AI Cuts Educational Storage Latency by 71% Using Teaching Cycles

Researchers from China have developed an intelligent storage optimization method for educational lakehouse architectures that leverages the predictable periodicity of academic workloads to significantly reduce data access latency and storage costs. Published in Discover Computing, the study introduces a system that integrates business semantics, such as academic calendars and course schedules, into storage management decisions. By treating instructional cycles as a natural temporal scale for data temperature, the proposed method enables proactive data migration across hot, warm, and cold storage tiers. The approach addresses a critical gap in current educational information systems, which often rely on static configuration strategies that fail to account for the multi-scale periodic access loads driven by pedagogical operations.

The research team, led by Geng Wei, Geng Xingyuan, and Hu Min, identified that traditional data warehousing and data lake technologies struggle to accommodate the growing scale and diversity of educational data. While lakehouse architectures combine the transactional guarantees of data warehouses with the cost-efficiency of data lakes, existing implementations lack mechanisms to perceive and predict workload variations driven by domain-specific business rules. This leads to performance bottlenecks during peak access periods, such as exam weeks, and resource waste during low-access periods, such as vacations. The study aims to formalize instructional cycles as domain knowledge and embed them into storage scheduling to achieve a dynamic balance between performance and cost.

The proposed method consists of three core modules that form a closed-loop feedback system. The first module, the Multi-Scale Teaching Period Load Model (MSTP-LM), encodes business information such as academic calendars and course schedules into machine-interpretable multi-granularity temporal features. Through semantic encoding and periodic embedding, the model translates structured operational information into a continuous, quantitative heat scale. This allows the storage system to perceive the progression of teaching activities and anticipate evolving trends in data value, effectively mapping discrete instructional events into a time-varying data heat function.

The second module, the Multi-Factor Dynamic Value Evaluation Model (MFD-VE), integrates access frequency, recency, and pedagogical relevance to quantitatively assess the time-varying characteristics of data across different educational stages in real time. Unlike traditional data temperature models that rely solely on general indicators like access frequency, MFD-VE captures domain-specific semantic characteristics. It employs an adaptive weighting mechanism that adjusts the importance of these factors based on the current instructional stage, ensuring that data strongly associated with active teaching activities is not erroneously demoted following periods of low access, such as holidays.

The third module, the Intelligent Storage Migration Decision Framework (DQN-SM), utilizes Deep Q-Networks to enable dynamic and adaptive migration of data across hot, warm, and cold storage tiers. This framework takes the features from MSTP-LM and the value metrics from MFD-VE as inputs, using a dual-stream network architecture to process heterogeneous data. By incorporating curriculum learning and reward shaping strategies, the DQN-SM framework balances access performance with long-term archival costs, allowing for proactive pre-warming of data before peak access periods arrive. The system also includes safety constraints and rollback mechanisms to prevent unreasonable migration operations from disrupting normal business operations.

The experimental evaluation was conducted using a dataset sourced from the academic administration platform of an applied-sciences university. The dataset comprises authentic operational logs spanning from September 2023 to December 2025, covering over 200 courses and more than 14,000 students. It involves over 800,000 data objects with a total volume of approximately 50 TB. The logs reveal distinct periodic patterns, with course-selection and examination weeks receiving significantly higher daily accesses than regular weeks, while vacation periods see a sharp decline. This empirical basis validates the necessity of teaching-cycle awareness in storage optimization.

Experimental results demonstrate that the MSTP-LM reduces the Mean Absolute Percentage Error (MAPE) in load forecasting to 5.8%, a significant improvement over raw time-series features. The MFD-VE achieves a data valuation accuracy of 93.2%, substantially outperforming baselines that rely solely on access frequency. These improvements indicate that integrating business-semantic information and pedagogical relevance allows the system to more accurately predict workload fluctuations and assess the dynamic value of educational data objects.

When compared to a rule-based baseline, the DQN-SM migration strategy reduces average access latency by 71.1%, from 128 ms to 37 ms. The system also reduces storage costs by 24.0% and boosts the Service Level Agreement (SLA) compliance rate to 99.1%. The migration volume remains low at 2.4 GB/h, indicating that the high-quality features provided by the preceding modules enable the DQN to make precise migration decisions with fewer operations. The prefetch hit rate reaches 85.3%, further validating the effectiveness of teaching-cycle awareness in proactive migration.

Ablation studies reveal that the teaching-cycle characteristics provided by MSTP-LM are the most critical component for the proactiveness of migration decisions. Removing this module causes the largest performance degradation, with latency rising by 40.5%. The removal of the MFD-VE module also significantly impacts performance, confirming the importance of dynamic value evaluation. The DQN training strategy, specifically the use of curriculum learning, is crucial for convergence quality, as removing it increases latency by 21.6%. These findings highlight the synergistic necessity of feature engineering and the intelligent decision framework.

The study acknowledges several limitations, including the fact that experiments were conducted on a single-university dataset, leaving generalizability to diverse institutions unvalidated. Evaluation was performed on a trace-driven simulator rather than a live production system, which may not fully capture factors such as network jitter and hardware failures. Additionally, the cost model does not include long-term operational costs such as hardware depreciation, and the DQN-based decision framework lacks interpretability, which could hinder operational trust in production deployments.

Future research will focus on expanding the method from single-university scenarios to diverse university environments, aiming to achieve breakthroughs in cross-scenario knowledge transfer mechanisms. The team plans to develop methods for aligning representations of heterogeneous educational data and scenario-adaptive parameter fine-tuning strategies. This will enable storage systems to rapidly adapt to the unique requirements of institutions with varying educational missions and instructional paces, ultimately driving the realization of global optimization in performance, cost, and energy consumption within a diversified educational ecosystem.

The dataset and code associated with this study are publicly available, allowing for reproducibility and further exploration by the research community. The work contributes to the broader field of intelligent storage management by demonstrating how domain-specific knowledge can be leveraged to enhance the efficiency and performance of data-intensive educational systems. As digital transformation in education continues to deepen, such adaptive storage solutions will be increasingly vital for managing the growing volume and complexity of educational data.

To evaluate the framework under controlled yet realistic conditions, the researchers deployed a replay-driven simulator that mirrored a physical cluster of four servers equipped with Intel Xeon Gold 6248R processors, Nvidia RTX 5090 graphics processing units, and tiered storage consisting of non-volatile memory express solid-state drives, serial ATA drives, and traditional hard disks. Benchmarking confirmed the simulator closely matched the university’s live environment, maintaining request-count and access-latency deviations below 3.5 percent. The research team enforced a strict chronological data split across the 28-month evaluation window, allocating 16 months for model training, six months for validation using Optuna hyperparameter optimization across 200 trials, and six final months across distinct academic periods to test cross-semester robustness without temporal leakage.

Beyond average speedups, the full Deep Q-Network system showed marked improvements in controlling tail latency across hundreds of thousands of student interactions. The architecture recorded a median access latency of 27 milliseconds, a 99th-percentile response time of 98 milliseconds, and a maximum recorded latency of 230 milliseconds. When matched against alternative algorithmic approaches under identical feature inputs, the discrete action space of the Deep Q-Network outperformed proximal policy optimization, which averaged 42 milliseconds and suffered from discretization overhead. It likewise surpassed a competitive random forest model, which achieved 44 milliseconds when supplied with cyclical features but lacked the long-term cumulative optimization benefits delivered by reinforcement learning.

Stepwise component analyses further clarified how each technical refinement drove overall performance gains. In workload forecasting, relying purely on raw time series yielded an initial error rate of 12.3 percent. Progressively embedding academic calendar milestones trimmed this prediction error to 8.1 percent, integrating granular course schedules lowered it to 7.0 percent, and adding cross-scale attention fusion brought it down to 6.2 percent before reaching the full model benchmark. Similar incremental advantages appeared in data classification, where tracking access frequency alone attained an accuracy of only 74.5 percent. Incorporating instructional relevance raised classification accuracy to 87.6 percent, adding relevance propagation along data dependency links elevated it to 90.8 percent, and applying stage-aware adaptive weighting achieved the final 93.2 percent accuracy.

Investigation of training dynamics demonstrated that the structured representations accelerated model convergence while dampening instability. The complete reinforcement learning framework converged within approximately 40,000 training episodes, effectively doubling the training speed of a baseline model lacking domain features, which required 80,000 episodes. Hyperparameter sensitivity testing revealed that the system performed best when setting its recency decay constant to seven days, matching the natural weekly rhythm of university coursework. Migration volume and latency remained stable within a 20 percent parameter perturbation range, confirming that the autonomous storage policy is robust against operational shifts.

Subject of Research: Intelligent storage optimization for educational lakehouse architectures using periodic workload modeling and deep reinforcement learning.

Article Title: Storage optimization method for lakehouse architecture in education systems with periodic workloads

Article References: Wei, G., Xingyuan, G., & Min, H. (2026). Storage optimization method for lakehouse architecture in education systems with periodic workloads. Discover Computing, 29(1), Article 787. https://doi.org/10.1007/s10791-026-10739-0

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10739-0

Keywords: lakehouse architecture, educational data management, deep reinforcement learning, storage optimization, periodic workloads, data migration, academic calendar semantics, tiered storage, Deep Q-Networks, load forecasting, data value evaluation, SLA compliance, access latency reduction, curriculum learning, reward shaping, educational big data, intelligent storage, workload prediction, cost optimization, proactive data migration