Survey of Performance Acceleration Techniques for Network Function Virtualization

Linguaglossa, Leonardo; Lange, Stanislav; Pontarelli, Salvatore; Rétvári, Gábor; Rossi, Dario; Zinner, Thomas; Bifulco, Roberto; Jarschel, Michael; Bianchi, Giuseppe

doi:10.1109/jproc.2019.2896848

Cited by 102 publications

(67 citation statements)

References 58 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Furthermore, the input traffic is also opaque to the owner as NICs usually adopt a kernel-bypass approach to reach high speed (cfr. Sec.III of [11]). As such, after resource allocation no optimization is possible on the owner's side: for example, the CPUs must be always active even in the case of zero traffic.…”

Section: Contextmentioning

confidence: 93%

On the Learnability of Software Router Performance via CPU Measurements

Shelbourne

Linguaglossa

Lipani

et al. 2019

Proceedings of the 15th International Conference on Emerging Networking EXperiments and Technologies

Self Cite

View full text Add to dashboard Cite

In the last decade the ICT community observed a growing popularity of software networking paradigms. This trend consists in moving network applications from static, expensive, hardware equipment (e.g. router, switches, firewalls) towards flexible, cheap pieces of software that are executed on a commodity server. In this context, a server owner may provide the server resources (CPUs, NICs, RAM) for customers, following a Service-Level Agreement (SLA) about clients' requirements. The problem of resource allocation is typically solved by overprovisioning, as the clients' application is opaque to the server owner, and the resource required by clients' applications are often unclear or very difficult to quantify. This paper shows a novel approach that exploits machine learning techniques in order to infer the input traffic load (i.e., the expected network traffic condition) by solely looking at the runtime CPU footprint.

show abstract

Section: Contextmentioning

confidence: 93%

On the Learnability of Software Router Performance via CPU Measurements

Shelbourne

Linguaglossa

Lipani

et al. 2019

Proceedings of the 15th International Conference on Emerging Networking EXperiments and Technologies

Self Cite

View full text Add to dashboard Cite

show abstract

“…Several hardware and software strategies can accelerate cloud services [31], [32]. Hardware acceleration can employ custom accelerators, such as FPGA [33] and GPU [34], and dedicated accelerators for service-specific hardware functions, such as DLBoost on Intel Cascade Lake processors [35], which can accelerate deep learning algorithms on both containers and VMs.…”

Section: Related Workmentioning

confidence: 99%

Hardware Acceleration for Container Migration on Resource-Constrained Platforms

et al. 2020

View full text Add to dashboard Cite

The computing capabilities of client devices are continuously increasing; at the same time, demands for ultra-low latency (ULL) services are increasing. These ULL services can be provided by migrating some micro-service container computations from the cloud and multi-access edge computing (MEC) to the client devices. The migration of a container image requires compression and decompression, which are computationally demanding. We quantitatively examine the hardware acceleration of container image compression and decompression on a client device. Specifically, we compare the Intel ® Quick Assist Technology (QAT) hardware acceleration with software compression/decompression. For scenarios with a local container image registry (i.e., without network bandwidth constraints), we find that QAT speeds up compression by a factor of over 7 compared to the single-core GZIP software and reduces the CPU core utilization by over 15% for large container images. These QAT benefits come at the expense of Input/Output (IO) memory access bitrates of up to 900 Mbyte/s (while the software compression/decompression does not require IO memory access). For scenarios with a remote container image registry, we find that the container push (compression) time savings increase with the network bandwidth, while the container pull (decompression) time savings level out for moderately high network bandwidths and slightly decrease for a very high network bandwidth. Furthermore, QAT acceleration achieves substantial power consumption reductions for container push compression for low to moderately high network bandwidths. Our evaluation results give reference performance benchmarks of the achievable latencies for container image instantiation and migration with and without hardware acceleration of the compression and decompression of container images.

show abstract

“…Rather than providing a detailed discussion of implementation and/or acceleration techniques, for which we refer to the survey in [64], we aim in this section to consider each switch design in relation to a number of technical aspects affecting packet processing performance. The objective is to gain insight on how to devise meaningful experimental scenarios.…”

Section: A Design Objectivesmentioning

confidence: 99%

Comparing the performance of state-of-the-art software switches for NFV

Zhang

Linguaglossa

Gallo

et al. 2019

Proceedings of the 15th International Conference on Emerging Networking Experiments and Technologies

Self Cite

View full text Add to dashboard Cite

With the ultimate goal of replacing proprietary hardware appliances with Virtual Network Functions (VNFs) implemented in software, Network Function Virtualization (NFV) has been gaining popularity in the past few years. Software switches route traffic between VNFs and physical Network Interface Cards (NICs). It is of paramount importance to compare the performance of different switch designs and architectures. In this paper, we propose a methodology to compare fairly and comprehensively the performance of software switches. We first explore the design spaces of seven state-of-the-art software switches and then compare their performance under four representative test scenarios. Each scenario corresponds to a specific case of routing NFV traffic between NICs and/or VNFs. In our experiments, we evaluate the throughput and latency between VNFs in two of the most popular virtualization environments, namely virtual machines (VMs) and containers. Our experimental results show that no single software switch prevails in all scenarios. It is, therefore, crucial to choose the most suitable solution for the given use case. At the same time, the presented results and analysis provide a deeper insight into the design tradeoffs and identifies potential performance bottlenecks that could inspire new designs.

show abstract

Survey of Performance Acceleration Techniques for Network Function Virtualization

Cited by 102 publications

References 58 publications

On the Learnability of Software Router Performance via CPU Measurements

On the Learnability of Software Router Performance via CPU Measurements

Hardware Acceleration for Container Migration on Resource-Constrained Platforms

Comparing the performance of state-of-the-art software switches for NFV

Contact Info

Product

Resources

About