Abstract
In the process of responding to user applications on highway structural health monitoring (SHM) data sharing platforms, the primary objective is to decrease network transmission costs and avoid redundant Input/Output (I/O) operations between storage servers and local disks. This paper investigates the problem of cost-saving response scheduling for SHM data applications. We develop a greedy response scheduler with a $(1 + \frac{1}{e-1})$ approximation ratio. Evaluation results demonstrate that our proposed solution significantly improves efficiency and reduces transmission costs compared to traditional scheduling methods.
The Network I/O Bottleneck
Highway Structural Health Monitoring (SHM) systems utilize high-frequency sensors (up to 2000Hz) to ensure infrastructure safety, generating massive datasets—often 5 to 6 TB per month. These platforms typically employ a two-tier architecture: a front-end server for user interaction and a secure, isolated back-end storage server.
As shown in Figure 3, network I/O transmission is the primary bottleneck, accounting for approximately 87.8% of total processing time. Standard scheduling approaches, such as First-Come-First-Served (FCFS), fail to account for data locality, leading to redundant data fetches from the back-end and severe performance degradation. Prior research has focused on data analysis for anomaly detection, but the challenge of efficient data sharing remained unaddressed until now.
Methodology: Greedy Overlap Scheduling
The authors frame the challenge as a "Cost-saving response scheduling" problem and prove it is NP-hard. However, they identify that the optimization function—minimizing I/O costs—is an increasing submodular function. This mathematical property allows for the implementation of a highly efficient Greedy Algorithm with a guaranteed performance bound.
The core heuristic of the scheduler is data overlap. As illustrated in Figure 6, the scheduler analyzes the queue of pending applications and calculates the overlap ($O_i$) between the requested data ($D_{ij}$) and the data already cached on the front-end server's local disk ($D_L$). By prioritizing requests with the maximum overlap, the system minimizes the amount of new data that must be fetched over the slow network link.
Experimental Evaluation
The scheduler was tested using both synthetic datasets and real-world sensor data from the Jilin Tonghua Highway in China. The real-world data included transverse, longitudinal, and vertical strain/stress measurements with file sizes up to 988 MB.
Key Results:
- I/O Reduction: The greedy scheduler reduced network I/O transmission sizes by an average of 42.16% compared to eight baseline strategies, including FIFO, size-based scheduling, and Q-Learning.
- Computational Efficiency: The algorithm operates with $O(n^2 \log n)$ complexity. In practical tests, it processed batches of 50 applications in milliseconds, making it ideal for real-time administrative use.
- Cache Strategy: Interestingly, the researchers found that a Random cache replacement policy performed best when paired with this scheduler, as SHM query patterns are often unpredictable and lack the temporal locality that favors policies like LRU.
Conclusion and Future Work
By prioritizing data locality through submodular optimization, the proposed greedy scheduler effectively mitigates the network I/O bottleneck in SHM platforms. While highly effective for batched requests, future work will focus on improving the user experience through "data sketches" and visual summaries. These tools will allow users to preview data before requesting large downloads, further reducing unnecessary network load at the source.
Summary updated