This paper presents a comprehensive analysis of diffusion model serving challenges in production cloud environments. We examine the unique computational patterns and resource requirements that distinguish diffusion model serving from traditional ML workloads, revealing fundamental systemlevel challenges from their multi-stage pipeline architectures. Our analysis is based on a dataset collected from a commercial image generation service processing 3.5 million requests across 300+ GPUs of production operation.

In this paper, we present ROCK, a novel system for efficiently serving thousands of LoRA adapters for multimodal models in cloud environments. Through extensive analysis of production workloads, we identify key challenges in current cloud-based image generation services: extreme request burstiness (up to$90 \times$normal rates), heterogeneous task characteristics, and inefficient adapter management that wastes 40 % of GPU memory and increases delays by$3 x$during peak times. ROCK addresses these challenges through a three-layer architecture that decouples hardware, adapters, and requests. Our system features dynamic heterogeneous queues that match tasks to appropriate resources based on multidimensional feature vectors, and a multilevel orchestration framework that intelligently manages adapter placement across heterogeneous storage. Experiments on a 64-GPU testbed demonstrate that ROCK reduces average response latency by$16-26 \%$, and achieves an 84.1 % cache hit rate for LoRA adapters-outperforming traditional approaches while reducing adapter update frequency by up to 77 %.

Although the emerging serverless paradigm has the potential to become a dominant way of deploying cloud-service tasks across millions of mobile and IoT devices, the overhead characteristics of executing these tasks on such a volume of mobile devices remain largely unclear. To address this issue, this paper conducts a deep analysis based on the OpenFaaS platform—a popular open-source serverless platform for mobile edge environments—to investigate the overhead of performing deep learning inference tasks on mobile devices. To thoroughly evaluate the inference overhead, we develop a performance benchmark, namedESBench, whereby a set of comprehensive experiments are conducted with respect to a bunch of simulated mobile devices associated with an edge cluster. Our investigation reveals that the performance of deep learning inference tasks is significantly influenced by the model size and resource contention in mobile devices, leading to up to$3\times$degradation in performance. Moreover, we observe that the network environment can negatively impact the performance of mobile inference, increasing the CPU overhead under poor network conditions. Based on our findings, we further propose some recommendations for designing efficient serverless platforms and resource management strategies as well as for deploying serverless computing in the mobile edge environment.

As Large Language Models (LLMs) continue to advance, their parameter sizes are growing exponentially-far outpacing hardware capabilities. This widening gap necessitates distributed computing through pipeline parallelism for efficient inference. However, the uneven distribution of requests across pipeline stages creates significant performance bottlenecks in real-world deployments. To address this challenge, we presentPlanck, a performance optimization framework specifically designed for distributed LLM inference.Planck implements fine-grained control through two key mechanisms: a progressive SLO allocation strategy that dynamically adjusts time constraints based on workload patterns, and stage-specific performance controllers that prevent bottlenecks before they cascade through the system. By intelligently balancing resources across pipeline stages,Planck effectively eliminates queue buildup-essentially preventing traffic congestion before it forms. Evaluation using diverse workloads in real cloud environments demonstrates thatPlanck reduces P99 tail latency by up to 18% and decreases the longest queue lengths by as much as 47.8% across pipeline stages, significantly improving both system responsiveness and resource utilization.

Pipeline parallelism is an important strategy for improving inference performance in Large Language Models (LLMs). However, we find that different stages of LLM pipelines exhibit distinct performance and request characteristics, posing challenges to system performance in online inference scenarios. To address this issue, we propose Planck, a performance optimization framework tailored for LLM pipeline inference. By balancing request traffic, queue length, and execution time at each stage, Planck introduces a progressive SLO (Service Level Objective) allocation method and a stage instance performance controller. Planck fine-grainedly allocates SLOs to each pipeline stage and dynamically adjusts according to request distribution to control queue length. Through optimizing queue lengths across different stages of the model pipeline, Planck effectively reduces waiting time and tail latency. Evaluations conducted on a real cloud cluster using diverse workloads demonstrate that Planck effectively reduces P99 latency and queue length for each pipeline stage.

With the advancement of new industrialization, production processes or scenarios characterized by automation and intelligence require monitoring by more powerful visual inspection tools to maintain production continuity, safety, and efficiency. Spike vision sensors, with their high temporal resolution and high dynamic range, can capture subtle changes of high-speed moving objects in complex industrial environments, providing precise data support for equipment maintenance, fault detection, and condition monitoring. However, in industrial scenarios, factors such as mechanical vibrations, electromagnetic interference, and extreme strong light affect the quality of sensor data, leading to higher noise levels and increased probability of data loss and errors. To address this, we propose a multi-stage processing method for spike data processing and image reconstruction. It first performs preliminary feature extraction and denoising through an anti-noise feature encoder. Building on this, it integrates features and enhances consistency through spatiotemporal information utilization. Finally, it reconstructs the images through an fine-tuned spike reconstruction model. Experimental results show that our method can effectively reduce noise and improve data quality, achieving good performance on industrial datasets. This provides higher quality data support for downstream tasks such as predictive maintenance and high-speed production inspection in industrial scenarios.