Serverless computing, with its ease of management, auto-scaling, and cost-effectiveness, is widely adopted by deep learning (DL) applications. DL workloads, especially with large language models, require substantial GPU resources to ensure QoS. However, it is prone to produce GPU fragments (e.g., 15%-94%) in serverless DL systems due to the dynamicity of workloads and coarse-grained static GPU allocation mechanisms, gradually eroding the profits offered by serverless elasticity. Different from classical serverless systems that only scale horizontally, we present introspective elasticity (IE), a fine-grained and adaptive two-dimensional co-scaling mechanism to support GPU resourcing-on-demand for serverless DL tasks. Based on this insight, we build Dilu, a cross-layer and GPU-based serverless DL system with IE support. First, Dilu provides multi-factor profiling for DL tasks with efficient pruning search methods. Second, Dilu adheres to the resourcing-complementary principles in scheduling to improve GPU utilization with QoS guarantees. Third, Dilu adopts an adaptive 2D co-scaling method to enhance the elasticity of GPU provisioning in real time. Evaluations show that it can dynamically adjust the resourcing of various DL functions with low GPU fragmentation (10%-46% GPU defragmentation), high throughput (up to 1.8× inference and 1.1× training throughput increment) and QoS guarantees (11%-71% violation rate reduction), compared to the SOTA baselines.

Graph matching is a widely studied and applied field, with Graph Edit Distance (GED) serving as a fundamental metric for evaluating the similarity between graphs. As an NP-hard problem in graph theory, GED focuses on determining the minimum cost required to transform one graph into another through a series of edit operations. Existing research on GED primarily falls into two categories: exact algorithms and inexact algorithms. This paper focuses on enhancing the accuracy of inexact GED computation and introduces the Greedy 2-Swap (G2S) algorithm. The G2S algorithm adopts a greedy strategy to explore graph mappings, consistently selecting mappings that minimize the edit distance. Our approach establishes an upper bound on the GED, aiming to achieve a lower upper bound for more precise results. To address the inherent limitation of greedy algorithms—their tendency to converge to local optima—we propose initialization and random rearrangement strategies to escape such states and enhance solution quality. In addition, we provide a comprehensive comparative analysis of our algorithm against state-of-the-art methods, including A-star and GNN-based approaches. We also investigate the impact of G2S on edit distance calculations. The experimental results highlight the superior performance of our algorithm across various metrics. Notably, the G2S algorithm combines simplicity with robust generalization across diverse graph datasets. Furthermore, its scalability is significantly improved through the integration of multiple optimization strategies.

A low-code development platform (LCDP) is an efficient way to reduce the learning curve for programming languages. It provides a graphical user interface and configurable environment for application building. Users can focus on the design and functionality parts rather than the technical details of programming languages. This can lead to a faster and more efficient development process. However, most of the low-code platforms are mainly used on web pages through traditional interactive devices like mouse and keyboard. To the best of our knowledge, we found there are currently few tools for creating visual interfaces using free-hand manipulation in AR. Interactions in current AR are mostly limited to the two-dimensional display space of screens. Thus, we propose NCARVis, a LCDP that is a novel visualization interface-creating system, it can help users create and show visualization without any programming skills. To provide an environment for creating visualization in NCARVis, we have proposed an initial prototype (MARLP) that changes the interaction method of only touching the 2D screen in the traditional AR scene and provides more interaction methods in AR. Users can freely explore virtual worlds through it, interacting in different directions, and have more creative space to place all diagrams than a 2D scene. We present two tasks for five users without any programming skills to evaluate the usability and effectiveness of NCARVis in visualization creation and design. And we compare NCARVis with a graphical low-code development platform to evaluate the advantages of NCARVis on learning cost. We plan to explore more interesting interactive experiences and visualization design in the future.