University of Crete (UOC)
Accelerating stateful network packet processing using graphics hardware
Abstract
dc:descriptionThe need for differentiated services (such as firewalls, network intrusion detection/prevention systems and traffic classification applications) that lie in the core of the Internet, instead of the end points, constantly increases. These services need to perform complex packet processing operations at upper networking layers, which, unfortunately, are not supported by traditional edge routers. To address this evolution, specialized network appliances (called "middleboxes") are deployed, which typically perform complex packet processing operations - ranging from deep packet inspection operations to packet encryption and redundancy elimination. Packet-processing implemented in software promises to enable the fast deployment of new, sophisticated processing without the need to buy and deploy expensive new equipment. In this thesis, we propose to increase the throughput of packet processing operations by using Graphics Processing Units (GPUs). GPUs have evolved to massively parallel computational devices, containing hundreds of processing cores that can be used for general-purpose computing beyond graphics rendering. GPUs, however, have a different set of constraints and properties that can prevent existing software from obtaining the improved throughput benefits GPUs can provide.This dissertation analyzes the tradeoffs of using modern graphics processors for stateful packet processing and describes the software techniques needed to improve its performance. First, we present a deep study into accelerating packet processing operations using discrete modern graphics cards. Second, we present a broader multi-parallel stateful packet processing architecture that carefully parallelizes network traffic processing and analysis at three levels, using multi-queue network interfaces (NICs), multiple CPUs, and multiple GPUs. Last, we explore the design of a GPU-based stateful packet processing framework, identifying a modular mechanism for writing GPU-based packet processing applications, eliminating excessive data transfers as well as redundant work found in monolithic GPU-assisted applications. Our experimental results demonstrate that properly architecting stateful packet processing software for modern GPU architectures can drastically improve throughput compared to a multi-core CPU implementation.
Degree
thesis:*- Grantor dc:publisher
- University of Crete (UOC)
- Year dc:date
- 2014
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Vasiliadis, Giorgos
- Βασιλειάδης, Γεώργιος
Subjects
dc:subject × 8Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/36132