Mostra i principali dati dell'item

dc.contributor.authorRoscigno, Gianluca
dc.date.accessioned2017-02-02T14:21:10Z
dc.date.available2017-02-02T14:21:10Z
dc.date.issued2016-05-24
dc.identifier.urihttp://hdl.handle.net/10556/2309
dc.identifier.urihttp://dx.doi.org/10.14273/unisa-725
dc.description2014 - 2015it_IT
dc.description.abstractThe era of Big Data is leading the generation of large amounts of data, which require storage and analysis capabilities that can be only ad- dressed by distributed computing systems. To facilitate large-scale distributed computing, many programming paradigms and frame- works have been proposed, such as MapReduce and Apache Hadoop, which transparently address some issues of distributed systems and hide most of their technical details. Hadoop is currently the most popular and mature framework sup- porting the MapReduce paradigm, and it is widely used to store and process Big Data using a cluster of computers. The solutions such as Hadoop are attractive, since they simplify the transformation of an application from non-parallel to the distributed one by means of general utilities and without many skills. However, without any algorithm engineering activity, some target applications are not alto- gether fast and e cient, and they can su er from several problems and drawbacks when are executed on a distributed system. In fact, a distributed implementation is a necessary but not su cient condition to obtain remarkable performance with respect to a non-parallel coun- terpart. Therefore, it is required to assess how distributed solutions are run on a Hadoop cluster, and/or how their performance can be improved to reduce resources consumption and completion times. In this dissertation, we will show how Hadoop-based implementations can be enhanced by using carefully algorithm engineering activity, tuning, pro ling and code improvements. It is also analyzed how to achieve these goals by working on some critical points, such as: data local computation, input split size, number and granularity of tasks, cluster con guration, input/output representation, etc. i In particular, to address these issues, we choose some case studies coming from two research areas where the amount of data is rapidly increasing, namely, Digital Image Forensics and Bioinformatics. We mainly describe full- edged implementations to show how to design, engineer, improve and evaluate Hadoop-based solutions for Source Camera Identi cation problem, i.e., recognizing the camera used for taking a given digital image, adopting the algorithm by Fridrich et al., and for two of the main problems in Bioinformatics, i.e., alignment- free sequence comparison and extraction of k-mer cumulative or local statistics. The results achieved by our improved implementations show that they are substantially faster than the non-parallel counterparts, and re- markably faster than the corresponding Hadoop-based naive imple- mentations. In some cases, for example, our solution for k-mer statis- tics is approximately 30× faster than our Hadoop-based naive im- plementation, and about 40× faster than an analogous tool build on Hadoop. In addition, our applications are also scalable, i.e., execution times are (approximately) halved by doubling the computing units. Indeed, algorithm engineering activities based on the implementation of smart improvements and supported by careful pro ling and tun- ing may lead to a much better experimental performance avoiding potential problems. We also highlight how the proposed solutions, tips, tricks and insights can be used in other research areas and problems. Although Hadoop simpli es some tasks of the distributed environ- ments, we must thoroughly know it to achieve remarkable perfor- mance. It is not enough to be an expert of the application domain to build Hadop-based implementations, indeed, in order to achieve good performance, an expert of distributed systems, algorithm engi- neering, tuning, pro ling, etc. is also required. Therefore, the best performance depend heavily on the cooperation degree between the domain expert and the distributed algorithm engineer. [edited by Author]it_IT
dc.language.isoenit_IT
dc.publisherUniversita degli studi di Salernoit_IT
dc.subjectDistributed computingit_IT
dc.subjectDigital image forensicsit_IT
dc.subjectBioinformaticsit_IT
dc.titleThe Role of Distributed Computing in Big Data Science: Case Studies in Forensics and Bioinformaticsit_IT
dc.typeDoctoral Thesisit_IT
dc.subject.miurINF/01 INFORMATICAit_IT
dc.contributor.coordinatoreCostagliola, Gennaroit_IT
dc.description.cicloXIV n.s.it_IT
dc.contributor.tutorCattaneo, Giuseppeit_IT
dc.identifier.DipartimentoInformaticait_IT
 Find Full text

Files in questo item

Thumbnail
Thumbnail
Thumbnail

Questo item appare nelle seguenti collezioni

Mostra i principali dati dell'item