Search

Low-Rank Adaptation (LoRA) has garnered increasing attention for effectively fine-tuning large language models (LLMs) with limited resources. Nonetheless, conventional approaches that cater to multiple LoRA models independently lead to redundant computations and suboptimal GPU utilization. This study tackles these inefficiencies by presenting Dynamic Operator Optimization, a sophisticated automated optimization methodology crafted to dynamically enhance the Segmented Gather Matrix-Vector Multiplication (SGMV) operator according to specific contexts. The distinctive design of SGMV facilitates the batching of GPU operations for diverse LoRA models, resulting in a notable enhancement in computational efficiency. The strategy exploits a Search Space Constructor to construct a structured search space, segmenting the program space into overarching structural outlines and intricate implementation particulars to ensure a varied and adaptable operator implementation. Moreover, an Optimization Engine fine-tunes these implementations through evolutionary search driven by a performance estimation cost model. This progressive optimization procedure ensures that SGMV implementations can dynamically adjust to varying scenarios to uphold superior performance. The findings illustrate that our design can elevate throughput by up to 1.46 times in cutting-edge multi-tenant LoRA deployments.

Over the past decade, network operators worldwide have regarded passive optical networks (PONs) as a successful access network system and the preferred technology for fiber-to-the-home/building/curb/distribution-point (FTTx) architecture. However, implementing the state-of-the-art PONs requires a high initial capital expenditure (CapEx) to deploy fiber in the field and acquires the necessary active devices. Thus, new architectures need to be developed to share the physical architecture and enable the multi-tenancy capability in a flexible way to satisfy different requirements from the operators and users. This research proposes to integrate the next-generation TWDM-PON with software-defined network (SDN) and the network function virtualization (NFV) to provide highly enhanced form of virtual unbundling of broadband access networks, enabling cost sharing and the dynamics of a competitive landscape. The simulation results show that the proposed architecture and mechanism can provide better performance in multi-tenancy condition, while guaranteeing the quality-of-service requirements in terms of the mean packet delay, jitter, throughput, and packet loss ratio.

The multi-tenancy concept in cloud data center (DC) networks paves the way towards advancements and innovation in the underlying infrastructure such as network virtualization. Multicast routing is essential in leveraging multi-tenancy to its full potential. However, traditional IP multicast routing is not suitable for DC networks due to the need to support a massive amount of multicast groups and hosts. State-of-the-art DC multicast routing approaches aim to overcome these scalability issues by, for instance, taking advantage of the symmetry of DC topologies and the programmability of DC switches to compactly encode multicast group information inside packets, thereby reducing the overhead resulting from the need to store the states of flows at the network switches. Although these approaches scale well with the number of multicast groups, they do not perform well with group sizes and, as a result, yield substantial traffic control overhead and network congestion. In this article, we present Bert, a scalable source-initiated DC multicast routing approach that scales well with both the number and size of multicast groups through the clustering of multicast group members where each cluster employs its own forwarding rules. Compared to the state-of-the-art approach, Bert yields much less traffic control overhead by significantly reducing packet header sizes and eliminating switch memory usage across the switches.

Nowadays, most applications hosted on public cloud data centers (DCs) disseminate data from a single source to a group of receivers for service deployment, data replication, software upgrade, etc. For such one-to-many data communication paradigm, multicast routing is the natural choice as it reduces network traffic and improves application throughput. Unfortunately, recent approaches adopting IP multicast routing suffer from scalability and load balancing issues, and do not scale well with the number of supported multicast groups when used for cloud DC networks. Furthermore, IP multicast does not exploit the topological properties of DCs, such as the presence of multiple parallel paths between end hosts. Despite the recent efforts aimed at addressing these challenges, there is still a need for multicast routing protocol designs that are both scalable and load-balancing aware. This paper proposes $\mathsf {Ernie}$ , a scalable load-balanced multicast source routing for large-scale DCs. At its heart, $\mathsf {Ernie}$ further exploits DC network structural properties and switch programmability capabilities to encode and organize multicast group information inside packets in a way that minimizes downstream header sizes significantly, thereby reducing overall network traffic. Additionally, $\mathsf {Ernie}$ introduces an efficient load balancing strategy, where multicast traffic is adequately distributed at downstream layers. To study the effectiveness of $\mathsf {Ernie}$ , we extensively evaluate $\mathsf {Ernie}$ ’s scalability behavior (i.e., switch memory, packet size overheads, and CPU overheads), and load balancing ability through a mix of simulation and analysis of its performances. For example, experiments of large-scale DCs with 27k+ servers show that $\mathsf {Ernie}$ requires a downstream header sizes that are $10\times $ smaller than those needed under state-of-the-art schemes while keeping end-host overheads at low levels. Our simulation results also indicate that at highly congested links, $\mathsf {Ernie}$ can achieve a better multicast load balancing than other existing schemes.

In 5G slice networks, the multi-tenant, multi-tier heterogeneous network will be critical in meeting the quality of service (QoS) requirement of the different slice use cases and in reduction of the capital expenditure (CAPEX) and operational expenditure (OPEX) of mobile network operators. Hence, 5G slice networks should be as flexible as possible to accommodate different network dynamics such as user location and distribution, different slice use case QoS requirements, cell load, intra-cluster interference, delay bound, packet loss probability, and service level agreement (SLA) of mobile virtual network operators (MVNO). Motivated by this condition, this paper addresses a latency-aware dynamic resource allocation problem for 5G slice networks in a multi-tenant, multi-tier heterogeneous environment, for efficient radio resource management. The latency-aware dynamic resource allocation problem is formulated as a maximum utility optimisation problem. The optimisation problem is transformed and the hierarchical decomposition technique is adopted to reduce the complexities in solving the optimisation problem. Furthermore, we propose a genetic algorithm (GA) intelligent latency-aware resource allocation scheme (GI-LARE). We compare GI-LARE with the static slicing (SS) resource allocation; the spatial branch and bound-based scheme; and, an optimal resource allocation algorithm (ORA) via Monte Carlo simulation. Our findings reveal that GI-LARE outperformed these other schemes.

The new fifth generation (5G) era has transformed previous mobile generations into fast, smart networks that will be more responsive and customizable. With network slicing, 5G networks can be dynamically adapted to the different needs of specific vertical industries. This capability has opened the opportunity to new business models whereby infrastructure owners can monetize their investment by leasing resources to third parties (i.e. tenants). In this respect, a challenging task for the owner of the radio access network infrastructure (i.e. the network provider) is the spectrum planning of multi-tenant scenarios. This paper proposes different alternatives of hiring capacity to the provider as well as a set of spectrum planning strategies, each giving a certain degree of flexibility to allocate resources per tenant. These strategies are evaluated in a 5G small cell multi-tenant network through snapshot-based simulations. The performance of the strategies is assessed in terms of scalability, spectrum isolation, utilization and efficiency.

Network slicing is a key feature of forthcoming fifth generation (5G) systems to facilitate the partitioning of the network into multiple logical networks customized according to different business and application needs. Network slicing is a fundamental capability for enabling a cost-effective deployment and operation of 5G, as it allows the materialization of multi-tenant networks in which the same infrastructure is shared among multiple communication providers, each one using a different slice. This paper proposes a Markovian approach to characterize the resource sharing in multi-tenant scenarios with diverse guaranteed bit rate services by considering a slice-aware admission control policy. After describing the Markov model and its implementation and discussing its suitability, the model is applied to study the performance attained in a scenario with two different slices, one for enhanced mobile broadband communications and the other for mission critical services. The system is analyzed under standard and disaster situations, thus illustrating the capability to properly manage the different multi-tenant and multi-service traffic loads.

The decoupling of infrastructure from services, which has been so far a mainstream paradigm in the computational and storage domain, is now becoming a paradigm also for mobile networks. Indeed, 5G must provide a variety of services with very diverse requirements, such as throughput, latency, or reliability, and decoupling infrastructure from service provisioning allows to deal with such heterogeneity. In this context, a new business model, involving two different stakeholders, Infrastructure Providers and Service Providers, has emerged. Besides addressing the technical issues, it is also important to study the economic feasibility and behavior of such new paradigm and the techno-economic interactions among the different stakeholders that play different roles in the mobile network market. In this paper, we propose a multi-leader multi-follower variant of the Stackelberg game to model the considered environment. The proposed model is then fed with realistic data and used to analyze the system behavior and the impact of the technological features of the stakeholders on their competitiveness.