Papers

Data-driven methods for earthquake and tsunami early warning rely on large ensembles of rupture scenarios and their resulting waveforms, but generating such datasets with repeated high-fidelity seismic and tsunami simulations is prohibitively expensive. We exploit the linear time-invariant structure of both dynamics to precompute elastic Green's functions and acoustic-gravity adjoint responses, reducing the source-to-waveform map to two consecutive convolution operators. We evaluate these convolutions with a distributed, FFT-accelerated GPU pipeline that partitions the large seafloor grid across GPUs and directly generates the final observation waveforms. We demonstrate the scalability of this pipeline for the Cascadia Subduction Zone with 963 subfaults, 2,416,530 seafloor grid points, 64 observation locations, and 256 timesteps, requiring 9.45 TiB of aggregate GPU memory. On 64 GB200 GPUs within one NVL72 domain, the pipeline generates waveforms in 24 ms per rupture once the response operators are resident, enabling large rupture ensembles to be evaluated within minutes.

AI inference clusters are increasingly constrained by instantaneous power, not just energy: grid operators condition new capacity on demand response, imposing time-varying power caps. Existing LLM serving systems optimize a static energy objective or shed fixed priority tiers under load; either way, goodput collapses when the power envelope moves. An LLM pipeline is not a uniform load: compute-bound prefill loses throughput almost linearly with GPU frequency, memory-bound answer decode sustains it down to $0.57\times$ nominal, and reasoning's thinking phase couples KV-cache capacity to scheduling – so a cap should be steered to where each watt costs the least performance. PowerSlider does so with a new Flex SLO contract that turns bounded user slack into an optimization constraint, prefill–think–answer disaggregation exposing per-stage frequency and KV control, and a Karush–Kuhn–Tucker (KKT) online solver re-solving within 7.7 ms of every cap change, backed by a consolidated fail-safe that power-gates drained instances when DVFS bottoms out on static power. On SGLang with production traces, \sys{} sustains 78.3% online goodput at a 30% cap reduction versus 47.6% for the best of five baselines ($1.64\times$), holds latency-critical tails within $1.3\times$ of nominal (baselines: $2.3$–$6\times$, up to $12\times$), and delivers 92% mean goodput through a replayed CAISO grid-emergency day bottoming at $0.41\times$ (54% at the trough; every baseline below 7%).

Chain-of-thought (CoT) prompting improves LLM reasoning by decomposing complex problems into intermediate steps, but its sequential nature increases decoding latency and memory usage. Mixture-of-Experts (MoE) models scale capacity through sparse expert activation, yet their full expert weights often exceed GPU memory and require costly GPU-CPU transfers. Existing runtimes treat all tokens uniformly, overlooking a key structural property of CoT traces: consecutive reasoning stages exhibit coherent and predictable expert activation patterns. Ignoring this stage-level regularity leads to inefficient caching and unnecessary data movement. We propose SAEM, a stage-aware MoE inference runtime that detects reasoning stage boundaries and exploits stage-level activation coherence to guide expert placement. SAEM combines stage-aware caching, expert-aligned token repacking, and in-situ CPU execution to reduce data transfer and kernel fragmentation. On mathematical and scientific reasoning workloads, SAEM achieves an average 1.33x throughput improvement over the strongest state-of-the-art caching and offloading baselines under constrained GPU memory, rising to 1.54x when calibration data matches the workload, demonstrating the effectiveness of stage-aware, locality-driven MoE inference for CoT reasoning.

Federated fine-tuning enables large language models to adapt on edge devices without centralizing private data, but practical deployments must address hardware instability and adversarial update corruption together. Thermally constrained clients may throttle, slow local training, or delay synchronous aggregation, while Byzantine clients and communication-layer adversaries can corrupt the updates used to form the global model. To address these challenges, we present Thermo-FL, a thermal-aware federated LoRA fine-tuning framework that uses device temperature as an active control signal for local adapter training and sparse update transmission. On the client side, Thermo-FL adjusts the active LoRA-layer fraction and transmitted update density as devices heat or cool, reducing workload under thermal stress. On the server side, Thermo-FL introduces TERRA, a robust aggregation pipeline for dynamically sparse LoRA updates that combines norm filtering, mask-aware directional validation, adaptive active-coordinate clipping, and mask-aware aggregation. We evaluate Thermo-FL using both a large-scale emulator and a Jetson-based physical testbed. In the emulator, Thermo-FL improves robustness under adversarial sparse aggregation and achieves the strongest BoolQ accuracy across clean and attack settings while remaining competitive on GSM8K. In the physical prototype, Thermo-FL stabilizes device temperature, reduces compressed upload size through bitmap sparse encoding, and preserves GSM8K utility under sign-flip/scale and MITM perturbations. These results show that secure edge LLM adaptation should jointly consider hardware behavior, workload regulation, sparse communication, and aggregation robustness.

Modern Approximate Nearest Neighbour Search (ANNS) applications operate over continuously evolving vector collections and require graph indexes that sustain high-throughput searches while incorporating insertions and deletions with high recall. However, most GPU graph indexes are static or provide limited update support. Updates require neighbour discovery, reverse-edge creation, pruning, and deletion-induced graph repair; executing these operations concurrently introduces redundant distance computations and conflicting accesses to shared adjacency lists. Background-rebuild-based deletion further incurs substantial computation, additional memory consumption, and interference with foreground queries. We present GrAND (GPU-based Dynamic Graph Indexes for Approximate Nearest Neighbour Search), a GPU-native collection of dynamic-update algorithms for two popular graph indexes, Vamana and CAGRA. GrAND consolidates graph repair across a batch, eliminating redundant pruning computations, and employs a lock-free find-and-replace strategy for parallel adjacency-list updates. For reliable in-place deletion, GrAND constructs an on-demand reverse graph on the GPU, accurately identifying incoming edges without permanently duplicating the index. We evaluate GrAND on seven real-world datasets across five streaming workloads, comparing it against SVFusion and FreshDiskANN-GPU (our GPU adaptation of FreshDiskANN). GrAND improves overall workload throughput by 2.2x-8.7x and 6.5x-25.4x, respectively, while maintaining high search throughput and recall over sustained updates.

High-performance GPU kernels underpin modern deep learning and scientific computing. As workloads become increasingly diverse and GPU hardware evolves rapidly, developing efficient methods for automated GPU kernel generation and optimization has become increasingly important. Existing LLM-based methods typically optimize within a fixed implementation space, limiting either optimization flexibility or search efficiency. We propose HIERA, a hierarchical search-space planning framework for GPU kernel optimization. HIERA constructs contract-augmented task specifications, selects an appropriate implementation space across PyTorch operators, CUDA libraries, and custom CUDA kernels, and uses profiling feedback and expert knowledge to guide structured iterative refinement. Experiments on KernelBench across multiple various workload levels and base LLMs show that HIERA delivers stronger overall implementation validity, sample efficiency, and optimization performance than existing training-free methods, while remaining competitive with the training-based CUDA-L1 without additional model training. A case study on a specialized stencil operator from scientific computing further achieves a \(1.53\times\) speedup over cuDNN, demonstrating the potentiality of the general framework beyond standard machine-learning workloads.

The transition of Earth-system models to exascale is often hindered by rigid, monolithic Fortran codebases and maintenance-heavy compiler directives. While high-level DSLs offer a solution, they frequently fail due to cumbersome integration. We present the integration of a Python-based ICON dynamical core into the original Fortran simulation code. Leveraging the GT4Py DSL and the Data-Centric (DaCe) optimization framework, we demonstrate that high-level Python can be seamlessly integrated into legacy infrastructure without performance loss. Our results challenge the assumption that Python orchestration introduces prohibitive HPC overhead. In production-grade global simulations, our Python dynamical core achieves a 20–30% performance improvement over the highly-optimized Fortran+OpenACC implementation, with a 10% improvement on the total time for a coupled setup. Driven by advanced data-flow optimizations and automated kernel fusion, this approach replaces hardware-entangled directives by generating optimized device code from a single, portable Python source. This work proves that Python can provide a sustainable, efficient, and hardware-agnostic future for global climate modeling.

Decentralized Federated Learning (DFL) promises trust-free collaborative learning by replacing the centralized parameter server with peer-to-peer model exchange. However, this architectural shift fundamentally reshapes the threat landscape. Without globally coordinated aggregation, DFL becomes particularly susceptible to backdoor attacks, in which malicious participants implant persistent hidden behaviors while maintaining high clean-task performance. In this paper, we argue that the robustness of DFL has been significantly overestimated. Existing studies rely on simplified threat models, non-adaptive adversaries, fragmented evaluation protocols, inconsistent communication topologies, and ad hoc training configurations, leading to an incomplete understanding of DFL security. To address these limitations, we present BackDFL, a unified benchmark for systematically evaluating DFL under realistic and adaptive backdoor attacks. Through extensive experiments, BackDFL exposes critical failure modes of decentralized learning. Our results demonstrate that both state-of-the-art Byzantine-robust DFL methods and adapted FL backdoor defenses fail under modest malicious participation rates (as low as 15%), especially in heterogeneous settings, while their robustness varies substantially across communication graph topologies.

Satellites are becoming programmable computing platforms capable of running increasingly demanding AI workloads. This shift raises a systems problem: how can AI services remain deployable, manageable, and recoverable after launch when compute capacity, connectivity, energy, and thermal headroom vary over orbital time? This paper develops a systems vision for AI infrastructure in space. We define it as the systems layer that manages AI capabilities across spacecraft, orbital networks, ground stations, and cloud backends, while treating orbital and physical state as part of the resource model. We synthesize relevant foundations from terrestrial AI infrastructure, satellite networking, and satellite edge computing, and examine the physical constraints that directly shape system design. We further ground this vision in three in-orbit case studies spanning the node, platform, and service levels. Telemetry from BUPT-1 satellite shows that usable compute capacity is bounded by thermal and energy envelopes. SateLight on BUPT-2 satellite reduces application-update transmission latency by 56.54% on average and up to 91.18%, with 100% update correctness. A stateful VLM serving case further shows that thermal interruptions make execution-state recovery a first-class systems problem. These observations motivate a research agenda for space-native resource management, lifecycle support, and sustained AI service across space and ground.

Modern open models are hybrids: most layers are linear-attention (Gated DeltaNet, GDN) layers carrying a small fixed-size recurrent state instead of a growing key-value (KV) cache. This makes ordinary decoding memory-efficient, but hurts speculative decoding. To verify a batch of draft tokens and then roll back the rejected ones, today's systems snapshot the full recurrent state at every draft position for GDN layers, and those snapshots cannot be shared across branches of a draft tree, so a wide, high-acceptance tree becomes memory-infeasible. We remove the snapshots. Using a tree-structured WY transform of the gated delta rule, we compute every draft node's output with a single triangular solve and reconstruct only the one accepted state on commit, storing a small pseudo-value matrix instead of per-node states; the derivation depends only on the gated delta rule, not on any other architectural detail. In serving benchmarks on two scales of one hybrid model family (Qwen3.5 35B and 397B) this cuts speculative recurrent-state memory and KV-cache pressure at identical acceptance length, turning the freed HBM into higher throughput and much lower time-to-first-token (TTFT) wherever memory binds, and costing a few percent where it does not. For tree width the same memory buys affordability: a wider, higher-acceptance draft becomes possible, though not yet a throughput win.

Edge nodes are capacity-constrained by design, yet many edge workloads can trade result quality for resource efficiency at runtime. Existing edge pricing mechanisms largely treat requests as fixed-configuration submissions and rarely exploit per-request quality flexibility under overload. We present PRICE, an incentive mechanism that couples a utilization-dependent price signal to per-request quality selection. As utilization increases, rising acceptance prices make resource-intensive variants less likely to be selected, shifting accepted requests toward lighter execution and allowing the node to serve significantly more requests while operating near capacity. Evaluation on real hardware under sustained overload shows that PRICE outperforms both fixed-allocation and dynamic-pricing baselines in accepted throughput and CPU utilization. The results are robust across pricing function families, task-duration distributions, and client populations. Result-quality flexibility is a powerful but underused control dimension for overload management at the edge, and pricing is an effective mechanism to exploit it.

Managing the energy consumption and power efficiency of parallel applications is a significant issue in both HPC environments and in the cloud. As emerging applications continue to push against the memory wall of modern machines, the growing imbalance between compute and data movement creates new opportunities to intelligently tune CPU power consumption. Unfortunately, existing frequency and voltage scaling techniques do not adequately capture fine-grained changes in memory access behavior, rendering the compute/data access imbalance invisible to the components of the system that could capitalize on it, thus leaving potential power savings on the table. In this paper, we propose MEMPOWER, a flexible, model-based approach to exposing compute/data movement imbalance that characterizes the fine-grained memory behavior of parallel workloads. This characterization then informs our automated software framework which can statically instrument the application binary with model-determined voltage/frequency transitions that balance fine-grained changes in memory access behavior with the costs of hardware transitions. Using MEMPOWER, we demonstrate a reduction in EDP of 6% to 42% on a range of HPC benchmarks with minimal impact on execution time when compared to the standard OS/hardware-managed power control mechanism.

Convolution is a principal computational bottleneck in deep neural networks, and its efficiency depends on tight integration between algorithms and GPU hardware. Existing GPU convolution methods suffer from large memory overhead, poor cache utilization, limited effectiveness across kernel sizes, or numerical instability. This work extends the im2win paradigm – a universal, memory-efficient convolution method with contiguous memory access for all kernel sizes – to run efficiently in full precision on CUDA cores and half precision on tensor cores. By introducing new kernel designs and optimizations such as zig-zag memory access and asynchronous data movement, im2win efficiently exploits hardware-accelerated half-precision matrix multiply-accumulate operations. Across twelve CNN benchmarks, im2win achieves up to 2.8x higher TFLOPS than its CUDA core implementation, 1.4x higher than cuDNN, and 6.4x higher than GEMM-based convolution with cuBLAS, while using as little as 53% and 35% of their memory, respectively. These results establish im2win as a unified, high-performance convolution framework for modern GPU architectures.

There continues to be growing interest in using standard language constructs for parallel and accelerated HPC computing, avoiding the need for (sometimes vendor-specific) external APIs. For Fortran applications, language features such as 'do concurrent' loops open the door for compilers to implement multi-threaded, GPU-accelerated, and even distributed multi-node code with only the standard language. Here, we explore the current status of using 'do concurrent' for GPU-accelerated Fortran applications across three major GPU vendors (NVIDIA, AMD, and Intel). Using a production application, we test their current capabilities, showing where the standard language alone can be used, and where augmenting the code with a directive-based API (e.g., OpenMP) is still desirable or required. Multi-GPU tests are performed with GPU-aware MPI libraries. We find that the three GPU vendors can now GPU-accelerate pure Fortran (zero directives), but that manual data movement directives can help with performance and compatibility. The results show that there is rapid advancement towards making GPU-accelerated scientific HPC code performance portable using the Fortran standard language.

Earth observation (EO) programmes generate data at volumes that exceed the transfer and storage capacity of most institutional networks. Public cloud platforms address this for well-resourced organisations, but institutions across the Atlantic basin face constraints in connectivity, sovereignty and funding that make on-premises infrastructure the only viable path. Cloud-native data formats enable efficient partial reads, yet their performance depends on the bandwidth of the underlying network fabric, a dependency rarely measured in isolation. This paper presents a replicable, network-centric architecture for on-premises EO data access, evaluated at its first operational deployment: the AIR Data Centre, founding node of the Atlantic Cloud. The system comprises a MinIO object storage cluster on a 100 GbE fabric, a PostGIS metadata catalogue and an OGC API-EDR access layer. We characterise the fabric under sustained parallel load, evaluate object storage throughput for EO-representative workloads, and compare measured performance against throttled baselines on identical hardware, isolating network bandwidth as the sole variable. Multi-site replication benchmarks with partner institutions characterise the federation primitive the model depends on. Network bandwidth is the dominant constraint on storage throughput for bulk EO data access up to a threshold; beyond it, endpoint memory topology rather than capacity governs how much bandwidth a system can use. For this hardware class that threshold lies above 10 Gbps per server. Below it, network capacity alone sets what the facility can deliver; above it, the return on further network investment depends on endpoint memory provisioning, which can be deferred and bought later.

The rapid proliferation of memecoins on blockchain platforms has increased the risk of fraudulent activities, particularly rug pulls. While previous studies have focused on Ethereum-based tokens, this paper shifts the spotlight to Solana, the leading blockchain for memecoins by trading volume and token count. Unlike Ethereum, where rug pulls often exploit smart contract backdoors, Solana memecoin rug pulls are predominantly driven by liquidity manipulation and social dynamics. This research pioneers large-scale rug pull early detection in the Solana ecosystem by assembling a dataset of 6.4 million tokens over 7 months. Market analysis reveals that a vast majority of these memecoins exhibit rug pull characteristics within one hour of launch, highlighting the urgency of short-horizon prediction. Despite the absence of code-level features, we demonstrate that classic machine learning models, particularly Gradient Boosting (XGBoost), achieve robust performance in detecting potential rug pulls using only the first 5 minutes of trading data. Furthermore, we evaluate cross-platform generalization between PumpFun and Raydium, revealing that multi-source data fusion significantly mitigates domain shift and improves detection reliability. This study advances the understanding of DeFi fraud on high-throughput chains and provides a practical framework for protecting investors.

Retrieval-Augmented Generation (RAG) lets Large Language Models (LLMs) pull in up-to-date, domain-specific information instead of relying only on what they were trained on. Yet most RAG systems still draw from centralized databases with limited oversight, making it difficult to verify where a document came from, whether it has been tampered with, or whether it should be trusted at all. This is a serious problem in domains where both the timeliness and accuracy of retrieved content are critical, such as healthcare, finance, logistics, and legal case law, where a wrong or manipulated document can directly lead to bad decisions. We present TrustRAG, a committee-based, blockchain-backed RAG system: before a document is used, it is certified by a committee of domain experts through a zero-knowledge protocol, and the committee's hidden scores are combined via secure multi-party computation into a trust score that any client can verify. These scores, along with the underlying document data, are maintained jointly across chains through hash commitments, so no document or score can be silently altered or dropped, and every ranking can be independently replayed and checked.

Modern web applications increasingly require computationally intensive processing, yet JavaScript, the dominant language of the web, has traditionally been limited to a single-threaded execution model. Node.js Worker Threads and browser Web Workers provide low-level mechanisms for parallel execution, but developers lack high-level abstractions that capture recurring parallel structures as reusable patterns. In this paper, we present ParaWeb, a TypeScript library that implements ten parallel programming patterns for server-side Node.js, client-side browser environments, and WebGPU compute shaders. ParaWeb provides three implementation variants for each pattern: a message-passing (MP) variant based on structured cloning via postMessage, a shared-buffer (Shared) variant that uses SharedArrayBuffer with typed array views, and a GPU variant that uses WebGPU compute shaders for hardware-accelerated execution. We describe the architecture, design decisions, and pattern-specific implementation strategies, and we evaluate the performance of all thirty implementations across three data sizes. Experimental evaluation results show that the CPU-based variants achieve speedups of up to 11.6x with 16 threads for compute-bound patterns, while the GPU variants reach speedups of up to 260x for compute-bound patterns with high arithmetic intensity such as Farm, Scatter, Reduce, and Map. A case study on five image-convolution filters further shows that GPU acceleration reaches up to 414x speedup over single-threaded CPU on non-separable kernels, with consistent scaling across 1024x1024$, 2048x2048$, and 4K images.

Kubernetes is the de-facto platform for container orchestration. Its scheduler combines resource capacities with label-based affinity and anti-affinity rules, and the interaction of these features can make the eventual placement of a pod. In this paper, we study the pod-deployability problem: given an initial cluster, a pod type, and a designated node, does some legal sequence of pod deployments and deletions cover the target pair? We give three complexity results. First, when dynamic constraints contain no affinity (anti-affinity is allowed), pod-deployability is decidable in polynomial time. Second, required affinity together with required anti-affinity makes the problem PSPACE-complete. Third, required affinity alone is already enough for PSPACE-completeness on a single node with one scalar capacity. The lower bounds encode, respectively, 1-safe Petri-net coverability and bounded black pebbling. These results isolate two independent sources of state-space complexity in Kubernetes scheduling: logical exclusion and resource-bounded prerequisite management.

Prefix caching avoids prefill only when a repeated request returns to a server that still holds the prefix KV. Cache-blind balancing disperses that reuse; fixed affinity preserves it but can overload a server. CacheRoute resolves this tradeoff with a periodic routing plan. It admits high-rate keys to a stable warm set and places their assignments by expected load. Hot keys may use more than one destination, although every key in our primary semi-synthetic aggregate uses exactly one. On Llama-3.3-70B in fp8 across 60 H100 GPUs, CacheRoute sustains 176+/-11 QPS at a 3.5-s p99 SLO, 2.3x the strongest of five baselines. Served KV-cache hit rate rises from 64.1+/-1.3% under cache-blind balancing to 93.2+/-0.5%. A second semi-synthetic aggregate and controlled 8B and burst experiments separate the effects of affinity and placement. Two 32B workloads provide the counterexamples: when affinity recovers too little KV work, its residual load skew reduces or erases the improvement. We therefore recommend gating any deployment with a shadow replay rather than enabling affinity from workload statistics alone.

Choosing tensor-parallel (TP) degrees and replica counts for an LLM serving fleet is difficult because performance is not monotonic in TP and the feasible choice can change with load. Exhaustive profiling resolves this uncertainty, but measures many configurations that do not affect the final resource allocation. We present FleetSieve, which selects measurements according to their expected effect on a resource-coupled, SLO-aware fleet decision. FleetSieve models capacity and tail latency jointly, compares conservative and optimistic allocations, and stops when their remaining decision gap is below a specified tolerance. On a fixed H100 measurement grid for a 31B-parameter open-weight model, FleetSieve reaches the oracle aggregate decision using 22,200 GPU-seconds, 6.9% less than uniform random profiling in the fixed comparison. Across 200 random reveal orders, its mean saving over random profiling is 5.4% (95% bootstrap CI: 3.5-7.2%). The fixed-comparison saving is 21.5% for Chat, while FleetSieve does not use the fewest GPU-seconds for Code. Joint capacity and tail modeling also avoids selecting a configuration whose 46.4-second completion p99 violates a 30-second SLO. In a 16-GPU allocation, an incorrect sparse-profile decision loses up to 1.93 requests/s and 12.4 percentage points of max-min fulfillment. Boundary repeats and BurstGPT measurements support the observed load-dependent tail-latency mechanism.

Fast state-machine replication (SMR) protocols in the crash-fault setting have attracted significant interest in both academia and industry. This interest stems from their advantages over leader-based protocols, including low execution latency for non-conflicting commands, high throughput, improved availability, and increased fairness. While Fast SMR is well studied in the crash-fault setting, the Byzantine fault-tolerant (BFT) setting remains much less understood, with only a small number of existing Fast BFT SMR protocols. In this paper, we present tight upper and lower bounds on the replication factor required for Fast BFT SMR. We also present a suboptimal protocol that illustrates a trade-off between replication factor and recovery efficiency.

Autonomous vehicles offload latency-sensitive perception tasks to nearby mobile edge computing (MEC) servers, where a missed safety-critical task is unsafe rather than merely degraded. Large language models (LLMs) are increasingly proposed as adaptive, explainable schedulers, yet evidence of when they help is scarce. We study deadline-aware, mixed-criticality scheduling on heterogeneous MEC servers, where time-critical (TC) tasks must be protected at a controlled cost to best-effort traffic, and ask whether a multi-agent LLM control layer improves on a strong heuristic. We answer in two steps. First we build the heuristic: a windowed contract-net auction that orders each admission window time-critical-first by earliest deadline and places tasks by earliest-finish-time. Across 60 instances on three topologies and 15 baselines under an identical online constraint, it attains a TC completion rate of 0.902, above every baseline (Holm-corrected p < 0.001; best baseline 0.838) and at 0.87 of a CP-SAT upper bound. Second, we add the LLM control plane. A controlled decomposition traces the scheduler's advantage to two ordinary factors, the batching horizon and time-critical-first ordering; the auction, the per-window LLM policy, and online adaptation add nothing while the load is stationary, where the heuristic is already near-optimal. Under a mid-run surge of safety-critical tasks the picture changes, and the LLM control plane gains significantly over both the static heuristic and the bandit. LLM orchestration therefore earns its cost only when non-stationarity opens headroom a fixed policy cannot use. We report control-plane latency and rationale, and release all code and seeded instances.

Retrieval-augmented generation (RAG) improves language-model responses by grounding generation in external passages, which comes with overhead: retrieved context lengthens the prompt, increasing prefill work, KV-cache footprint, memory traffic, latency, and energy. Context compression offers a natural remedy by pruning retrieved text before generation. However, state-of-the-art context-compression methods are typically used with a fixed compression budget, or with the rate selected offline and then applied at inference time. This static view ignores both workload variation and the live state of the edge device. On an edge SoC, compression is not free: the compressor itself runs on the same SoC and consumes latency and energy that can offset any generation savings. This paper proposes a vision for telemetry-informed adaptive compression in edge RAG, grounded in experimental evidence. We characterize the compression tradeoff on the NVIDIA Jetson AGX Thor using Llama and Qwen generators, Natural Questions and HotpotQA datasets, and LLMLingua-2 compression. Our measurements show that generation dominates the RAG budget for larger models, reaching roughly 90% of per-query latency and 91% of GPU energy for 7B-8B generators. Exploring the impact of the compression rate reveals an adaptive operating region: mild compression can miss energy opportunities, and overly aggressive compression can hurt inference quality. Intermediate compression can reduce GPU energy by up to 53.2%, and SoC energy by up to 48.2%, with negligible quality loss. We argue for runtime policies that dynamically manage compression, guided by workload features and edge telemetry.

Lazy container-image pulling promises to eliminate the dominant cost of starting a model-serving pod by mounting the image immediately and fetching content on demand. We evaluate this promise for model delivery on Kubernetes, using KServe with two production lazy-pulling systems – eStargz/stargz-snapshotter and AWS SOCI – against eager baselines, on artifacts from 2 to 140 GB including real fp16 weights. Lazy pulling delivers its headline: cold time-to-first-prediction becomes size-independent (16.9–17.6s, versus 24.5–573.0s eager). But the cost is deferred, not eliminated: a full read of a 14 GB model through the lazy mount takes 105.3s, slower than the 72.4s eager pull it replaced, and the two systems pay at opposite lifecycle ends (SOCI prefetches nearly the full image before Ready; eStargz defers nearly everything to first read). More consequentially, we characterize a failure mode eager pulling structurally cannot exhibit: under sustained legitimate reads with default configuration, the snapshotter's node-level cache exhausts its finite volume and already-running pods begin failing reads of model files. At the earliest stage of exhaustion, an instrumented serving pod passed every Kubernetes-visible and application-level check for 196s while its snapshotter was already logging real failures; under heavier pressure, 67–94% of model files fail, scaling monotonically with residual cache occupancy. A live pod self-heals if cache space is freed under it, but a snapshotter-daemon restart under a live pod leaves permanently stale file handles in a pod still reported Running. We derive placement, monitoring, and cache-sizing guidance for serving platforms and operators.

Vizing's theorem states that any graph of maximum degree $\Delta$ can be properly edge-colored with $\Delta + 1$ colors (which is optimal in general). A recent breakthrough result by Bernshteyn showed that such a $(\Delta + 1)$-edge coloring can be found deterministically in $poly(\Delta,\log n)$ rounds in the LOCAL model of distributed computing, where $n$ denotes the number of vertices of the input graph [J. Comb. Theory 2022]. Since then, the exponent in the $poly(\log n)$-part of the runtime has been improved by Christiansen [STOC 2023] and Bernshteyn and Dhawan [J. Comb. Theory, Series B, 2025]. However, the algorithms used in all of these works use large messages, leaving open the question for efficient algorithms in the more restrictive CONGEST model. We answer this question by presenting the first $poly(\Delta,\log n)$-round algorithm for $(\Delta + 1)$-edge coloring in the CONGEST model. Our algorithm is deterministic and the $n$-dependency of its runtime, $\tilde{O}(\log^5 n)$, matches the best published dependency in the LOCAL model.

We prove matching upper and lower bounds on the space complexity of simulating a large shared register using smaller shared registers. We focus on the case where both the simulated and base registers are single-writer, which means they can be accessed concurrently by multiple readers but only by a single writer. To strengthen our lower bounds, we prove that they hold even when the base registers are atomic and the simulated register is regular. Furthermore, the lower bounds hold for obstruction-free implementations, which means they also hold for lock-free and wait-free implementations. If $m$ is the number of values representable by the large register and $b$ is the number of values representable by each base register, our first lower bound says that any obstruction-free implementation that has an invisible reader requires at least $\lceil \frac{m-1}{b-1} \rceil$ base registers. A reader is considered invisible if it never writes to base registers. This lower bound is asymptotically tight for the invisible-reader case and represents an exponential improvement over the previous best known lower bound. For the general case, which allows any combination of visible and invisible readers, we prove a $\lceil \min(\frac{m-1}{b-1}, r+\frac{\log{m}}{\log{b}}) \rceil$ space lower bound, where $r$ is the number of readers. To show that this lower bound is asymptotically tight, we develop a wait-free algorithm for simulating a multi-word atomic register from atomic base registers using $\Theta(r + \frac{\log{m}}{\log{b}})$ space. Combining this algorithm with known invisible-reader constructions gives a $\Theta(\min(\frac{m}{b}, r + \frac{\log{m}}{\log{b}}))$ space upper bound. This improves upon the previously known space upper bound of $\Theta(\min(\frac{m}{b}, r\frac{\log{m}}{\log{b}}))$.

Modern Intel AI PCs ship capable integrated GPUs and NPUs with 16+ GB of unified memory, and they spend considerable time idle. That is not enough memory to fit a large model such as a 70B-parameter LLM. We show that a handful of AIPCs, working together over an ordinary network, can serve models beyond the capability of any single one. We use pipeline parallelism: a model is split by layer into per-stage shards, each pre-compiled into an OpenVINO graph, so that every machine runs one shard and passes activations to the next. Three techniques make this fast enough to be useful. First, we recover the speed of the unsplit model: a naive per-stage export runs well below monolithic inference because it misses an OpenVINO GPU optimization, and injecting a beam_idx Gather into each shard triggers that optimization (the IndirectKVCache fusion) and brings the shards to parity. Second, we leverage speculative decoding on stateful OpenVINO models. Third, the pipeline serves several users at once by interleaving their requests across the stages, each request carrying its own cache (micro-batching). Together, a two-node Llama 3.1 8B INT4 pipeline serves two concurrent users at 1.79x the single-user throughput of the unsplit model on the same hardware, and the gap widens under simulated wide-area latency. The same design scales to a 70B model that no single fleet member can hold: a four-node deployment of Lunar Lake AI PCs on Intel Tiber Cloud serves a single user at interactive speed, with output token-for-token identical to the same four-node pipeline decoding without speculation. Code, raw benchmark logs, and reproduction scripts ship as a self-contained package at https://github.com/labscommunity/pipeline-sharded-inference-paper (in the top-level reproduction/ directory).

A connected vehicle has roughly 100 milliseconds to decide whether an incoming Basic Safety Message is real or fabricated. If a false emergency braking alert reaches the planning pipeline in time, the car brakes - a safety failure triggered by a security failure. Existing intrusion detection systems are not designed to handle that coupling. They operate per vehicle, per message, with static rules - blind to attack patterns that only emerge across a fleet or over time, and blind to the fundamental tension between dropping a suspicious message and dropping a real emergency alert. We propose a three-tier multi-agent architecture that treats this timing constraint as a hard design requirement, not a performance target. At the vehicle level, an onboard agent classifies each incoming V2X message into one of four actions - Accept, Drop, Quarantine, or Escalate - within a 10-millisecond budget, deliberately biased toward Escalate when uncertain, passing ambiguous cases to the roadside edge agent rather than risking a dropped legitimate alert. The edge agent operates across a roadside unit zone with a 50-millisecond budget, fusing threat assessments from multiple vehicles and resolving safety-security conflicts using complementary sensor observations. The cloud tier refines detection models through Byzantine fault-tolerant federated learning and redistributes updated weights to the fleet. Every timing constraint derives directly from the 100-millisecond Basic Safety Message cycles mandated by SAE J2735 and ETSI EN 302 637-2. No existing framework simultaneously assigns standards-grounded latency budgets to all three deployment tiers while treating safety-security conflict resolution as a first-class design constraint. Remaining open problems - adversarial poisoning at the edge and the absence of regulatory frameworks for autonomous security response - are discussed as future work.

We study the Minimum Dominating Set (MDS) problem in the sleeping CONGEST model (Chatterjee, Gmyr, and Pandurangan, PODC 2020), a generalization of the standard CONGEST model, in which a node may sleep in some rounds and can only compute, send messages, or receive messages when it is awake. The awake complexity of an algorithm in this model is the worst case number (over all inputs and all nodes) of rounds a node is awake for during the execution of the algorithm. While there are several $O(\log \Delta)$-approximation algorithms (in expectation) for MDS that run in $O(\log^2 \Delta)$ rounds, all of these have $\Omega(\log^2 \Delta)$ awake complexity. Whether this awake complexity can be improved is the question that drives our work. We present the first $O(\log \Delta)$-approximation algorithm for MDS with $o(\log^2 \Delta)$ awake complexity; our algorithm runs in $O(\log^2\Delta)$ rounds with $\tilde{O}(\log\Delta)$ awake complexity. We can reduce the awake complexity further, but at the cost of approximation: we present, for any $1<\alpha\le\Delta$, an algorithm in the sleeping CONGEST model that computes an $O(\alpha\log\Delta)$-approximate dominating set in expectation in $\tilde{O}(\log\Delta\cdot \log_{\alpha} \Delta)$ rounds with $\tilde{O}(\log_{\alpha} \Delta)$ awake complexity. Our results depend on a generalization of the CONGEST model SetCover algorithm of Grunau, Mitrovi'c, Rubinfeld, and Vakilian (SODA 2020) that we develop. This generalization computes an $O(p\cdot q\cdot\log_p\Delta)$-approximate dominating set in $O(\log_p\Delta\cdot\log_q\Delta)$ rounds for parameters $1<p,q\le\Delta$. Our sleeping CONGEST algorithms apply a variety of techniques including sampling-based estimation and scheduling using virtual binary trees to the aforementioned 2-parameter SetCover algorithm.