondeterministic storage systems (which weaken the assumptions of Terascale and Petascale predecessors) can provide useful capabilities that are difficult to implement in today's parallel file systems, such as high-performance writes, dynamic load-balancing, enriched flexibility of moving data and creating replicas, and potentially other features. However, locating data and reading data back from nondeterministic storage systems are challenging tasks. Fundamentally, data must remain flexible in its location ("in motion") throughout its life, in order to remain available/durable. However, such data must also be found promptly in order for it to remain useful throughout its life cycle. In this dissertation, we present a new approach for locating data in Exascale storage systems. With our lightweight approach, we designed two search-based data location services to enable free data placement, movement, and replication in nondeterministic Exascale storage systems. We evaluated our protocols and algorithms through extensive simulations of HPC storage environments. The results we obtained showed that our methods achieved much higher search efficiency than pure flooding search, while offering comparable search speed and search coverage. The main contribution of this dissertation is that we explored and demonstrated the viability of nondeterministic data placement in Exascale storage systems by showing our solutions for finding data quickly and efficiently. Our solutions can be directly applied in existing nondeterministic Exascale storage systems to find data and thereby enable reads. This work answers the important question of whether a nondeterministic file system is a viable strategy because we can indeed locate data placed freely. Specific results related to managing fault-tolerant overlay networks are important collateral findings of this work, which have potential applicability in other aspects of peer-to-peer computing and Exascale networking.
Recommended Citation
Sun, Zhiwei, "Lightweight Data Location Services for Exascale Storage Systems" (2013). All ETDs from UAB. 3071.
https://digitalcommons.library.uab.edu/etd-collection/3071