Return to search

Accelerating Audio Data Analysis with In-Network Computing

Digital transformation will experience massive connections and massive data handling. This will imply a growing demand for computing in communication networks due to network softwarization. Moreover, digital transformation will host very sensitive verticals, requiring high end-to-end reliability and low latency. Accordingly, the emerging concept “in-network computing” has been arising. This means integrating the network communications with computing and also performing computations on the transport path of the network. This can be used to deliver actionable information directly to end users instead of raw data.

However, this change of paradigm to in-network computing raises disruptive challenges to the current communication networks. In-network computing (i) expects the network to host general-purpose softwarized network functions and (ii) encourages the packet payload to be modified. Yet, today’s networks are designed to focus on packet forwarding functions, and packet payloads should not be touched in the forwarding path, under the current end-to-end transport mechanisms. This dissertation presents fullstack in-network computing solutions, jointly designed from network and computing perspectives to accelerate data analysis applications, specifically for acoustic data analysis.

In the computing domain, two design paradigms of computational logic, namely progressive computing and traffic filtering, are proposed in this dissertation for data reconstruction and feature extraction tasks. Two widely used practical use cases, Blind Source Separation (BSS) and anomaly detection, are selected to demonstrate the design of computing modules for data reconstruction and feature extraction tasks in the in-network computing scheme, respectively. Following these two design paradigms of progressive computing and traffic filtering, this dissertation designs two computing modules: progressive ICA (pICA) and You only hear once (Yoho) for BSS and anomaly detection, respectively. These lightweight computing modules can cooperatively perform computational tasks along the forwarding path. In this way, computational virtual functions can be introduced into the network, addressing the first challenge mentioned above, namely that the network should be able to host general-purpose softwarized network functions. In this dissertation, quantitative simulations have shown that the computing time of pICA and Yoho in in-network computing scenarios is significantly reduced, since pICA and Yoho are performed, simultaneously with the data forwarding. At the same time, pICA guarantees the same computing accuracy, and Yoho’s computing accuracy is improved.

Furthermore, this dissertation proposes a stateful transport module in the network domain to support in-network computing under the end-to-end transport architecture. The stateful transport module extends the IP packet header, so that network packets carry message-related metadata (message-based packaging). Additionally, the forwarding layer of the network device is optimized to be able to process the packet payload based on the computational state (state-based transport component). The second challenge posed by in-network computing has been tackled by supporting the modification of packet payloads.

The two computational modules mentioned above and the stateful transport module form the designed in-network computing solutions. By merging pICA and Yoho with the stateful transport module, respectively, two emulation systems, i.e., in-network pICA and in-network Yoho, have been implemented in the Communication Networks Emulator (ComNetsEmu). Through quantitative emulations, the experimental results showed that in-network pICA accelerates the overall service time of BSS by up to 32.18%. On the other hand, using in-network Yoho accelerates the overall service time of anomaly detection by a maximum of 30.51%. These are promising results for the design and actual realization of future communication networks.

Identiferoai:union.ndltd.org:DRESDEN/oai:qucosa:de:qucosa:86497
Date19 July 2023
CreatorsWu, Huanzhuo
ContributorsFitzek, Frank, Rein, Stephan, Franchi, Norman, Technische Universität Dresden
Source SetsHochschulschriftenserver (HSSS) der SLUB Dresden
LanguageEnglish
Detected LanguageEnglish
Typeinfo:eu-repo/semantics/publishedVersion, doc-type:doctoralThesis, info:eu-repo/semantics/doctoralThesis, doc-type:Text
Rightsinfo:eu-repo/semantics/openAccess

Page generated in 0.0021 seconds