Accurate forecasting of large-scale cloud server workloads is a critical challenge in modern microservice-based data centers, where complex inter-node dependencies, temporal latency effects, and multi-dimensional fluctuation patterns often undermine the effectiveness of traditional prediction models. To address these challenges, we present Collaborative Fluctuation-aware Learning based on Long Short-Term Memory (CFL-LSTM), a unified framework that combines dynamic correlation mining, causal temporal-spatial alignment, and gated attention-based multi-channel LSTM fusion. The dynamic correlation mining module extracts fluctuation-driven features and quantifies inter-node relationships with polarity-aware delay compensation. The temporal-spatial alignment module corrects auxiliary sequence misalignments to preserve causal consistency. The gated attention fusion mechanism adaptively integrates multi-node representations for fine-grained forecasting. Comprehensive experiments conducted on both the Alibaba Cluster Trace and GAIA datasets demonstrate that CFL-LSTM consistently outperforms competitive statistical and deep learning baselines across multiple error metrics. By incorporating attention-guided feature selection and polarity-aware alignment into a scalable end-to-end architecture, CFL-LSTM offers a novel and deployable solution for robust and accurate cloud workload prediction.

The rise of serverless computing has prompted service providers to offer fine-grained function configurations, such as allocating memory in 1MB increments, to support a wide range of function types and adapt to fluctuating workloads. Implementing a Federated Learning (FL) aggregator using serverless functions significantly minimizes the cost associated with idle waiting time. However, for serverless Federated Learning as a Service (FLaaS) providers, determining cost-effective function configurations is challenging. This is primarily because predicting future workload patterns, such as the number of parties involved, is difficult and often inaccurate. In this paper, we use Alibaba Cloud Function Compute as a case study to analyze serverless pricing and address key challenges in optimizing function execution costs within the context of logic-tree-based FL aggregation. We then propose an offline optimal algorithm based on dynamic programming, assuming the future workload pattern is known, as a benchmark. To accommodate online workload arrivals, specifically the variable number of parties in each round, we propose a deep Reinforcement Learning based adaptive logic Tree Generation method (RLTreeGen), which dynamically constructs cost-efficient aggregation architectures for each round of FL training. We evaluate its cost performance using real-world traces and Alibaba Cloud Function Compute’s pricing model. The experimental results demonstrate that RLTreeGen outperforms other online benchmark algorithms and closely approaches the offline optimal in all tested scenarios.

Infrastructure-as-a-Service (IaaS) cloud platforms offer resources with diverse buying options. Users can run an instance on the on-demand market which is stable but expensive or on the spot market with a significant discount. However, users have to carefully weigh the low cost of spot instances against their poor availability. Spot instances will be revoked when the revocation event occurs. Thus, an important problem that an IaaS user faces now is how to use spot instances in a cost-effective and low-risk way. Based on the replication-based fault tolerance mechanism, we propose an online termination algorithm that optimizes the cost of using spot instances while ensuring operational stability. We prove that in most cases, the cost of our proposed online algorithm will not exceed twice the minimum cost of the optimal offline algorithm that knows the exact future a priori. Through a large number of experiments, we verify that our algorithm in most cases has a competitive ratio of no more than 2, and in other cases it can also reach the guaranteed competitive ratio.