Charalampos Stylianopoulos, Data- och informationsteknik

Hardware-Aware Algorithm Designs for Efficient Parallel and Distributed Processing


The introduction and widespread adoption of the Internet of Things, together with emerging new industrial applications, bring new requirements in data processing. Specifically, the need for timely processing of data that arrives at high rates creates a challenge for the traditional cloud computing paradigm, where data collected at various sources is sent to the cloud for processing. As an approach to this challenge, processing algorithms and infrastructure are distributed from the cloud to multiple tiers of computing, closer to the sources of data. This creates a wide range of devices for algorithms to be deployed on and software designs to adapt to.


In this thesis, we investigate how hardware-aware algorithm designs on a variety of platforms lead to algorithm implementations that efficiently utilize the underlying resources. We design, implement and evaluate new techniques for representative applications that involve the whole spectrum of devices, from resource-constrained sensors in the field, to highly parallel servers. At each tier of processing capability, we identify key architectural features that are relevant for applications and propose designs that make use of these features to achieve high-rate, timely and energy-efficient processing.


In the first part of the thesis, we focus on high-end servers and utilize two main approaches to achieve high throughput processing: vectorization and thread parallelism. We employ vectorization for the case of pattern matching algorithms used in security applications. We show that re-thinking the design of algorithms to better utilize the resources available in the platforms they are deployed on, such as vector processing units, can bring significant speedups in processing throughout. We then show how thread-aware data distribution and proper inter-thread synchronization allow scalability, especially for the problem of high-rate network traffic monitoring. We design a parallelization scheme for sketch-based algorithms that summarize traffic information, which allows them to handle incoming data at high rates and be able to answer queries on that data efficiently, without overheads.


In the second part of the thesis, we target the intermediate tier of computing devices and focus on the typical examples of hardware that is found there. We show how single-board computers with embedded accelerators can be used to handle the computationally heavy part of applications and showcase it specifically for pattern matching for security-related processing. We further identify key hardware features that affect the performance of pattern matching algorithms on such devices, present a co-evaluation framework to compare algorithms, and design a new algorithm that efficiently utilizes the hardware features.


In the last part of the thesis, we shift the focus to the low-power, resource-constrained tier of processing devices. We target wireless sensor networks and study distributed data processing algorithms where the processing happens on the same devices that generate the data. Specifically, we focus on a continuous monitoring algorithm (geometric monitoring) that aims to minimize communication between nodes. By deploying that algorithm in action, under realistic environments, we demonstrate that the interplay between the network protocol and the application plays an important role in this layer of devices. Based on that observation, we co-design a continuous monitoring application with a modern network stack and augment it further with an in-network aggregation technique. In this way, we show that awareness of the underlying network stack is important to realize the full potential of the continuous monitoring algorithm.


The techniques and solutions presented in this thesis contribute to better utilization of hardware characteristics, across a wide spectrum of platforms. We employ these techniques on problems that are representative examples of current and upcoming applications and contribute with an outlook of emerging possibilities that can build on the results of the thesis.

​Charalampos Stylianopoulos tillhör avdelningen för Nätverk och system vid Data- och informationsteknik.

Opponent

Angelos Bilas, Department of Computer Science, University of Crete & FORTH, Greece.

Länk till publikation

Följ online

https://chalmers.zoom.us/j/61222582506
Kategori Disputation
Plats: Zoom, link above
Tid: 2020-10-08 10:00
Sluttid: 2020-10-08 12:00

Publicerad: må 14 sep 2020.