Wenhao Zhu

Efficient multi-layer Cartesian grid generation and computation are vital for computational fluid dynamics simulations in complex geometries, where traditional methods face challenges with lengthy grid generation times and numerical dissipation in multi-layer lattice Boltzmann method (LBM). This study proposes an advanced multi-layer LBM computational framework that enhances both accuracy and efficiency by integrating parallel multi-layer grid generation algorithm with a buffer-driven bicubic interpolation algorithm to solve these problems. The proposed parallel Cartesian grid generation algorithm introduces a load-balancing strategy for adaptive grid generation, which automatically adjusts the distribution of cells that intersect with geometry boundaries, balancing the number of intersected cells in different domains across up to three dimensions. To address the challenges of data transfer between different grid layers in multi-layer LBM, a new buffer-driven bicubic interpolation method is introduced. By creating a buffer zone between coarse and fine grids, the interpolation method extends the range of data sources, allowing more neighboring grid points to be used in computations. Numerical experiments demonstrate that these algorithms lead to faster multi-layer grid generation and excellent computational performance in both 2D and 3D cases.

The regularized lattice Boltzmann method (RLBM) is an improvement of the lattice Boltzmann method (LBM). The advantage of RLBM is improved accuracy without increasing computational overheads. The paper introduces the method of multi-layer grids, the multi-layer grids have different resolutions which can accurately solve problems in computational fluid dynamics (CFD) without destroying the parallelism of RLBM. Simulating fluid flow usually requires a large number of grid simulations. Therefore, it is necessary to design a parallel algorithm for RLBM based on multi-layer grids. In this paper, a load-balancing-based grid dividing algorithm and an MPI-based parallel algorithm for RLBM on multi-layer grids are proposed. The load balancing-based grid dividing algorithm ensures that the workload is evenly distributed across processes, minimizing the discrepancies in computational load. The MPI-based parallel algorithm for RLBM on multi-layer grids ensures accurate and efficient numerical simulation. Numerical simulations have verified that the proposed algorithms exhibit excellent performance in both 2D and 3D experiments, maintaining high stability and accuracy. The multi-layer grids method is significantly better than single-layer grids in terms of CPU runtime and number of grids required. Comparative analysis with the OpenMP multi-threading method on the multi-layer grid RLBM shows that the proposed algorithm in this paper achieves superior speedup and efficiency.

Cloud computing technology is widely embraced due to its ability to meet diverse computing resource demands from users. As the user base grows, the challenge for cloud service providers to dynamically allocate resources to applications according to real-time customer needs becomes more daunting. Existing auto-scaling mechanisms, such as Horizontal Pod Autoscaler (HPA), lack flexibility in handling transient and irregular fluctuations in resource demands, making it difficult to respond quickly. This research focuses on auto-scaling solutions for applications in high-concurrency scenarios. The study enhances the MAPE structure of traditional auto-scaling models and introduces a trend detection module in a proactive auto-scaling engine based on predictive algorithms. This module detects trends during transient irregular request volume fluctuations, corrects prediction algorithm results inconsistent with the current trend, and provides more reasonable resource allocation for programs. Experimental results, using both real access peak data and simulated data in a Kubernetes environment, indicate that the proposed auto-scaler effectively prevents resource shortages in high-concurrency network environments compared to other auto-scaling mechanisms. It ensures the performance and availability of applications while reducing resource wastage.

基于笛卡尔网格的格子Boltzmann方法具有降低计算复杂度，易于自适应和处理复杂边界条件等优势，然而非贴体的特性导致在计算复杂几何或复杂流动时通常需要过大的网格量。为了缓解这个问题，对复杂几何进行局部的网格加密是一个有效的手段。其中，实现局部网格加密的关键在于快速网格生成和不同尺度网格之间的映射关系。本文通过设计新的数据结构和基于射线交点算法，提出了一种新的快速网格生成算法，并将其应用在两种基本单元下的多层网格格子Boltzmann方法中，从而进一步发展高效的网格生成技术。考虑射线法的算法特性，经预处理后的射线，通过新的数据结构可大幅度减少需要检索的物面单元数量。对于两种基本单元下的几何自适应多层网格，格心格式通过叉树结构构建不同层之间的网格映射关系，格点格式通过背景网格重叠确定映射关系。在两种格式的网格结果中，格心格式的多层网格受加密次数的影响相对较小，在加密9次的千万网格点量级下达到每秒约589万网格点的生成效率。格点格式受加密次数的影响较大，但在千万量级下仍达到每秒约66万网格点的生成效率，比目前已有方法提高了约40倍的效率。最后通过二维圆柱绕流问题、三维圆球绕流和三维潜艇的绕流求解验证了网格算法的准确性。

In this study, we propose a user behavior-based query optimization strategy to address the commonly observed suboptimal query performance and user experience in the field of video retrieval. This strategy primarily involves analyzing user behavioral data during the video retrieval process to expand query keywords, enhancing their relevance and diversity, thus optimizing search results. We achieved this by augmenting the original dataset to generate simulated user behavioral data and Ground Truth data. Subsequently, an automated query expansion algorithm effectively applied this data to video retrieval. Through experimental assessment on the video retrieval dataset (MSR- VTT), we validate the effectiveness of this method. The results demonstrate that this user behavior- driven query expansion approach significantly improves the accuracy of video retrieval and user satisfaction.

Caching plays a crucial role in network systems, and an appropriate caching strategy can effectively reduce network load and enhance user experience during video browsing. The proliferation and development of smartphones have led to a surge in global mobile data traffic. This paper targets the viewing characteristics of short video platform users, and proposes two adaptive video cache update mechanisms, TelLRU and TelLRU(sigmoid), aiming to achieve optimal performance by dynamically adjusting the cache video block size according to the video playback duration. This strategy is an improvement on the LRU cache replacement algorithm, adding a section of the video file of a new request to the cache. During video playback, the length of the video block changes dynamically according to a specific algorithm, and the LRU replacement strategy replaces the least recently used video block. Experimental results show that this strategy can significantly improve the cache hit rate of mobile video-on-demand platforms, reducing delay time and delay proportion.