Péter Orosz

A case study on correlating video QoS and QoE

Finding the correlation among Quality of Experience (QoE) for video, measured Quality of Service (QoS) parameters in the network, and objective video performance metrics is a challenging task. This paper provides some analysis results on this issue. Our motivation is that streaming media content gets dominant position in the global traffic mix within the next few years. With the evolution of personal devices, demand for High Definition (HD) resolution contents is dynamically increasing. Traversing real-time media across public packet-switched networks is a complex task, especially if quality of service should be sustained. The issue gets more complicated when the traffic is forwarded through heterogeneous infrastructures. Media content with various resolutions and bit rates show different sensitivity to transmission anomalies. Our paper investigates the correlation between subjective quality assessment (i.e., Mean Opinion Score, MOS evaluation), measured QoS parameters (packet loss, jitter) and objective video performance metrics (Video Quality Metric - VQM, Structural Similarity - SSIM, Peak Signal Noise Ratio - PSNR) in the context of real-time HD video streaming (i.e., IPTV and MobileTV). In twelve scenarios, packet-level perturbations were emulated in our laboratory testbed during the transmission of short video sequences with three different resolutions (480p, 720p and 1080p). Later the videos were evaluated using subjective and objective assessment methods.

Software-Based Packet Capturing with High Precision Timestamping for Linux

Widely used network measurement applications, such as tcpdump and Wireshark, use the same common libpcap packet capture library. Libpcap assigns a 10-6 second precision timestamp to all processed frames. Higher physical bandwidth implies shorter inter-arrival times between consecutive frames. Therefore timestamp precision must be proportional to the link speed. Latest version 1.0.x of libpcap provides 10-6 second native resolution, however pcap format supports a larger 2 x 32-bit timestamp value for each stored packets. On Gigabit Ethernet or faster networks, timestamp resolution that works in the microsecond domain may not enable us to precisely reproduce the time-domain relation between consecutive frames. Therefore overall analysis of the data transmission could drive to a false result. Independently from one other, five impact factors could directly bias the generation of timestamps: hardware architecture, NIC driver operation mode, clock source, kernel queue handler and the libpcap itself. In an idealized case generated timestamps are always converging and suitably close to the real arrival or transmission time of each frames so to conserve the original inter-arrival time values. For packet capturing with libpcap, it is assumed that timestamping performed when a frame is enqueued to the kernel’s input packet queue. Accordingly these timestamps represents the time moment when a frame reaches the input queue. Libpcap must retrieve these timestamps from the kernel. Timestamp resolution of network measurement applications must be increased according to the requirements of advanced high speed data networks. In our paper we are going to show an alternative libpcap-based solution that features nanosecond precision timestamping.

Towards estimating video QoE based on frame loss statistics of the video streams

This paper describes a unique path of estimating the Quality of Experience (QoE) for streaming video. Instead of following the widely researched idea of correlating video degradation with Quality of Service (QoS) metrics of the network, we propose to merely analyze the nature of packet losses. The majority of todays streaming video traffic is using compressed formats, where the videos are composed from a series of key, predicted, and bi-predictive frames (I frame, P frame, B frame, respectively). The content of these frames gets packetized, and sent over the network as video traffic. Losing packets that belong to I, P and B frames lead to QoE degradations of different severity. The degrading effect of these packet losses depend on the type of the video (two extremes can be slowly moving balloons versus fireworks in the sky), the type of frame, the volume of the lost packets, and further factors. The aim of this paper is to introduce a measurement and analysis method for correlating packet loss patterns of key, predicted and bi-predictive frames with objective quality metrics, such as the SSIM (Structural SIMilarity index). Furthermore, measurement results on the effect of such packet losses for various video types are also presented in this paper.

C-GEP: Adaptive network management with reconfigurable hardware

Carrying out network monitoring tasks remains a continuous challenge, partially because the line rate reaches and exceeds 100 Gbit/s. Besides the increasing data rate, the advent of programmable networks necessitates efficient solutions for supporting packet processing tasks in an adaptive way. Introducing a modification of a protocol or any new protocol in such a flexible infrastructure implies a novel management approach incorporating network monitoring equipment with reconfigurable architecture. The requirement for high throughput and high level of reconfiguration together put Field Programmable Gate Array (FPGA) technology into the focus of high performance networking. In this paper, we introduce a programmable, multi-purpose network platform called C-GEP that is based on a reconfigurable architecture. The system consists of two main building blocks: a high performance FPGA-based custom hardware platform and a firmware dedicated for network monitoring. We present the architecture focusing on the system-level integration of specific packet processors. The integration of processing building blocks into one high performance system has great challenges. These are primarily related to specific, limiting factors of system resources - which we discuss also in this paper.

C-GEP: 100 Gbit/s capable, FPGA-based, reconfigurable networking equipment

Programmable networking platforms are in the spotlight since the advent of SDN (Software Defined Networking). It is a great challenge to create such a platform - especially with reconfigurable hardware and line-rate capabilities reaching and exceeding 100 Gbit/s. These requirements together put FPGA (Field Programmable Gate Array) technology into the focus of high performance networking. In this paper, we introduce a highly flexible, programmable, multi-purpose networking platform, which is capable of hosting multiple 1 and 10 Gbit/s Ethernet interfaces - beside their 40 or 100 Gbit/s interface. The hardware of the introduced C-GEP platform is reconfigurable, even on-the-fly; due to the FPGA technology. C-GEP can host a wide range of high-speed network specific applications - including monitoring, switching and media conversion -, and it is aligned with the SDN principles. The system consists of two main building blocks: a high performance FPGA-based custom specific hardware platform and the firmware tailored to the actual task. The architecture is briefly introduced by its hardware and firmware setup, then some of the core functionalities, such as packet processing, filtering, and switching are presented.

A NetFPGA-based network monitoring system with multi-layer timestamping: Rnetprobe

In this work, we are about to introduce a high performance NetFPGA based network measurement system, called Rnetprobe that implements a dual, multi-layer timestamping method for QoS analysis. The multi-layer timestamping idea came from the demand to perform end-to-end active measurements that enable improved cross-layer analysis to evaluate QoS for time sensitive services. Accordingly, for each captured packet one timestamp is generated at MAC level, which is provided by our FPGA hardware module, while another timestamp is tagged to the packet on its processing path as soon as it becomes available to the OS kernels network stack. This second timestamp is software based and therefore reflects to the packet processing characteristics of the OS kernel. In this paper, we are introducing the mentioned multi-layer timestamping method and are evaluating its performance.

C-GEP: Platform demo for 100 Gbit/s network monitoring

As the line rate reaches and exceeds 100 Gbit/s, the usage of hardware-accelerated networking equipment is getting a natural choice when evaluating newly developed network-related features. The requirement for high throughput and high level of reconfiguration together put Field Programmable Gate Array (FPGA) technology into the focus of high performance networking. In this demonstration, we show some of the capabilities of a new, 100 Gigabit Ethernet Evaluation Platform, called C-GEP. Since the hardware is reconfigurable, the platform can host a wide range of high-speed network specific applications, and it is aligned with the Software Defined Networking (SDN) principles. Our demonstration contains two different implementations of 100 Gbit/s-capable applications: a traffic generator and a traffic monitor engine. Our aim here is to show the feasibility of C-GEP for high-speed networking evaluations.

Estimation of Media QoE without reference

QoE analysis of media services over IP networks (e.g., VoIP, IPTV, OTT, etc.) is a complex task, since it is based on the subjective human opinion. The subjective assessment requires a large number of volunteers to periodically provide feedback about the perceived quality of a service. Accordingly, there is a requirement for an objective alternative to the inherently subjective, perception-based service quality assessment approaches. The pivotal point of the presented solution is that the actual service quality is calculated from the momentary value of the QoS metrics measured at packet-level and it is presented on the MOS scale. Using an estimation-based approach, we are working on an objective NR type QoE estimation mechanism that needs no reference content from the sender side. To determine an estimation function, we performed combined (subjective and objective) assessments to build a reference dataset of 3-tuples of MOS, jitter and loss values. Applying polynomial regression and investigating canonical correlation, we are searching a low-degree two-variable polynomial to hash objective QoS metrics (jitter and loss, respectively) to the subjective MOS score of the service quality.

Line-rate packet processing in hardware: the evolution towards 400 Gbit/s

Network packet parsing and packet forwarding are general tasks for all routing devices. However, the requirement for line-rate packet processing, independently from the transmission technology, is a common demand against core network equipments. In this paper, we investigate programmable hardware architectures (i.e., Field Programmable Gate Array FPGA) as central building blocks of the data plane for state-of-the-art 100 Gbit/s network devices. We reveal the benefits and drawbacks of the available hardware architectures (such as Network Processors, Application-Specific Integrated Circuits (ASIC) and FPGAs, respectively). After showing the general packet processing steps on programmable hardware, we describe the problem space of line-rate packet processing in relation to the evolution of transmission technologies, i.e., 1, 10, 100 Gbit/s Ethernet and beyond. Moreover, we present design trade-offs, such as operational frequency, data path width and resource requirement, covering the 1 to 400 Gbit/s throughput range and we propose best practices for their hardware designs.