Ensemble Temporal Prediction Model-as-a-Service (ETP-MaaS) has become crucial in fields like financial modeling and cloud monitoring. Existing solutions fail to co-optimally address a two-fold challenge of dynamic collaboration and heterogeneity, treating models as independent entities and employing simplistic worker allocation rules. However, at the model level, data volatility means that optimal performance requires identifying and weighting constantly shifting subgroups of base models, not just individual ones; at the system level, these model groups must be efficiently mapped to a pool of heterogeneous and dynamically available workers. To this end, we introduce WIET, an efficient ETP-MaaS system that co-optimizes model weighting and worker allocation. For adaptive weighting, WIET identifies evolving group behaviors among base models and propose a novel group temporal locality-enhanced weighting method. Additionally, WIET develops an efficient, multi-dimensional worker allocation method powered by hybrid heuristic optimization, effectively reducing bottlenecks and resource waste. Experiments show WIET consistently outperforms state-of-the-art methods in terms of accuracy, latency, and resource usage across various workloads and tasks.

Spin-Transfer Torque Magnetic Random-Access Memory (STT-MRAM) has emerged as a compelling candidate for embedded memory due to its non-volatility, high endurance, and low power consumption. However, manufacturing process limitations result in low tunnel magnetoresistance ratios (TMR), leading to narrow read windows. Therefore, high-performance trimming is required, yet existing self-trimming schemes face challenges in balancing trimming time, accuracy, and flexibility. To address these issues, this paper proposes a Dynamic Stepsize Single-write Pattern Trimming (DSPT) algorithm that uses dynamic step sizes and a single-write pattern to rapidly lock the read window boundary range, followed by binary search for precise boundary determination, thereby reducing trimming time while maintaining high accuracy. A parallel variant, DSPT_P, is further proposed, allowing each sense amplifier (SA) to shift gears independently in advance and introducing an Address Reliability Limit (ARL) for result assessment. Finally, a BuiltIn Self-Test (BIST) architecture is proposed to implement the proposed trimming schemes.

Recently, AI inference services have adopted ensemble architectures, which are widely recognized and used for their advanced performance. However, the existing ensemble inference services are mainly created and managed using the platform-as-a-service model with the static ensemble service architecture and rigid persistent resource allocation. This makes the highly heterogeneous inference requests rely on a fixed combination of basic learners and manual homogeneous resource management, resulting in insufficient precision, waste of resources, and high management costs. Serverless computing, represented by Functions-as-a-Service (FaaS), realizes transparent on-demand resource allocation to developers, which is suitable for ensemble inference services. Therefore, we propose a serverless proactive elastic scheduling solution PESEI for ensemble inference services. First, a two-level hierarchical dynamic ensemble service framework with joint model precision and overhead sensing is proposed to ensure inference precision while improving cost efficiency; second, a proactive elastic resource allocation algorithm with dynamic sensing of workload patterns is proposed to optimize the quality of service and cost efficiency for heterogeneous base learners; based on this, a prototype system is designed and developed to support the autonomous management of ensemble inference services, which implements the benign combination and adaptation of dynamic service architecture and elastic resource management. Real cluster experiments on public datasets demonstrate the effectiveness and robustness of PESEI, providing a new solution for building serverless ensemble inference services.