東華大學圖書館 |

API Design and Middleware Optimization for Big Data and Machine Learning Applications.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	API Design and Middleware Optimization for Big Data and Machine Learning Applications./
作者:	Guo, Jia.
出版者:	Ann Arbor : ProQuest Dissertations & Theses, : 2021,
面頁冊數:	191 p.
附註:	Source: Dissertations Abstracts International, Volume: 84-02, Section: B.
Contained By:	Dissertations Abstracts International84-02B.
標題:	Computer engineering. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29338526
ISBN:	9798834085249

API Design and Middleware Optimization for Big Data and Machine Learning Applications.
Guo, Jia.

API Design and Middleware Optimization for Big Data and Machine Learning Applications. - Ann Arbor : ProQuest Dissertations & Theses, 2021 - 191 p.

Source: Dissertations Abstracts International, Volume: 84-02, Section: B.

Thesis (Ph.D.)--The Ohio State University, 2021.

This item must not be sold to any third party vendors.

The past decade has witnessed the success of big data processing frameworks, which provide simple interfaces to parallelize and scale applications efficiently. Comparing the design of MapReduce, Spark, and Reduction Object paradigm, we identified that the design of pattern-based Application Programming Interfaces (APIs) can significantly impact the captured application types as well as middleware performance. Therefore, by concluding common patterns in popular big-data and machine learning applications, we want to build new frameworks with both expressive interfaces and efficient middleware, to achieve better parallelism, locality, programmability, fault tolerance, and coverage of applications.To approach this, Chapter 2 studies the impact of API design on programmability and middleware performance of MapReduce(-like) frameworks. Specifically, we introduce two different variations of the original MapReduce API and efficient implementations of all three APIs. Through performance comparison and modeling, we identify that though MapReduce and similar frameworks have demonstrated high programmability, they fall short in terms of performance. We show that Reduction-Object-based APIs, which only require small additional effort from programmers, can provide high performance.Following this work, in Chapter 3, we built a high-throughput stream processing framework that offers a high-level API to the users (similar to Reduction Object), is fault-tolerant, and is also more efficient and scalable than current solutions. Particularly, a cost-efficient MPI/OpenMP-based fault-tolerant scheme is incorporated so that the system can survive node failures with only a modest degradation of performance. A comparison against state-of-the-art streaming frameworks shows our system boosts the throughput of test cases by up to 10X and achieves desirable parallelism when scaled out.In the fast-evolving Internet of Things (IoT) scenario, we envision the potential of leveraging established pattern-based APIs to resolve the challenges of (1) automatically distributing the work between different devices to reduce application latency and (2) parallelizing operations on each multi-core device. Motivated by the popularity of MapReduce(- like) frameworks, in Chapter 4, we develop a pattern-based high-level programming API targeting computer vision applications for the Edge/Fog paradigm with parallelism within devices. Based on this API, parallelization, workload distribution, and optimizations that account for IoT devices' resource limitations, are implemented. Our evaluation shows a 17-45% speedup over OpenCV and desirable load-balancing performance.A more recent trend is to deploy accurate and robust convolution neural networks (CNN) models on edge devices to cooperate with or replace traditional pattern recognition algorithms. Chapter 5 proposes an efficient sparse inference strategy combining direct sparse convolution and fusion. We improve the locality of intermediate results while reducing fusion overhead through loop re-ordering and tiling. We also demonstrate a scalable implementation that utilizes both multi-core and SIMD parallelism. Experiments show that our scheme significantly outperforms existing implementations on an edge device, while also scaling better on multi-core platforms. On top of this design, Chapter 6 brings in state-of-the-art pattern-based sparsity, which uses pre-defined kernel patterns, leading to a high pruning rate while maintaining low accuracy loss. With regrouping of similar patterns and re-structured fusion design, we show that the pattern-based approach and fusion can work together to further improve the inference of sparse CNN on edge devices.

ISBN: 9798834085249Subjects--Topical Terms:

621879
Computer engineering.
Subjects--Index Terms:

Framework

API Design and Middleware Optimization for Big Data and Machine Learning Applications.
LDR:04963nmm a2200421 4500 001 2352285
005 20221118093906.5
008 241004s2021 ||||||||||||||||| ||eng d
020 $a 9798834085249
035 $a (MiAaPQ)AAI29338526
035 $a AAI29338526
040 $a MiAaPQ $c MiAaPQ
100 1 $a Guo, Jia. $3 3433762
245 1 0 $a API Design and Middleware Optimization for Big Data and Machine Learning Applications.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2021
300 $a 191 p.
500 $a Source: Dissertations Abstracts International, Volume: 84-02, Section: B.
500 $a Advisor: Agrawal, Gagan;Teodorescu, Radu.
502 $a Thesis (Ph.D.)--The Ohio State University, 2021.
506 $a This item must not be sold to any third party vendors.
520 $a The past decade has witnessed the success of big data processing frameworks, which provide simple interfaces to parallelize and scale applications efficiently. Comparing the design of MapReduce, Spark, and Reduction Object paradigm, we identified that the design of pattern-based Application Programming Interfaces (APIs) can significantly impact the captured application types as well as middleware performance. Therefore, by concluding common patterns in popular big-data and machine learning applications, we want to build new frameworks with both expressive interfaces and efficient middleware, to achieve better parallelism, locality, programmability, fault tolerance, and coverage of applications.To approach this, Chapter 2 studies the impact of API design on programmability and middleware performance of MapReduce(-like) frameworks. Specifically, we introduce two different variations of the original MapReduce API and efficient implementations of all three APIs. Through performance comparison and modeling, we identify that though MapReduce and similar frameworks have demonstrated high programmability, they fall short in terms of performance. We show that Reduction-Object-based APIs, which only require small additional effort from programmers, can provide high performance.Following this work, in Chapter 3, we built a high-throughput stream processing framework that offers a high-level API to the users (similar to Reduction Object), is fault-tolerant, and is also more efficient and scalable than current solutions. Particularly, a cost-efficient MPI/OpenMP-based fault-tolerant scheme is incorporated so that the system can survive node failures with only a modest degradation of performance. A comparison against state-of-the-art streaming frameworks shows our system boosts the throughput of test cases by up to 10X and achieves desirable parallelism when scaled out.In the fast-evolving Internet of Things (IoT) scenario, we envision the potential of leveraging established pattern-based APIs to resolve the challenges of (1) automatically distributing the work between different devices to reduce application latency and (2) parallelizing operations on each multi-core device. Motivated by the popularity of MapReduce(- like) frameworks, in Chapter 4, we develop a pattern-based high-level programming API targeting computer vision applications for the Edge/Fog paradigm with parallelism within devices. Based on this API, parallelization, workload distribution, and optimizations that account for IoT devices' resource limitations, are implemented. Our evaluation shows a 17-45% speedup over OpenCV and desirable load-balancing performance.A more recent trend is to deploy accurate and robust convolution neural networks (CNN) models on edge devices to cooperate with or replace traditional pattern recognition algorithms. Chapter 5 proposes an efficient sparse inference strategy combining direct sparse convolution and fusion. We improve the locality of intermediate results while reducing fusion overhead through loop re-ordering and tiling. We also demonstrate a scalable implementation that utilizes both multi-core and SIMD parallelism. Experiments show that our scheme significantly outperforms existing implementations on an edge device, while also scaling better on multi-core platforms. On top of this design, Chapter 6 brings in state-of-the-art pattern-based sparsity, which uses pre-defined kernel patterns, leading to a high pruning rate while maintaining low accuracy loss. With regrouping of similar patterns and re-structured fusion design, we show that the pattern-based approach and fusion can work together to further improve the inference of sparse CNN on edge devices.
590 $a School code: 0168.
650 4 $a Computer engineering. $3 621879
650 4 $a Computer science. $3 523869
650 4 $a Design. $3 518875
650 4 $a Information technology. $3 532993
650 4 $a Information science. $3 554358
653 $a Framework
653 $a Big data
653 $a Machine learning
653 $a API
653 $a Middleware
653 $a Parallelism
690 $a 0984
690 $a 0464
690 $a 0389
690 $a 0489
690 $a 0723
690 $a 0800
710 2 $a The Ohio State University. $b Computer Science and Engineering. $3 1674144
773 0 $t Dissertations Abstracts International $g 84-02B.
790 $a 0168
791 $a Ph.D.
792 $a 2021
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=29338526