The scalable linear algebra package (ScaLAPACK) is a critical library for parallel computing on distributed-memory systems, enabling the development of numerous scientific applications that depend on robust linear algebra operations. However, for specific computations such as the Cholesky decomposition, the native ScaLAPACK routines are not communication-optimal and fail to fully leverage the capabilities of modern parallel architectures. This paper proposes the Parallel Cholesky Factorization (PCF) optimization algorithm designed to address these limitations within the ScaLAPACK framework. The PCF algorithm enhances performance and load balancing by strategically differentiating data partitions across processes. It involves a temporary redistribution of computational workloads to a root process, which performs a concentrated calculation before redistributing the results. This approach ensures a more balanced utilization of CPU resources. Experimental evaluation is conducted on both Intel and Kunpeng processor platforms. The first set of experiments demonstrates that the PCF algorithm achieves an average performance improvement of 30
Fully homomorphic encryption (FHE) enables computation over encrypted data without decryption, ensuring data confidentiality throughout the processing pipeline. However, the complexity and heterogeneity of FHE schemes like CKKS, BFV, and TFHE pose challenges for unified hardware design. In this paper, we propose the Homomorphic Processing Unit (HPU), a general-purpose platform supporting multiple FHE schemes. Unlike fixed-function accelerators, HPU is implemented as a RISC-V extension and introduces a dedicated homomorphic instruction set architecture (H-ISA), comprising micro-instructions for minimal compute kernels and macro-instructions for integration with general-purpose models. Core operations like the number-theoretic transform (NTT) and automorphism are abstracted into a unified instruction layer, mapped onto specialized homomorphic compute units. HPU integrates a collaborative computation and storage design, utilizing a polynomial-level instruction set and a polynomial-granular memory management unit (MMU) to optimize memory and reduce data movement. By combining a complete ISA with a general-purpose computing core, HPU can achieve efficient computing, control, and scheduling. We evaluate the HPU prototype using four multi-FHE benchmarks on field-programmable gate array (FPGA) and 7 nm ASIC. The key results are as follows. 1) HPU achieves up to 1596x and 1174x speedup over CPU for CKKS-like and TFHE operations, respectively. 2) Compared with SOTA FPGA solutions for CKKS and TFHE, HPU achieves 1.36x and 1.2x improvement, respectively. 3) HPU’s ASIC implementation achieves 1.36x speedup over state-of-the-art TFHE accelerators for long short-term memory (LSTM).
With the rapid rise of domestic accelerators and the growing demand for efficient cryptographic algorithm implementations, optimizing the performance of domestic commercial cryptography on homegrown hardware platforms has become a critical research challenge. This paper focuses on utilizing the Chinese Hygon Deeplearning Computing Unit (DCU) as an acceleration platform and proposes a high-performance SM4 encryption implementation. First, we design a collaborative CPU/DCU computing framework, where the CPU handles task scheduling while the DCU performs parallel execution of computationally intensive tasks. Additionally, we leverage the data copy engine of the DCU and employ HIP (heterogeneous-computing interface for portability) stream-based data transmission to enhance computational efficiency. Second, we propose an optimized hierarchical memory access strategy by merging multiple sets of computational data for unified transfer and employing a hierarchical storage scheme based on data characteristics to improve memory access efficiency. Furthermore, we employ GCN (Graphics Core Next) ISA (Instruction Set Architecture) inline assembly to achieve constant-time computation, which enhances both security and computational performance. Finally, we support both single-key and multi-key encryption modes, optimizing memory access and parallel computation for each. The experimental results show that our optimized SM4 implementation HI-SM4 achieves a throughput of 161.81 Gbps in the multikey encryption mode and up to 730.88 Gbps in the single-key encryption mode. Compared with the best existing SM4 implementations on CPU, FPGA, and GPU platforms, our approach achieves performance improvements of 28.22x, 4.50x, and 1.36x, respectively. Our research presents a feasible framework for efficiently implementing indigenous cryptographic algorithms on domestically developed hardware accelerators, thereby contributing to the advancement of high-performance cryptographic computation.
Metaproteomics, utilizing high-throughput liquid chromatography-mass spectrometry (LC-MS), offers a profound understanding of microbial communities. Quantitative metaproteomics further enriches this understanding by measuring relative protein abundance and revealing dynamic changes under different conditions. However, the challenge of missing peptide quantification persists in metaproteomics analysis, particularly in data-dependent acquisition (DDA) mode, where high-intensity precursors for MS2 scans are selected. To tackle this issue, the match-between-runs (MBR) technique is used to transfer peptides between LC-MS runs. Inspired by the benefits of MBR and the need for streamlined metaproteomics data analysis, we develop SEMQuant, an end-to-end software integrating Sipros-Ensemble’s robust peptide identifications with IonQuant’s MBR function. The experiments show that SEMQuant consistently obtains the highest or second highest number of quantified proteins, with notable precision and accuracy. This demonstrates SEMQuant’s effectiveness in conducting comprehensive and accurate quantitative metaproteomics analyses across diverse datasets and highlights its potential to propel advancements in microbial community studies. SEMQuant is freely available under the GNU GPL license at https://github.com/Biocomputing-Research-Group/SEMQuant .
We introduce the Air-Writing General Text Benchmark, termed AWGT-Benchmark (AWGT), a video-based, multilingual, and multigranularity dataset designed to advance air-writing text recognition (AWTR) across diverse scenarios. Although AWGT bridges the gap between conventional text recognition and AWTR scenarios, it also poses new challenges for computer vision and natural language processing. AWGT comprises 226 648 video frames across four subsets in Chinese and English, with finger-writing trajectories captured using RGB cameras in real-world scenarios to ensure realism and diversity. Based on AWGT, we propose a two-stage recognition framework that first extracts finger motion trajectories and subsequently converts them into trajectory images for character recognition. This design effectively suppresses background noise and emphasizes structural character details. Experimental results on AWGT show that widely used text recognition models suffer substantial performance degradation when applied to air-writing tasks, highlighting their limited ability to handle dynamic finger movements and cluttered visual contexts. AWGT provides a unified evaluation protocol and an experimental baseline, reveals key limitations of existing methods, and offers insights for developing more robust and adaptable cross-scenario recognition systems. Furthermore, it contributes to advancing research in intelligent human-computer interaction. The dataset is made publicly available to facilitate future research: https://www.scidb.cn/detail?dataSetId=0c676ce0d6fb41149c72613ab5ba968b version=V1 code=j00247 .
To develop a medical generative system, current research has primarily focused on integrating medical knowledge into language models (LMs) through training them on medical corpora or incorporating external medical resources, such as knowledge graphs and terminologies. However, approaches rely on implicit knowledge integration from textual narratives or standardized knowledge sources (e.g., Unified Medical Language System), making the acquired knowledge opaque and difficult to control during inference. This lack of transparency limits the traceability of knowledge accumulated from historical training data. To address these limitations, we propose constructing a personalized medical knowledge base that extracts discrete knowledge points from real doctor-patient dialogues and encodes them as semantically condensed vectors. Building on this, we introduce a novel dense retrieval-augmented medical dialogue generation framework, called RagMedge, which retrieves relevant knowledge embeddings conditioned on the input context and incorporates them into response generation. Experimental results demonstrate that RagMedge achieves state-of-the-art performance, highlighting the effectiveness of leveraging personalized medical knowledge for improved dialogue generation.
Recent advancements in financial large language models (FinLLMs) have shown strong performance in high-resource languages like English and Chinese, but are limited in low-resource settings, particularly in Southeast Asia (SEA), where labeled data resources are extremely scarce and challenging to annotate. Existing benchmarks primarily focus on high-resource finance, neglecting low-resource finance. To address these issues, we introduce CroFinBen, to the best of our knowledge, the first multilingual benchmark specifically designed to bridge the language-resource gap between high- and low-resource finance. It includes four key financial NLP tasks—financial sentiment analysis (FinSA), financial stock prediction (FinSP), financial text summarization (FinTS), and financial text classification (FinTC)—across both high-resource languages (English and Chinese) and low-resource Southeast Asian (SEA) languages (Indonesian, Malay, Thai, Filipino, and Vietnamese), comprising over 50 000 samples from 16 datasets, providing a comprehensive and balanced evaluation of LLMs. Unlike others that rely on full translations or overlook local context, CroFinBen incorporates localized annotations to reflect financial terms and cultural nuances in SEA languages. Evaluating 25 LLMs shows significant performance differences, with no clear proficiency in either high- or low-resource languages, especially for existing language-biased fine-tuned FinLLMs. The 1800B large-parameter closed-source GPT-4o excels, while DeepSeek-V3 and ChatGPT-3.5 also perform well. By bridging language-resource barriers, CroFinBen enhances the fairness and robustness of FinLLMs, providing strong support for improving performance in global financial scenarios. Our data resources are available at https://jcst.ict.ac.cn/en/supplement/afc97c89-905b-4397-941c-60dd0c720248 .
Although large language models (LLMs) have been widely deployed across numerous applications, they can generate harmful or illicit content, posing substantial safety risks. Evaluating such risks requires effective evaluation methodologies using high-quality benchmarking datasets. This study introduces LLMSafety-Choice in this regard, a multilingual benchmark for content safety evaluation containing 11 911 multiple-choice questions in both Chinese and English, covering four safety domains and eight categories per language (nine in total). We further introduce a systematic metamorphic-testing approach, defining seven metamorphic relations for LLMSafetyChoice, for LLM safety evaluations. Through an extensive empirical study involving 1 408 evaluation scenarios (11 LLMs×(8 categories×2 languages)×(1 constructed benchmark + 7 transformations)), we reveal key insights into model behavior under safety-critical conditions and demonstrate that metamorphic testing effectively uncovers subtle safety vulnerabilities. The benchmark and evaluation results are publicly available at https://anonymous.4open.science/r/LLMMetamorphic-08C9/ .
While large language models for code (Code LLMs) excel at generating functionally correct code, existing benchmarks neglect a crucial aspect: adherence to explicit time complexity constraints. We introduce the Complexity-Constraint Code Evaluation (C3E), a novel benchmark evaluating both functional correctness and complexity compliance across feasible and infeasible scenarios. C3E enables precise differentiation between asymptotic complexity classes and tests model robustness against theoretically impossible constraints. Our proposed Complexity Alignment Score (CAS) integrates correctness and complexity adherence into a unified metric, assessed through theoretical analysis rather than costly executions. Experiments reveal a striking gap in state-of-the-art models: GPT-4o achieves 81 https://github.com/wahaha12321/C3E .
Magnetic resonance imaging (MRI) can generate images with varying contrasts and acquisition times depending on imaging parameters. Utilizing a high-resolution contrast with a short acquisition time as a reference for the super-resolution (SR) of low-resolution contrasts with long acquisition times is effective for the rapid acquisition of high-quality images. However, existing methods mainly process features in the spatial domain, and overlook potential features in the frequency domain. This paper proposes Multiscale Spatial-Frequency Collaborative Fusion Network (MSCFN), which jointly leverages information in the spatial and frequency domains for SR. A global-local fusion block optimizes global structural features and local texture details at different scales, and an adaptive low-high frequency fusion module utilizes the complementary nature of multiple contrasts to decompose reference images into high- and low-frequency components and adaptively fuse them to enhance feature integration. Experimental results indicate that MSCFN outperforms existing multicontrast MRI SR methods. The code is publicly available at https://github.com/crystal177/MSCFN .
Predicting query performance is essential for database tasks such as resource allocation and scheduling. However, existing methods, designed for single-node systems, often fail in distributed analytical databases. This is because they overlook key features such as data partitioning, cross-node data transfer, and parallel query execution. To address these challenges, we propose a novel approach, PPredictor (Performance-Predictor), for predicting query performance in analytical distributed databases that is grounded in graph representation models. First, we introduce a graph model that encodes both data partitions and query execution plans within distributed databases. In this model, vertices represent partitioned tables, whereas edges capture their relationships, such as partitioned tables located on the same data node and data transfers between partitioned tables within an execution plan. Second, we present a prediction model that effectively uses a graph attention mechanism network to encode graph features and uses deep learning techniques for performance prediction. Third, considering dynamic workloads and various database environments, we introduce an incremental learning method based on feature replay. Extensive experiments conducted on real-world datasets demonstrate that our approach significantly outperforms state-of-the-art methods.
Discovering possible candidate elements and capturing their connections to form the output triplets constitute the core challenge of aspect sentiment triplet extraction (ASTE). However, the information encapsulated by prevailing methods within a solitary sentence may often prove insufficient, particularly in complex scenarios characterized by uncommon aspect, opinion terms or intricate syntax patterns. To mitigate these limitations, we advocate incorporating inter-sentence information retrieval to enrich intra-sentence representations within ASTE. One existing study has proposed a method dubbed Retrieval-Based Aspect Sentiment Triplet Extraction via Label Interpolation (RLI), which retrieves triplets from the corpus to augment the representations of a candidate aspect-opinion pair and further improve sentiment prediction. Nevertheless, obtaining data with standard triplets might be challenging in practice. Therefore, we propose an approach, namely Multi-Task ASTE with the Corpus-Enhanced Graph (MACG), to conduct sentence-level retrieval and extract helpful information from unlabeled similar sentences. Specifically, we design a corpus-level enhanced graph to capture inter-sentence information, alongside a local graph preserving intra-sentence information. A graph neural network is subsequently employed to adaptively learn enhanced representations of the target sentence for ASTE. RLI and MACG collaboratively form a comprehensive methodological framework, which is effective in both scenarios with and without standard triplet labels. Extensive experiments on two benchmarks demonstrate the superiority and flexibility of retrieving inter-sentence information, which underscores their potential to advance ASTE by leveraging neighboring information.
Graph neural networks have achieved remarkable performance across a wide range of complex graph-based tasks. While subgraph sampling (SS) methods substantially improve per-epoch efficiency for large graphs, they introduce higher gradient variance, which hinders convergence and reduces accuracy. Moreover, applying corrections to reduce SS variance impact faces the issue of diminishing returns, thereby limiting training efficiency. To address these challenges, we propose ECHO+, a novel dual-level correction framework designed to accelerate training while maintaining accuracy comparable to that of node sampling. ECHO+ employs a lightweight, variance-guided strategy search during preprocessing to reduce SS variance. During training, it operates on two levels: a coarse level that adaptively schedules correction to reduce residual SS variance, and a fine level that leverages a batch-loss driven early stopping mechanism to improve overall training efficiency. Experimental results reveal that ECHO+ achieves rapid convergence with high accuracy, delivering a speedup of up to 12.4x over the node sampling baseline while maintaining comparable performance. Furthermore, ECHO+ outperforms existing SS baselines, achieving up to 4.3x faster convergence.
Large-scale machine learning (ML) clusters deploy hundreds of models packaged as large container images, making scheduling and image pulling dominant contributors to deployment latency. Current systems adopt a hybrid-granularity strategy: they schedule using image-level metadata, reuse on-node layer caches, and lazily fetch content at the file or chunk level. This fragmentation leads to suboptimal cross-image reuse. We show that an intermediate unit, software granularity, can be uniformly applied at scheduling, caching, and pulling to improve deployment performance. We present Containerlet, a software-scoped layer that reorganizes standard images into one private layer plus multiple reusable per-software units that reside in compatible registries and are pulled by unmodified runtimes. Containerlet promotes cross-image reuse while bounding the number of units that the scheduler must consider. We design a file backtracking method to convert standard images into Containerlet images, which includes one private layer and multiple Containerlets. We also extend Kubernetes and container runtime to support software-granular scheduling, caching, and lazy pulling. End-to-end deployment experiments on diverse ML images and production cluster traces show consistent gains, reducing deployment overhead by 18
Beasiswa PIP di SMP Negeri 2 TILAMUTA merupakan pemberian berupa bantuan keuangan yang diberikan kepada perorangan yang bertujuan meringankan beban biaya siswa demi keberlangsungan pendidikan yang ditempuh. Kendala yang di hadapi oleh pihak sekolah yaitu terdapat banyak kriteria yang harus di perhitungkan sehingga kesulitan dalam menentukan prioritas calon penerima PIP dan perhitungan yang di lakukan oleh bagian staff guru pengelola PIP masih secara manual dengan data-data yang ada sehingga perhitungan pun di lakukan berulang-ulang untuk mendapatkan hasil yang akurat dan perhitungan pun memakan waktu yang cukup lama. Ketidakefisienan dalam menghitung data tersebut mengakibatkan sasaran penerima beasiswa PIP kurang tepat.
Boundary value problems (BVPs) have wide applications in computer graphics and mechanical engineering. This paper presents a quasi-bounding method for progressively solving several boundary value problems, where initial solutions are easy to be obtained for rapidly solving the accurate solution. Given an equation F(u, t) = 0,t ∈ [a, b], and several boundary value constraints as well, two systems consisting of n + 1 equations are derived for rapidly searching two polynomials fn, 1(t) and fn, 2(t) of degree n, which satisfy F(fn, 1,t) ≤ 0 ≤ F(fn, 2, t),t ∈ [a, b] in the cases when certain conditions are satisfied. From the middle value theorem, the solution u⋆(t) is bounded by fn, 1(t) and fn, 2(t), t ∈ [a, b]. These two bounding polynomials fn, i(t), i = 1, 2, are taken as initial values for progressive refinements of approximation error in two ways, i.e., B-Spline form of the same degree with more knots, and Bézier form of a higher degree. Numerical experiments show that the new method can be applied to more generalized BVPs, and achieves better computational stability, much better approximation with less error and better computational efficiency than prevailing methods, even by using a small degree n.
Computility networks (CNs) enable large-scale computation scheduling and have emerged as a new computing paradigm. CNs have a broader service scope and more complex infrastructure than cloud and edge computing. Consequently, resource allocation and task scheduling in CNs face numerous challenges, such as unifying diverse computility resources in existing heterogeneous clouds, adequately incorporating network resource providers within the CN framework, and pricing computility resources. In this study, first, we extract computility and network resources to construct a task scheduling model for CNs. To maximize the number of tasks successfully scheduled, we represent this problem as a mixed-integer programming model that involves multiple roles, tasks, and resource constraints. Unlike most approaches, we explicitly incorporate network resource providers into the model. Second, we propose a double auction mechanism named Computility Double Auction (Computility_DA) to address the task scheduling problem in CNs. Specifically, we derive feasible solutions for task scheduling and network flow using optimization methods and then determine the final winners and payment pricing solution based on matching theory. Furthermore, we demonstrate that the proposed mechanism has economic properties such as individual rationality, truthfulness, and budget balance. Experimental results demonstrate that compared with existing algorithms, Computility_DA significantly increases the number of scheduled tasks and the utility and revenue for participants.
Penelitian ini dilatarbelakangi oleh permasalahan dalam pelayanan administrasi kependudukan di Dinas Kependudukan dan Pencatatan Sipil (Dukcapil) Kabupaten Boalemo, di mana tingginya volume permohonan dokumen seperti akta kelahiran sering kali menyebabkan antrean panjang dan inefisiensi. Prosedur penentuan prioritas yang berjalan saat ini masih bersifat konvensional dan subjektif, sehingga rentan terhadap bias dan ketidakpastian waktu tunggu bagi masyarakat. Tujuan dari penelitian ini adalah untuk merancang dan mengimplementasikan Sistem Pendukung Keputusan (SPK) menggunakan metode Simple Additive Weighting (SAW) guna menentukan prioritas pelayanan dokumen administrasi kependudukan secara objektif. Sistem ini dikembangkan dengan menggunakan bahasa pemrograman PHP dan basis data MySQL, serta diuji melalui teknik White Box dan Black Box untuk memastikan fungsionalitas dan logika sistem berjalan dengan benar. Hasil penelitian menunjukkan bahwa penerapan metode SAW mampu memberikan rekomendasi prioritas permohonan berdasarkan kriteria yang telah ditentukan, seperti surat keterangan lahir, dokumen kependudukan orang tua, dan urgensi lainnya. Dengan adanya sistem ini, diharapkan proses pelayanan di Dukcapil Kabupaten Boalemo menjadi lebih transparan, akuntabel, dan efisien, sehingga dapat meningkatkan kepuasan masyarakat terhadap layanan publik.
pan manusia, khusus-nya di teknologi informasi yang sangatberpengaruh di zaman saat ini, dan segala sesuatu sudah menjadi hal yanglumrah serba mudah untuk melakukan aktivitas dalam dunia teknologiinformasi. Quick Response Code atau yang lebih dikenal dengan sebutan qrcode merupakan kode dua dimensi sebagai pengembangan dari kode batangatau barcode. sebuah aplikasi absensi dan presensi siswa pada SD Negeri 05Tilamuta menggunakan qr code. Pada sekolah SD Negeri 05 Tilamutamemang belum pernah dilakukan penelitian terkait yang dilakukan olehorang lain tersebut, pada Sekolah Dasar Negeri 05 Tilamuta, khususnya padakelas V dan VI sd sampai saat ini masih kurang dalam memanfaatkanteknologi absensi berbasis android, guru dalam melakukan absensi hanyamenggunakan absen manual kepada siswa di SD Negeri 05 Tilamuta.
