High-performance computing (HPC) systems generate massive scientific datasets, often stored in remote data repositories. Limited bandwidth and resource-constrained end-points pose challenges for efficient large data transfer. Error-bounded lossy compression addresses this by reducing data sizes (higher bit-rates) while controlling distortion. However, different compressors exhibit distinct rate-distortion behaviors even under the same error bounds. Thus, selecting an optimal compressor before transfer is essential to meet ndpoint-specific requirements e.g., maximizing data reduction at a fixed distortion. Existing trial-and-error approaches require multiple costly full-scale compression runs to reach at target requirements, making them impractical for such online use. To address this, we propose OPTRD, a compressor-agnostic framework that efficiently models rate-distortion trade-offs across multiple lossy compressors by analyzing spatial data traits at reduced resolutions. Evaluated using 3 state-of-the-art lossy compressors on 30 scientific datasets from 4 HPC applications, OPTRD incurs only similar to 5% average estimation error and achieves over 100x runtime speedup compared to trial-and-error methods, significantly improving optimal compressor selection during such data transfer use cases.
Public availability of data about the supply and utilization of Dockless Electric Scooters (DESs) is needed to develop smart mobility applications to promote the use of sustainable, shared, and equitable transportation. Prior studies have analyzed some attributes of DES usage, but they are limited in three aspects. 1) Data collection methods: Prior studies relied on DES providers to provide usage data to authors. Unfortunately, convincing DES providers to provide complete usage data is extremely hard, if not impossible. Thus, this approach is not scalable and further limits the ability to study DES usage trends only in the regions and only at the times for which the DES providers provide the data. 2) Global DES usage analysis: Prior studies have not analyzed how DES usage varies in different parts of the world. 3) Aspects of the usage analysis: Prior studies have not analyzed important DES usage attributes such as the availability of DES in minority-concentrated areas, temporal trends of DES trips from different types of locations, etc. This paper fills the gap in prior work by addressing these three limitations. For data collection methods, we present novel techniques to continuously collect DES usage data in any city worldwide without relying on DES providers. For global DES usage analysis, we study the spatiotemporal DES usage trends of 12 DES providers in 11 cities worldwide. For the aspects of the usage analysis, we provide extensive analysis of various aspects of DES usage that have previously either never been studied or studied only briefly. These aspects include DES supply trends, competition among providers, supply and trip hotspots, equitable DES availability, and so on. The methods and findings that we present in this paper aid in understanding DES usage across the world and offer actionable insights to DES stakeholders.
GPU-accelerated molecular dynamics (MD) simulations have enabled significant advances in modeling biomolecular systems and extended timescales. While system size, typically measured by atom count, is often used to estimate simulation performance, the impact of biophysical system composition remains underexplored. This study investigated GPU-accelerated simulation throughput using GROMACS on an NVIDIA Titan RTX GPU across eight atom count-controlled systems (approximately 82,000 atoms) with varying biophysical compositions. The benchmark set included water-only, protein-ligand, membrane-only, and membrane-protein systems, matched to the standard benchMEM system for controlled comparison. MD systems with high water and protein content consistently demonstrated superior GPU performance exceeding 30% compared to lipid-rich systems, despite equivalent atom counts. Membrane-protein systems exhibited modest improvements over membrane-only systems, likely due to more favorable domain decomposition properties. These findings provide quantitative insight into the influence of system composition on GPU performance and offer practical information for optimizing MD simulation systems in terms of biological complexity and computational efficiency.
This paper investigates a multi-user downlink communication system where an unmanned aerial vehicle (UAV)-mounted intelligent reflecting surface (IRS) assists a base station employing rate-splitting multiple access (RSMA) under hardware impairments. A sum-rate maximization problem is formulated, jointly optimizing the common rate allocation, beamforming vectors, and IRS phase shifts. To tackle the non-convexity, an alternating optimization framework is developed in which each subproblem is efficiently solved using successive convex approximation (SCA). Simulation results demonstrate that the proposed design outperforms conventional benchmarks.
Deploying AI models on low-power hardware remains limited by severe constraints in memory, compute throughput, and energy efficiency. This project explores whether the ESP32-S3, a low-cost microcontroller with SIMD acceleration, can run a lightweight AI model efficiently.
With the widespread application of large language models (LLMs), their advantages in context analysis and semantic inferencing are becoming increasingly prominent, and we are attempting to introduce them into the realm of DNS repair. As is well known, DNS, as the core of Internet infrastructure, has complex policies and a fragile system, where even a small misconfiguration can lead to catastrophic service failures. Especially in large-scale networks, analyzing detected errors and generating repair solutions often requires operators to invest a significant amount of time and effort. This paper proposes a scalable framework, DNSHolmes, designed to leverage LLMs for generating DNS configuration repair solutions. Specifically, this method first addresses the numerous potential root causes in large-scale errors by abstracting them into a small number of State Equivalence Classes (SECs). It then adopts a deterministic finite automaton (DFA) to compute a Critical Path Graph (CPG) for each class, precisely isolating the minimal set of records responsible for the failure. Crucially, the CPG serves as a focused, verifiable context for a Large Language Model (LLM), guiding it to generate accurate patches while overcoming the fundamental context-length limitations that help LLM with effective repair reasoning. Our evaluation on large-scale public datasets and a real-world campus network demonstrates that DNSHolmes can reduce operational effort by 78.4% and achieve efficient repair in large-scale DNS configuration.
Traditional Database (DB) systems use a DB buffer, a page-based cache management system, to load data and indexes from block storage devices into byte-addressable main memory. However, this approach is inefficient in terms of space and I/O when key-value pair sizes are significantly smaller than the page size. Inserting a single key-value pair results in reading and writing an entire page, consuming a full page's worth of memory in the buffer. Moreover, the entire page is immediately loaded even when just a single key-value pair is inserted into the page. Also, an infrequently accessed page is likely to be evicted to disk before any subsequent accesses occur. We present TurboIndex, a hybrid cache management scheme that combines a record-based cache with the traditional page-based cache to address these inefficiencies. TurboIndex first accumulates key-value pairs from cold pages in the record-based cache. It then identifies hot pages, those that are likely to benefit from page-based caching, and migrates them collectively from the record-based cache to the page-based cache. This strategy increases effective cache capacity without significantly increasing memory usage, while also improving performance by reducing disk I/O. TurboIndex achieves up to a 4.5x improvement in pure write workloads and at least a 1.8x gain on the write-heavy YCSBA benchmark.
Learned indexes have received significant attention for their potential to dramatically outperform traditional treebased indexes in both speed and space efficiency. Their core strength lies in using predictive models to estimate the position of a key within a sorted array. To handle complex key distributions and support frequent insertions in dynamic workloads, learned indexes typically organize multiple models hierarchically in a tree structure. These indexes perform best when a model can accurately predict a key's location. However, existing learned indexes often require traversing several models to reach the one responsible for a target key. Moreover, if the prediction is imprecise, an additional local search (the last-mile search) is needed. These overheads before and after model execution can significantly degrade performance, sometimes approaching that of conventional indexes like B+-trees. In this paper, we propose LASER, a new learned index design that tackles these inefficiencies by leveraging two common patterns in real-world workloads. First, LASER exploits key access locality, where some key ranges are more likely to be accessed. Second, it detects linear key distributions, enabling precise prediction without last-mile search. LASER adaptively promotes models covering long key ranges to the top of the model tree, reducing the number of model traversals. For key ranges with perfectly linear distributions, it employs models that guarantee a direct hit (on the exact key position), eliminating the need for further search. We implemented LASER and conducted extensive evaluations. The results show that LASER outperforms state-of-the-art learned indexes such as LIPP and ALEX, as well as traditional indexes, like ART, by up to 1.6 to 5.5 times.
Mobile Crowd Sensing (MCS) has revolutionized the way we collect environmental and urban data by leveraging a tremendous number of distributed sensors equipped on wearable devices like smartphones, smart watches, and so on. Gathering continuous data from these sensors allows us to generate a complete spatiotemporal field for the desired measurements. However, obtaining continuous sensing data at all locations and times is often infeasible due to resource limits, cost constraints, and practicality. To address this problem, we proposed a Learned Gumbel-Based Active Sparse MCS framework utilizing an Encoder-Decoder Time Series Transformer that reconstructs the full spatiotemporal field for a given time cycle and a small subset of available sensors. Notably, an active sensor selection layer is included, which leverages the Gumbel distribution to create noise from learned weights, helping to dynamically select the most effective sensors for the next sensing cycle. We verified the effectiveness of our proposed model by comparing it to random selection, attention-scoring, and statistical methods. Our Gumbel-Based selector achieves higher performance in both firstand second-order norm metrics.
Efficiently scheduling Machine Learning (ML) training tasks in a GPU data center presents a significant research challenge. Existing solutions commonly schedule such tasks based on their demanded GPU utilization, but simply assume that the GPU utilization of each task can be approximated as a constant number (e.g., by using the peak value), even though the ML training tasks commonly have their GPU utilization varying significantly over time. Using a constant number to schedule tasks can result in an overestimation of the needed GPU count and, therefore, a high capital expense for GPU purchases. To address this, we design CorrGPU, a correlation-aware GPU scheduling algorithm that considers the utilization correlation among different tasks to minimize the number of needed GPUs in a data center. CorrGPU is designed based on a key observation from the analysis of real ML traces that different tasks do not have their GPU utilization peak at exactly the same time. As a result, if the correlations among tasks are considered in scheduling, more tasks can be scheduled onto the same GPUs, without extending the training duration beyond the desired due time. For a GPU data center to be constructed based on an estimated ML workload, CorrGPU can help the operators purchase a smaller number of GPUs, thus minimizing their capital expense. Our hardware testbed results demonstrate CorrGPU's potential to reduce the number of GPUs needed. Our simulation results on real-world ML traces also show that CorrGPU outperforms several state-of-the-art solutions by reducing capital expense by 20.88 %.
Semantic segmentation in complex natural environments remains a challenging task due to diverse object appearances, background clutter, and class imbalance. This study evaluates state-of-the-art segmentation models, categorized into UNet-based, encoder-decoder, and transformer-based architectures, using a dataset focused on aquatic plant segmentation in a natural environment. The model is evaluated using Dice coefficient, F1 score, IoU and overall accuracy, along with computational complexity measured by parameter count and GFLOPs. Results indicate that backbone-enhanced models outperform base configurations by providing deeper and more robust feature representations. Then, explainable AI (XAI) techniques are used to evaluate the model attention and ensure the precise focus of the network on ecologically meaningful regions. These insights not only highlight the importance of architectural depth and input diversity but also demonstrate how XAI can support iterative model refinement. As a result of these combined efforts, segmentation performance is improved from a baseline dice coefficient and accuracy of 50.76 and 85.07 % to a maximum 75.57 and 94.94 % respectively. This comprehensive analysis provides valuable insights of selecting suitable models in resource-constrained environments and advances the development of robust segmentation systems.
Website fingerprinting (WF) attacks undermine the privacy promised by anonymizing networks such as Tor by inferring the websites a user visits from encrypted-traffic side-channels. Recent criticisms of the single-tab assumption have shifted attention to the more realistic multi-tab setting, where concurrent page loads create severe noise. Existing multi-tab studies rely on direction sequences that ignore temporal structure and therefore provide only limited discriminative power. Our experiments show that packet-level timestamps do carry extra signal, yet their raw form is fragile under overlapping tabs and timing-obfuscation defences. We propose the Packet-Burst Counts (PBC) feature-a 4xL matrix that, for each time interval, stores the counts of upstream packets, downstream packets, upstream bursts and downstream bursts. PBC preserves coarse temporal structure while discarding noisy fine-grained timings, striking a balance between expressiveness and robustness. Building on PBC, we design PBC-MWF, an end-to-end framework that (i) uses a residual CNN to learn local embeddings and (ii) applies an adaptive sparse transformer to capture global correlations while suppressing tab-overlap noise. Unlike prior work, which evaluates on datasets with a fixed number of tabs and reports top-k accuracy, we additionally merge datasets with varying tab-count and perform threshold-based inference; the threshold is tuned on the validation set before testing. To the best of our knowledge, PBC-MWF is the first WF framework to simultaneously address the multi-tab setting's challenges of fine-grained webpage identification and resilience against WF defences. Evaluations on three public multi-tab datasets demonstrate PBC-MWF's enhanced robustness: compared against nine baselines, it surpasses the best prior method-improving F1 by up to 12.4 % on site-level, 8.4 % on page-level, and over 10 % under defenses.
Determining the maximum possible flow (max-flow) between a source vertex and a sink vertex in a network with given capacities on every edge is a fundamental graph problem found in many domains. The Push-Relabel (PR) algorithm is the leading approach for solving the max-flow problem and has been parallelized for GPUs. However, existing approaches suffer from scalability issues and may perform poorly on large graphs. This paper focuses on improving the efficiency of the PR algorithm on GPUs, presenting several implementation and parallelization improvements compared to the state of the art. Our improvements include fast global relabeling on the GPU, worklists to minimize wasted work and CUDA threads, and a two-level parallelization scheme to improve load balancing. Our approach is, on average, $5.75 x$ faster than the fastest prior GPU implementation and $5.08 x$ faster than the fastest CPU code.
Identity privacy leakage through the wireless interface in mobile networks represents a persistent security challenge and a long-standing concern that network designers have aimed to address. Despite the remediation introduced in 5G standards, identity privacy attacks targeting the wireless interface continue to pose a potential threat. In this paper, we present an identity mapping attack based on time series similarity in 4G and 5G networks. This attack enables an adversary with no privileges to map a victim's social media account to their RNTI by sending a single image message to the victim and measuring the similarity of time series extracted from the generated downlink traffic. To improve the attack success rate, we specifically design an elastic similarity measure for time series, tailored to the properties of the data collected during the attack. We investigate the feasibility of the attack under various scenarios, achieving a success rate of 83% for a single attempt and nearly 100% when conducting two or three attempts. Our work provides new insights into the vulnerability of 4G/5G standards to identity privacy attacks.
Deep Learning-based Side-Channel Analysis (DLSCA) has demonstrated its potent attack capability against cryptographic systems, enabling the extraction of secret information from leaked physical signal traces. Unfortunately, the availability of real traces for classifier training is often limited due to restricted data collection or protective countermeasures. The generation and enhancement of signal trace is crucial in practical DL-SCA analysis. Existing methods often struggle to accurately model the feature distribution of real traces, and their applicability to emerging post-quantum cryptographic (PQC) algorithms remains unexplored. In this paper, we propose DMSCT, an attention-enhanced diffusion model framework for side-channel trace generation. DM-SCT precisely models the feature distributions of real traces through a conditional injection mechanism and an enhanced attention mechanism. We evaluate DM-SCT across three PQC algorithms: NTRU, Saber, and Kyber, demonstrating its ability to faithfully replicate the leakage features of original side-channel traces. By integrating DM-SCT into DL-SCA, we validate its practical efficacy. Compared to training with real traces alone, DM-SCT improves the accuracy of recovering individual key coefficients by 0.62% to 1.6%. Furthermore, DM-SCT reduces the required number of real training traces by 20% to 25%, outperforming state-of-the-art trace generation methods.
Internet service providers often fingerprint the type of application, e.g., video streaming or video conferencing, from a network traffic trace, for security and traffic engineering purposes. Several traffic classification methods have been proposed to apply machine learning techniques to packet timestamp and packet size information in traffic traces to perform such fingerprinting. However, classification accuracy drops when the machine learning methods are trained on traces that do not capture diverse end point and underlying network conditions. In this work, we compare a number of augmentation techniques that attempt to mitigate this problem, and evaluate their effectiveness when we systematically change specific end point and underlying network conditions. We find that the Rosetta traffic classification method that uses packet size sequences benefits much more from current augmentation methods than the mini-FlowPic method that uses both packet timestamps and sizes. We believe that there is a critical need for customized augmentations that enrich traffic traces, while being faithful to behaviors of real users and real protocols.1
Scientific simulations can produce petabytes of data in a single run. Even when aggressively compressed, such data cannot be stored locally on a scientist's workstation. This paper introduces omni-homomorphic compression (oHC), a technique to generate extremely compressed versions of the data that are still useful and that can be processed locally without the need for decompression. Since the main data is kept separate, it can be preserved at any user-specified error bound and retrieved at will. Compared to other lossy and/or homomorphic compressors, oHC delivers not only orders of magnitude higher compression ratios but also some of the highest throughputs. For example, running on an RTX 4090 GPU with an error bound of 1E-2 on single-precision data, it provides an average of 270 and 253 GB/s throughput for compression and decompression, respectively. When extracting the arithmetic mean from a large dataset, oHC's extremely compressed version is 95,000 times faster than the fastest homomorphic compressor from the literature.
Previous research on the ethical decision-making capabilities of large language models (LLMs) has largely relied on classification-based approaches, focusing on binary judgments of moral acceptability. While prior findings indicate strong alignment between LLMs and human judgments in morally unambiguous scenarios, the reasoning processes behind these decisions remain underexplored. This paper introduces an explainability-driven workflow to assess LLM moral reasoning by prompting models to provide both binary ethical judgments and accompanying justifications. We evaluate the semantic similarity of these justifications across varied prompting styles using latent semantic analysis. Gemini achieved the most consistent moral justification similarity across frameworks, with LLaMA3 close behind and LLaMA2 and Mistral showing greater variability, especially in deontology.
Most HPC resources that are available to a wide public userbase (for example XSEDE/ACCESS resources) have fairly basic user guides for building and running applications on the systems. Large performance differences, however, can occur if users are not experienced with detailed job configuration flags and their specific application behavior. This performance difference affects job quality-how well the job used its resources to accomplish the user's goals. However, a one-dimensional idea of job quality cannot capture all of the intricacies around how or why users might select a particular configuration. This paper presents four metrics for four different views of job quality, and then presents results from experiments over several applications that show the variability in quality that results from a variety of job configuration tuning parameters. This work can point the way towards guidelines for how users can potentially improve their own quality goals of their applications and thereby use the HPC resources most beneficially. One significant result we saw across applications is that high thread counts with low processes-per-node counts are generally not the best configurations; most applications do better with low thread counts and more processes per node.
In the era of Industry 4.0 (4IR), the Internet of Things (IoT) drives the transformation of conventional operation mode into intelligent systems through interconnecting smart devices to monitor, analyze, and optimize the target applications. In order to achieve energy-saving data transmission, a routing protocol, called RPL, has been specified for resource-challenged IoT devices and networks. In the context of 4IR, the IoT technology is being widely used for mission-critical systems, and the data collected by IoT devices might contain privacy-sensitive information. In addition, the IoT gateway could be compromised due to the lack of necessary and persistent physical and/or logical security protection. Hence, the protection of data security and privacy becomes a crucial factor for RPL-based IoT systems to realize their quality of service requirements and objectives successfully. In this paper, we propose a secure and privacy-preserving data aggregation approach, called SPARDA, for IoT devices and networks running the RPL routing protocol. SPARDA is realized with physical unclonable function, homomorphic encryption, and trapdoor function, and is perfectly integrated with the RPL routing protocol to prevent the malicious IoT gateway from either accessing, falsificating, or corrupting the real-time data from IoT devices throughout the data gathering and summarization phase. We choose an automatic security protocol verification tool, widely known as AVISPA, to analyze and verify the security specification of SPARDA. We also conduct an experimental study to evaluate the performance of SPARDA by comparing with benchmark methods. The experimental results indicate that not only does SPARDA protect IoT networks from malicious gateway attacks, but it also outperforms existing schemes in terms of computation and storage overheads while satisfying all critical security and privacy criteria.
