Design and analysis of ultra-low latency fronthaul and backhaul networks for 5g

Resumen

Conventional communications between human mobile users are not the only type of activities envisioned for post-5G and future 6G networks. Traditional Human-to-Human (H2H) communications are now joined by Human-to-Machine (H2M), and Machine-to-Machine (M2M) communications, in a Tactile Internet world. 6G is expected to put an end to smartphone-centric era, introducing new system paradigms. New services and applications such as telepresence, remote control of haptic machines, online schools/working, remote surgery, Multi-access Edge Computing (MEC) are envisioned to be part of the future daily life. Cloud-based services (including the cloud-based signal processing of cellular signals), video traffic, enterprise applications, and virtualization have led to a rise of the demand in metropolitan areas. The so-called Total Cost of Ownership (TCO) applied to cellular networks, comprises Capital Expenditure (CAPEX) and Operational Expenditure (OPEX). The former refers to the cost of construction of a network. Network planning, site acquisition, purchase of hardware and software, installation of the powering and cooling are some examples that would be defined as CAPEX. The OPEX, refers to the costs that arise while operating the network. Among these, it is worth highlighting the cost of electricity, site rental, maintenance, etc. It seems obvious that, under traditional deployment schemes, the CAPEX and OPEX increase importantly as more stations are added to the network. This is due to the fact that bases stations represent one of the most expensive elements in the network. Moreover, many authors claim that a huge percentage (~ 72%) of the total power consumption required to operate is at the cell sites. Adding new macro base stations in such traditional scenarios is expensive or even an impossible task. Most of the times, these base stations require a huge capital expenditure (housing, signal processing equipment, adequate powering infrastructure, etc). Traditional cellular architectures are becoming exhausted and unable to cope with the requested performance and new approaches such as the Cloud Radio Access Networks (c-ran) appear as natural replacements. This a suggests the separation between the radio and the processing equipment that now will be located at different places of the network. However, all the implementation details about how to meet the bandwidth, latency, jitter or even how to measure these parameters remain open. One of the goals of this thesis is the advance in the State of the Art regarding the characterization of the aggregated fronthaul traffic with the tools of queueing theory. Following an incremental contribution scheme, we review the estimations provided by classical queueing models and explored others that are not so well-known. We firstly assess real world traditional deployments like ringstar topologies implemented with Ethernet switching equipment. Namely, we study a real-world unidirectional and bidirectional ring-star topology in terms of feasibility to transport delay constrained fronthaul data. We derive the theoretical expressions for the propagation and queueing delays, assuming a G/G/1 queueing model. Then, we characterize the mixture of fronthaul flows, studying the rate of convergence of the arrivals squared coefficient of variation. Particularly, we look into the aggregation of functional Split B flows in the same fronthaul network. We investigate the accuracy of the theoretical estimations for different load states in the network. The main conclusions are that the simulations demonstrate that a centralized implementation is feasible if we adopt the concept of functional split in the baseband processing chain. Real-world deployments are achievable using Split B and the rate of convergence of the arrivals squared coefficient of variation (C^2) is different depending on the packet size. When aggregating 150 fronthaul flows, for the same total data, C^2 can be reduced by a factor of 4 by using 3-packet bursts, instead of 12. Also, theoretical estimations given by the G/G/1 queueing model are close to the simulator outputs for system loads >= 0.4, when more than 50 flows are combined. Moreover, the average aggregated queueing delay is far from exceeding the envisioned restriction of 50 μs. The worst average queueing delay is 10 times lower in the bidirectional ring case. However, the study of the percentiles of the queueing delay remains pending and serves as a motivation for the next contribution. Additionally, more delay sources (switching, packet processing, etc) shall be added in ulterior studies. Then, the G/G/1 queueing model is further extended with the Kingman’s Exponential Law of Congestion, paying special attention to the percentile values of the queueing delay. This expands the previous study by proposing high queueing delay percentiles as the key metric for the fronthaul network dimensioning. Namely, the aggregation of ecpri flows (Splits Iu and IId) in a packet aggregator is explored for different values of the radio channel’s bandwidth. We derive a set of rules for Ethernet-based fronthaul network dimensioning, using high delay percentiles as the key design metric, instead of conventional average delays often seen in the literature. Simulations revealed that the Kingman’s Exponential Law of Congestion provides accurate estimations for the queueing delay percentiles in the case of aggregating a number of ecpri fronthaul flows. Namely, tests were made for functional splits Iu and IId. We show how the 90-th or 99-th delay percentiles are substantially higher than the average values. In some cases, these are between one or two orders of magnitude higher. Thus, the dimensioning requires typically larger overprovisioning factors of capacity. Additionally, we observe that the transmission of multiple legacy legacy 20 Mhz LTE channels using such functional split can be realized with 40 Gb/s transponders guaranteeing 99-th delay percentiles below 9 μs. We conclude that conventional 10G, 40G and 100G transponders can cope with multiple legacy 10-20 Mhz radio channels with worst-case delay guarantees. Conversely, scaling to 40 and 100 Mhz channels will require the introduction of 200G, 400G and even 1T high-speed transponders. Next, we further develop this incremental study of the aggregation of fronthaul streams. The use of extreme latency percentiles is explored as a useful tool for the design of fronthaul networks. We suggest that theoretical worst-case delay-based designs could be too pessimistic in scenarios where achieving the maximum feasible BBU-RRH distance is required. Consequently, the proposal is to use very high packet delay percentiles as an alternative to the maximum theoretical delay in order to stretch the range of the fronthaul links at the expense of a higher frame loss ratio (FLR), within the limits established by eCPRI and IEEE 802.1cm. Functional splits Iu / IId from the eCPRI standard are analyzed making used of the G/G/1 and N*D/D/1 queueing models, which are also compared to the simulation option in terms of the tightness of their predictions in the extreme latency percentiles scenarios. The results support that an extra latency budget may be available for propagation, while meeting the requirements established by IEEE 802.1cm. Particularly, while the G/G/1 model and simulations can produce satisfactory results for moderate percentiles, both saturate in the context of high system loads and extreme percentile values. Only the N*D/D/1 queue is appropriate for the extreme percentiles. Conversely, the N*D/D/1 queue is able to model the behavior of a packet-switch fronthaul aggregator using the eCPRI standard for 5G New Radio (NR) fronthaul streams and can be used as a tool to dimension the length of the links. A better modeling of these percentiles enables us to comply with the defined FLR in IEEE 802.1cm by interpreting the gap between this estimation and maximum worst-case delay as an extra delay budget. This extra budget becomes more relevant at high loads. Experiments revealed that additional propagation delay budget can be gained at 100 Gb/s under the appropriate conditions. In general, the extracted rule of thumb is that the higher the load, the more extra latency budget we can obtain, proportionally to the maximum worst-case delay, since the gap between the percentiles and the maximum theoretical queueing delay becomes wider. Experiments show that the fronthaul links’ lengths can be increased by 60% and 10% for 50 Mhz and 100 Mhz NR channels, respectively, while keeping the latency budget and frame loss ratio within the IEEE 802.1cm limits. Alternatively, this extra budget could be used to aggregate more RRHs at the same aggregation point, or we could even think about dynamically switching to more resource-demanding functional splits on certain RRHs, if needed. Recent research efforts focus on the development of new photonic technologies, enabling capacities of Tb/s for the metropolitan area network. Sliceable Bandwidth Variable Transceivers (S-BVTs) appear as a promising technology to achieve this goal. This thesis also studies some applications and use cases of new optical multi-Tb/s equipment for future metropolitan area networks. Namely, the dimensioning of edge data center resources from the perspective of achieving the blocking probabilities as low as 10^-6 is explored. Two different scenarios/topologies are analyzed: the overflow over a paired same-level data centers and the overflow over a centralized site. We present a methodology to compare both schemes based on teletraffic engineering. The results show that a proper distribution of computing resources in the centralized overflow approach can outperform the costs of a distributed strategy, requiring fewer processors and much smaller data centers at the central office. The SDN-based dynamic multi-destination Tb/s capability of EU PASSION project S-BVTs seems particularly suitable to deal with the computing demands overflown from the edge data centers. Moreover, multiple strategies are feasible by using a single transceiver per node and an automatically switched optical network. For the particular conditions assumed in this study, a total saving of 8% can be achieved in terms of the amount of computational resources that must be provided to reach blocking probabilities of 10^-3. However, both the gap between the cost of the analyzed approaches as well as the threshold that makes one scenario a better choice than the other are hugely dependent on how the blocking probabilities are divided between tiers. The savings are higher when the backup data center has a lower blocking probability than the local one. Whenever the total blocking probability is shared equally between the local and the central data centers, the savings are never above 1%. Next, different strategies to implement cost-effective CDN caching are presented. Carrying the backup traffic from one data center to another with a permanent optical circuit based on Fixed Transceivers (FT) features low utilization and no statistical multiplexing gain on the path, which makes the backup network resources costly. Several strategies to implement cost-effective CDN caching are presented to solve this. Both fixed and bandwidth variable solutions are compared, with the study of three different scenarios from the perspective of availability. The main findings are that the only scalable solution for high service availability of low latency CDN caching is making other man data centers within the target latency budget backup other data centers of usually lesser reliability. S-BVTs can be key devices to improve backup-network scalability in terms of IT resources and transceivers, thanks to their capability to adapt to the actual traffic demand and to obtain multiplexing gains at the optical layer. Assuming that the three compared scenarios are dimensioned to give service to 40 metropolitan areas, Scenarios B and C provide a ~ 46% reduction in the number of resources in the local and centralized data centers. This is possible thanks to the statistical multiplexing. However, the number of needed transceivers does not scale well with the number of edge CDNs for Scenarios A and B. This is not the case for Scenario C, where the flexibility of the S-BVT helps reduce this requirement. From the wavelength occupancy point of view, the S-BVT is the best choice as it can fit the real load of the network with finer granularity (50 Gb/s). Regarding the latency of each solution, even though that Scenarios B and C suffer from a higher delay, 125 μs is not a heavy burden for most applications. Finally, this thesis investigates the convergence of legacy and future radio access technologies in the context of a heterogeneous post-5G and 6G environments. We address the characterization of the fronthaul traffic so that we may support its coexistence with other types of traffic and services. A Heterogeneous Cloud Radio Access Network (H-CRAN) serves as the reference network comprising all the envisioned traffic types, services, and data rates. We investigate the resource allocation and association of users with different delay requirements in a shared-backhaul fiber-wireless (FiWi) enhanced Heterogeneous Cloud Radio Access Network (H-CRAN) with Multi-access Edge Computing (MEC) and offloading. A network simulator and its building blocks are described and presented as the tool to assess the performance of the proposed solution. This description includes the basics and the building blocks of the simulator (channel and interference models) as well as the definition of many concepts such as the heterogeneous networks, the construction of the coverage maps, the uplink and downlink decoupling, etc. Then, a global optimization problem is proposed for the entire network, supporting three access technologies: LTE-A, WiFi, and Cloud Radio Access Network nodes. Additionally, a decentralized algorithm is proposed to orchestrate all the elements of the network and compared to a more traditional solution scenario. In this part of the thesis, several migration questions are answered, regarding the integration of legacy and future technologies in a single network. We show how a FiWi network topology based on a 25 Gbps EPON is able to maximize the utility of all users while meeting the delay requirements of different services, with the appropriate orchestration. Furthermore, we explain how to decompose the global optimization problem via a full dual decomposition technique. This enables divide the orchestration problem in smaller chunks that can be easily handled by the elements of the deployment (i.e., cells, users, etc), which leads to a flexible and seamless operation of the network. Due to that, the distributed algorithm is computationally simple as it only requires broadcasting the Lagrange multipliers that arise while solving the optimization problem. These, work as a sort of price indicator, transmitting information about the state and needs of base stations and users at any particular time. Simulations show that this solution outperforms the classical received-power criteria in terms of average delay, power consumption and delay thresholds compliance. Namely, the results suggest that the distributed approach can achieve up to an 80% improvement in terms of average delay. More importantly, the decentralized solution is able to organize the user association and network resources in the appropriate way so as to comply with the delay requirements. This is not the case for the traditional baseline scenario and orchestration solution. As a side effect of this improved orchestration, the energy consumption reduction obtained using the proposed approach ranges from 28-56%, for Human-to-Machine traffic, to 30-50% for Human-to-Human traffic. This is due to the fact that, thanks to the optimized operation of the network, the user devices need to turn their radios on less time than usual to complete their communication processes.