The advancement of artificial intelligence (AI) has the potential to revolutionize network communication. The use of advanced feature extraction in semantic communication can enhance transmission capacity. However, relying solely on unimodal visual characteristics derived from images through these approaches may result in inaccuracies in decoding under low signal-to-noise ratio (SNR) conditions. This paper introduces LaMoSC, a semantic communication system driven by large language models (LLMs) that uses multimodal features to reconstruct raw visual information, thereby improving transmission quality. The system proposes an LLM-driven multimodal fusion semantic communication framework, which aims to expand unimodal transmission systems and enhance generalization ability. LaMoSC has designed an end-to-end encoding-decoding network that integrates visual and textual multimodal feature inputs. The design deeply integrates modal features using the attention mechanism. Comprehensive comparisons with state-of-the-art baselines across various datasets demonstrate the robustness of the proposed method, particularly highlighting its superiority in low SNR conditions. LaMoSC outperforms Deep-JSCC and multi-level semantic aware communication system (MLSC) by 5.5% and 2.6%, respectively, under low SNR conditions, such as 4 dB. Its exceptional generalization capacity sets it apart from other methods.

It is important to minimize the energy consumption of large-scale, geographically distributed edge data centers (EDCs). While modern processing units (PUs) have energy-saving features like Dynamic Voltage and Frequency Scaling (DVFS) and Per-Core Power Gating (PCPG), optimization is still complex and requires a holistic approach. This paper presents a new decentralized, three-timescale, online optimization approach that enables multicore micro data centers (MDCs) to optimize their per-PU power states, per-enabled-PU voltage-frequency levels and offloading schedules at three different timescales. The key idea is that we employ multi-timescale Lyapunov optimization to decouple the energy minimization between workload scheduling and result delivery at a small timescale and PU configuration at large timescales. Another important aspect is that we apply the primal decomposition to decouple the PU configuration between a per-enabled-PU voltage-frequency level at an intermediate timescale and a per-PU power state at a large timescale. Experiments demonstrate that the proposed approach improves energy efficiency significantly by up to 4.5 times in our considered lightly loaded situations where DVFS alone does not work effectively, compared to existing benchmarks.

Cloud and fog computing are two emerging Internet-based collaborative technologies for big data analytics. An interesting question arising is whether the two technologies can resonate with significant gains of energy efficiency, especially in the case where advanced cloud platforms with Dynamic Voltage and Frequency Scaling (DVFS) are considered. This paper answers the question by formulating the optimization of a cloud with and without the assistance of fog, and deriving asymptotically optimal distributed solutions for the two cases. We also identify the critical condition under which fog computing helps the cloud to reduce the time-averaged queue lengths. The condition depends on the configurations of the fog, and the configurations of the connections between the fog and cloud. Extensive simulations exhibit good consistency with our analysis of the conditional benefits of fog computing. Evident from experimental datasets, the proposed fog-assisted cloud platform is able to increase the time-averaged energy efficiency by about 32.2%, and decrease the time-averaged queue length by around 37.0%, compared to a fog-coordinated counterpart where fog nodes only dispatch data and do not process the data.

Dynamic Voltage and Frequency Scaling, and Adaptive Body Biasing are increasingly adopted hardware techniques to improve energy efficiency of multi-core servers by adjusting reconfigurable supply and body bias voltages. Existing algorithms cannot fulfill the potential of the techniques because random variations of workload and background traffic can lead to coupling of voltage configurations over time and hinder effective real-time reconfigurations. This paper proposes a new approach which enables multi-core servers to optimize in real-time their configurations under random traffic variations. The approach asymptotically minimizes the time-averaged energy consumption of cloud computing while maintaining platform stability in a fully decentralized fashion. Lyapunov optimization is employed to decouple and separately optimize the voltage configuration, inter-and intra-server offloading schedules among servers and over time. The voltage configuration which is non-convex is proved to increasingly exhibit convexity with growing workloads. The optimality loss from the non-convexity asymptotically diminishes. Simulations show our approach dramatically reduces the power if the cloud is lightly loaded, or converts the power to processing capacity otherwise. Embraced by theoretical breakthroughs, the approach can potentially revolutionize cloud computing.

This paper proposes a new distributed real-time optimization for MapReduce-style framework in emerging cloud platforms supporting dynamic speed scaling functions. Distinctively different from the existing MapReduce parallelism strategy with fixed specific data chuck sizes, the new approach is able to dynamically dispatch input data of adequate sizes and synthesize interim processing results according to applications and the state of data centers (DCs). The key idea is to decouple the optimizations of data dispatching, processing, and result aggregation without loss of optimality, by employing stochastic optimization techniques. Another important aspect is that we optimize the subproblem of data processing to leverage the energyand speed-configurability of DCs, by optimally deciding the number of servers to be activated at every DC and the CPU speeds of the activated servers. Evident from extensive simulations, the proposed approach is able to increase the throughput-cost ratio by up to 94.3%, as compared to existing initiatives, and substantially improve the throughput in the case of high-rate data streams.

Modern Internet of Things (IoT)-aware business processes include various geographically dispersed sensor devices. Large amounts of raw data acquired from sensors need to be regularly transmitted to the targeted processes in enterprise data centers, resulting in a significant increase in network load and latency. It is necessary to execute such processes in a distributed way. The existing work has proposed different algorithms to partition a given process for distributed execution; however, they cannot satisfy the decentralized nature of IoT-aware business processes. Moreover, up to now, there is few work that studies uncertain optimal deployment problems in which traffic data for guiding subsequent deployment derives from experts' empirical knowledge. This paper proposes a novel location-based fragmentation algorithm and α-optimal deployment solution to deal with the mentioned problems, where α is the given confidence level. A hardware-in-the-loop simulation platform based on NS-3 was built. Based on this platform, an integrated monitoring process was deployed that ran on different virtual computers, and process fragments communicated with each other via a simulated network. The experimental results show that the proposed approach can reduce network traffic and round-trip time.