Files in this item

FilesDescriptionFormat

application/pdf

application/pdfGHOSH-DISSERTATION-2018.pdf (5MB)
(no description provided)PDF

Description

Title:Efficient data reconfiguration for today's cloud systems
Author(s):Ghosh, Mainak
Director of Research:Gupta, Indranil
Doctoral Committee Chair(s):Gupta, Indranil
Doctoral Committee Member(s):Vaidya, Nitin; Olson, Luke; Elmore, Aaron
Department / Program:Computer Science
Discipline:Computer Science
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:Ph.D.
Genre:Dissertation
Subject(s):reconfiguration
partitioning
replication
prefetching
compaction
nosql databases
interactive analytics engines
tiered storage systems
Abstract:Performance of big data systems largely relies on efficient data reconfiguration techniques. Data reconfiguration operations deal with changing configuration parameters that affect data layout in a system. They could be user-initiated like changing shard key, block size in NoSQL databases, or system-initiated like changing replication in distributed interactive analytics engine. Current data reconfiguration schemes are heuristics at best and often do not scale well as data volume grows. As a result, system performance suffers. In this thesis, we show that {\it data reconfiguration mechanisms can be done in the background by using new optimal or near-optimal algorithms coupling them with performant system designs}. We explore four different data reconfiguration operations affecting three popular types of systems -- storage, real-time analytics and batch analytics. In NoSQL databases (storage), we explore new strategies for changing table-level configuration and for compaction as they improve read/write latencies. In distributed interactive analytics engines, a good replication algorithm can save costs by judiciously using memory that is sufficient to provide the highest throughput and low latency for queries. Finally, in batch processing systems, we explore prefetching and caching strategies that can improve the number of production jobs meeting their SLOs. All these operations happen in the background without affecting the fast path. Our contributions in each of the problems are two-fold -- 1) we model the problem and design algorithms inspired from well-known theoretical abstractions, 2) we design and build a system on top of popular open source systems used in companies today. Finally, using real-life workloads, we evaluate the efficacy of our solutions. Morphus and Parqua provide several 9s of availability while changing table level configuration parameters in databases. By halving memory usage in distributed interactive analytics engine, Getafix reduces cost of deploying the system by 10 million dollars annually and improves query throughput. We are the first to model the problem of compaction and provide formal bounds on their runtime. Finally, NetCachier helps 30\% more production jobs to meet their SLOs compared to existing state-of-the-art.
Issue Date:2018-11-12
Type:Thesis
URI:http://hdl.handle.net/2142/102420
Rights Information:2018 Mainak Ghosh
Date Available in IDEALS:2019-02-06
Date Deposited:2018-12


This item appears in the following Collection(s)

Item Statistics