Papers

The default scheduler of Kubernetes, the state-of-the-art container orchestrator, uses fast, local placement decisions. Unfortunately, this design leads to resource fragmentation, reduced cluster usage, and overprovisioning. External solvers can compute global placement plans, but enforcing these plans in upstream clusters is hard. Kubernetes provides no native cross-node preemption, uncoordinated concurrent scheduling leads to inconsistencies, and replacing the default scheduler would sever deployments from upstream cycles. We present OPSche, an open-source Kubernetes Scheduling Framework plugin where external solvers can drive cluster-wide placement decisions in concert with the default scheduler. OPSche atomically validates and enforces solver-produced plans through coordinated framework hooks and supports three trigger modes: scheduling-failure, periodic, and stable-queue – resp. triggered when a workload cannot be placed, at fixed time intervals, when the set of pending workloads stabilises. Each mode has a blocking variant for a finer tuning of placement quality, latency, and disruption. We pair OPSche with a constraint-based optimisation solver, showing its feasibility across a broad set of cluster configurations and reporting improvements of resource usage by up to 3.0% and scheduling latency by more than a second.

Traditional complex event processing (CEP) systems focus on extracting information out of simple events in the input event streams to detect complex event patterns. In the context of CEP, a computing unit that executes a complex data transformation task is called an operator. Typically, a CEP system consists of an operator graph that could be distributed over the cloud-edge continuum. However, an operator graph distributed over heterogeneous nodes comes with its own set of challenges. In fact, the nodes of an operator graph are typically subjected to dynamically changing workloads. Unbalanced load distributions across heterogeneous nodes may lead to processing overloads in individual operators that do not have enough resources to properly adapt to an increasing workload. As a result, a framework that can detect and adapt to dynamically changing workload in real time across the various nodes of the cloud-edge continuum becomes paramount. To this end, the goal of this work is to propose a new extensible self-adaptive CEP framework with load regulatory features that help operators continue working even in face of an overload, thus facilitating the deployment of operators in a heterogeneous and/or resource-limited environment.

In gossip learning, a network of nodes trains a shared model collaboratively, without a central coordinator, by repeatedly exchanging parts of their local models. The state-of-the-art protocol, Partitioned Token Gossip Learning (PTGL) of Hegedüs et al., splits the weight matrix into S fixed partitions and disseminates them using a token-based fairness mechanism coupled with per-neighbor metadata exchange. We revisit partition scheduling by analogy with peer-to-peer live streaming, where model partitions act as video chunks and partition age acts as chunk scarcity. The analogy yields a design space of two-stage selection strategies (partition first, or neighbor first), from which we instantiate ten concrete protocols collectively called Stream Learning. Our main finding is that the simplest of these protocols, which transmits the locally least-trained partition to a uniformly random neighbor (Ri), matches PTGL on fault-free workloads while requiring neither token counters nor metadata exchange. Under an adversarial 30% permanent crash of the best-performing nodes, Ri matches or outperforms PTGL across all complete-graph configurations tested, with the gap reaching 5.53% on HAR and 5.41% on MNIST in the most heterogeneous regime (Dirichlet $\beta$ = 0.1). In our experiments, partition fairness, captured by a single local rule on partition age, accounts for the gap; token-based rate control and utility maximization do not improve over this rule and, under heterogeneity, sit below it.

In the era of Industrial Internet of Things (IIoT) and Cyber-Physical Systems (CPS), Federated Learning (FL) offers a promising decentralized intelligence paradigm for Video Anomaly Recognition (VAR). This task is vital for maintaining high-fidelity Digital Twins and ensuring safety in mission-critical environments. However, the inherent data heterogeneity across distributed edge clients leads to a fundamental challenge known as semantic misalignment, where clients learn divergent feature representations of "normal" and "abnormal" events. The problem becomes particularly pronounced in VAR, where the presence of diverse and fine-grained anomaly categories leads each client to develop distinct semantic interpretations of abnormality. Existing federated methods primarily focus on binary anomaly detection and fail to address this misalignment, preventing effective fine-grained recognition. In this paper, we introduce FedVAR, a weakly-supervised FL framework explicitly designed for VAR. Leveraging the rich representations of Vision-Language Models (VLMs), FedVAR employs a prototype-based alignment mechanism that creates a shared semantic anchor for all clients to re-center and align their visual and textual feature spaces. This process enforces a consistent representation of "normality" across the decentralized network, directly mitigating semantic misalignment and enabling robust prompt-learning of anomaly direction vectors with minimal communication overhead. We conduct extensive experiments on challenging benchmarks under various non-IID data partitioning schemes, unseen domains, and novel anomaly classes. The results demonstrate that FedVAR consistently outperforms state-of-the-art federated baselines, establishing a robust framework for distributed intelligence in video-based CPS.

Long-context training is increasingly important for large language models, and linear attention and state space models have become popular for improving long-context efficiency. However, efficiently parallelizing long-sequence training for recurrent and hybrid models remains challenging. We present StateFlow, a sequence pipeline parallelism system for models with linear recurrence. StateFlow partitions each sequence into chunks and schedules their execution while propagating boundary states and gradients across chunks, thereby reducing activation lifetimes and improving training throughput. StateFlow further uses profile-guided nonuniform chunking to balance recurrence and softmax attention computation in hybrid models, and overlaps state transitions that expose limited parallelism with surrounding computation. Applying StateFlow to models with up to 32B parameters and 256K context length, we achieve up to \(2.22\times\) throughput improvements and \(2.45\times\) memory reduction compared to conventional pipeline parallelism, enabling otherwise infeasible configurations.

This paper proposes a method and its implementation for emulating FP64 matrix multiplication (DGEMM) by constructing, on FP4 (E2M1; 2 exponent bits and 1 mantissa bit) Tensor Cores, Ozaki schemes I and II, which realize high-precision matrix multiplication on low-precision arithmetic units. Prior implementations were based on INT8 and FP8, and the use of the faster FP4 had not been realized. The key property is that every FP4 value becomes an integer when doubled, and that shifting this integer set by multiples of 13 covers all integers. Converting an arbitrary integer into base-13 FP4 limbs by this property keeps intermediate sums error-free in FP32 accumulators, which makes FP4 Tensor Cores usable for Ozaki schemes I and II. By the same principle, the integer GEMM of INT8 Tensor Cores can also be emulated bit-exactly on FP4 Tensor Cores. When FP4 Tensor Cores have twice the throughput of FP8, Ozaki scheme II on FP4 theoretically achieves slightly higher performance than its FP8 counterpart. This paper further proposes kernel implementation optimizations raising the attained fraction of peak performance, obtaining a measured speedup on top of the theoretical advantage. We verify this on an RTX PRO 6000 Blackwell, achieving performance competitive with that of an existing FP8-based implementation of Ozaki scheme II, and actually exceeding it at a large problem size ($16384^3$).

Weight quantization for large-language-model inference must balance adaptive reconstruction levels with representations regular enough for efficient GPU execution. Uniform integers constrain each group to a linear grid. Low-bit floating-point formats use a fixed exponent-mantissa structure, while learned codebooks gain flexibility at the cost of irregular decoding and additional metadata. We introduce CubicQuant, a parametric non-uniform scalar format that preserves a dense integer code stream while adapting reconstruction levels within each weight group. A monotonic cubic curve, specified by two shape parameters and one scale, maps uniformly spaced magnitude codes to non-uniform levels. The family spans 1-8-bit weight payloads, contains symmetric uniform integer quantization as an exact special case, and has effective width B + 64/G bits per weight for payload width B and group size G. We derive population distortion under Uniform, Gaussian, and Laplace distributions, formulate continuous and Dynamic-A8-carrier-aware fitting objectives, and describe direct packed-weight GPU execution. For finite groups of G=128 with 15,360 samples per distribution, W4 CubicQuant reduced reconstruction RMSE relative to optimally clipped four-bit uniform integer quantization by 3.90% on Uniform, 13.49% on Gaussian, and 28.14% on Laplace samples. Relative to the best enumerated four-bit finite floating-point format, the reductions were 3.90%, 9.44%, and 6.27%. Preliminary H200 kernel measurements show a workload-dependent crossover: model-dtype execution is faster for narrow GEMV, while Dynamic A8 becomes favorable as row count grows. The results establish the format's representational promise and direct executability; downstream model quality and cross-device end-to-end performance remain open evaluation questions.

Modern High Performance Computing systems depend on static heuristics and manual administration for job scheduling and reservation management. Deep Reinforcement Learning (DRL) has shown promising scheduling performance but requires historical training data and fixes the optimization goal at training time, forcing operators to retrain whenever priorities shift. We introduce MARS (Monte Carlo Tree Search-based Adaptive and Responsive Scheduler), a training-free HPC scheduler whose optimization goal is configurable through a reward function rather than baked into a learned model. MARS uses a lightweight discrete-event simulator to explore the future consequences of scheduling decisions within a strict time budget, adapting to the configured reward at each scheduling cycle. We evaluate MARS on year-long production workloads from two systems at Argonne Leadership Computing Facility – 4,360-node Theta and 560-node Polaris—under two reward functions: wait-time minimization (MARS-CW) and utilization maximization (MARS-CU). Unlike DRL and heuristics, which only react to the current queue or wait for backfill to find holes, MARS exploits look-ahead to proactively drain the system and plan around future reservations, packing the system to avoid the fragmentation and utilization drop that typically precede reservation windows. MARS-CW reduces tail wait time by 64% on Theta and 43% on Polaris over the production WFP heuristic, while MARS-CU recovers utilization in the 48 hours leading into maintenance, demonstrating that MARS can target either objective via reward reconfiguration.

The reasoning and agentic capabilities of large language models have expanded the range of applications they support, from short interactive exchanges to long, compute-heavy requests. LLM serving platforms today define response-latency service-level objectives, even though requests within the same service can differ by orders of magnitude in input length, generation length, execution cost, and the availability of reusable KV-cache state. As a result, requests governed by the same service level objective have different urgency: after accounting for the time required to execute them, some have substantial latency headroom while others have almost none. We define this headroom—the difference between a request's service level objective and its predicted remaining service time—as its per-request latency budget. We present Cascade, an LLM serving system that estimates and continuously updates this budget from request characteristics, KV-cache state, and current system load. Unlike prior SLO-aware schedulers that use deadlines to govern request ordering alone, Cascade uses a single per-request budget to jointly coordinate request scheduling and KV-cache management across the memory hierarchy. Its scheduler prioritizes requests with little remaining budget, while its memory manager uses the same budget to decide whether non-resident KV state should be restored or prefetched from a deeper tier, retained in HBM, or recomputed. By directing queueing and data-movement overhead toward requests that can absorb it, Cascade improves SLO-satisfied goodput while preserving fairness across heterogeneous request classes. On production traces across three large language models, Cascade improves goodput by up to2.4x and reduces SLO violations by 40% relative to the default vLLM first-come, first-served scheduler.

Two mobile agents, modeled as identical deterministic finite automata (DFA) navigating in synchronous rounds in a graph with unlabeled nodes, have to meet at some node. The well-researched task of meeting in a graph is known as rendezvous. Agents start at adversarially chosen distinct nodes in possibly different rounds. An instance of the rendezvous problem is the underlying graph, together with the initial nodes $u$ and $v$ of the agents. Such an instance is feasible, if there exists a DFA (possibly working only for this instance), such that its identical copies starting at nodes $u$ and $v$, with an arbitrary delay, accomplish rendezvous. A DFA is RV-universal for a class of instances, if it guarantees rendezvous of its copies starting at the designated nodes with arbitrary delay, for all feasible instances of this class. Our goal is to investigate the existence of RV-universal DFA. We start by observing that if agents cannot mark nodes in any way then there does not exist a RV-universal DFA even for the class of instances where the underlying graph is a line. Hence we allow the use of identical pebbles to mark the nodes by the agents. We consider stationary pebbles that can be dropped by agents at nodes but cannot be picked up, and movable pebbles that can be dropped by agents and later picked up. We observe that, even in the more powerful scenario of movable pebbles, if agents are equipped with any finite number of pebbles, there is no RV-universal DFA for the class of all instances. Hence we restrict attention to instances where the underlying graph is a tree. Our main contribution are two contrasting results showing that movability of pebbles is a crucial feature. We first prove that for any finite number of stationary pebbles there is no RV-universal DFA for trees, and then we design a RV-universal DFA for trees, where each agent is equipped with a single movable pebble.

Full-graph GNN training delivers high accuracy but scales poorly on multi-server clusters due to heavy, irregular inter-node embedding exchanges. We present SNI-GNN, a SmartNIC-assisted full-graph training system that reduces communication while preserving accuracy by predicting remote embeddings in-network. SNI-GNN deploys a lightweight linear-trend predictor on SmartNICs to refine cached historical embeddings, coupled with an importance-based boundary-node sampling policy and an asynchronous DPU–GPU data pipeline with intermediate-result reuse. We provide error and convergence bounds showing that predictor bias remains controlled under bounded second-order dynamics and yields standard non-convex convergence with inexact gradients. Implemented on NVIDIA BlueField-3, SNI-GNN integrates with state-of-the-art full-graph systems, cuts communication by 21–45%, achieves 1.3–3.6$\times$ end-to-end speedups over BNS-GCN and up to 1.29$\times$ over baseline SANCUS, with accuracy loss $\leq 0.01$, and scales efficiently to 16 GPUs on graphs with up to tens of millions of edges. These results indicate SmartNIC-based in-network prediction is a practical complement to partitioning and compression techniques for communication-efficient full-graph GNN training at scale.

Large-language-model inference is a fast-growing electricity load whose marginal carbon intensity varies by more than an order of magnitude across grid regions and across the day, making request placement an attractive lever: no retraining, no hardware change. We report a live validation of carbon-aware inference routing on multi-region GPU testbeds driven by marginal operating emissions rate (MOER) signals, with three properties uncommon in prior work: a blind baseline that is an actual production pressure-based router rather than uniform placement; per-request energy attributed from GPU telemetry (NVIDIA DCGM) via measured concurrency curves rather than nameplate TDP; and carbon settlement of every request against historical MOER, not only the forecast that drove the decision. The central live result is feasibility: a MOER signal steered inference across regions with no observed dispatch failures, as a strict and reversible overlay on the production router. To size the effect, we replay a year of hourly MOER across a grid-diverse CONUS fleet. In the primary modeled configuration, carbon-aware placement reduces modeled GPU-attributable operational emissions by 50.9% versus round-robin (95% block-bootstrap CI 48.5-53.3%). Because the replay dispatches against historical MOER rather than a forecast, this is an upper bound under that configuration; forecast error would reduce operationally realized savings. Before session pinning, hourly lowest-MOER routing contributes about 22.4 percentage points, roughly 40% of the 54.0% placement reduction, beyond a static annual-mean-MOER policy. These are modeled results for one fleet and historical year, not a universal estimate. We also record a practical observation: when comparing regions, rank by absolute MOER rather than the percentile signal-index, which is normalized within each region and answers a temporal, not a spatial, question.

Priority-differentiated services are a standard way for applications to offer different levels of performance, but database systems still often treat all sessions the same way. When database capacity is fixed, meaning replicas cannot be added on demand, and the workload becomes contended, this creates a difficult trade-off between over-provisioning resources and letting lower-priority users experience much larger slowdowns. In such settings, we propose enforcing priority by controlling how client sessions are assigned to database replicas. We present PLB, a priority-aware load balancer implemented as a JDBC driver that enforces priority differentiation through replica assignment under fixed resources. PLB partitions replicas by user group, premium versus freemium, and uses load-based borrowing so that higher-priority users can use idle capacity when available, while degradation for lower-priority users remains controlled. We evaluate PLB on a replicated read-only cluster under OLAP workloads. Compared with static dedicated per-priority partitions, PLB keeps utilization above about 70% in settings where fixed partitions can reduce cluster-wide CPU utilization to about 35%, while maintaining latencies close to those of the best dedicated allocation. Compared with a fully shared round-robin pool, PLB lowers high-priority median latency by about 12% on average, with improvements of up to 28%, while keeping the low-priority median overhead around 11% and overall median latency close to round-robin.

AI training workloads are growing rapidly, making their time, energy, and infrastructure costs increasingly important. In shared cloud clusters, training and fine-tuning jobs compete with co-running workloads for network resources, while network mechanisms and ML training choices are typically optimized separately: networking controls how bytes move, whereas ML systems control when and how much communication occurs. We argue that this separation leaves end-to-end performance on the table. We present ML-for-ML, a cross-layer perspective in which network-side and ML-side knobs are selected jointly under a shared time-to-target-loss objective. Our preliminary prototype shows that by co-optimizing the ML and network parameters, we reach the target loss up to 42% faster.

Modern LLM infrastructure increasingly manages tensors not only as computation data, but also as persistent states shared across distributed components. Existing systems optimize individual tensor management tasks, such as model weight loading, KV cache management, and checkpoint synchronization, by deeply integrating task-specific mechanisms with execution engines, networks, or storage backends. However, this specialization creates isolated silos that hinder the reuse and composition of tensor management strategies across evolving LLM workloads. In this paper, we identify tensor lifecycle management as a missing abstraction layer in LLM infrastructure and propose Tensor-as-a-Service (TaaS), which decouples tensor state management from computation logic. We design and build TensorCast, a distributed tensor management layer that provides first-class tensor abstractions, programmable lifecycle primitives, and a runtime that separates tensor management policies from execution mechanisms. This enables developers to write tensor management programs using TensorCast APIs while transparently leveraging distributed execution and data movement. We integrate TensorCast with vLLM and SGLang and evaluate it across diverse tensor lifecycle workloads, including model weight materialization, weight synchronization, KV cache management, and programmable request routing. Our results show that TensorCast achieves competitive performance with specialized tensor management systems while enabling new cross-component optimization policies. A programmable policy implemented with TensorCast improves median TTFT by up to 93.2% under highly concurrent multi-turn agent workloads.

We report operational experience full-fine-tuning a 32.76B-parameter dense model (Qwen3-32B) on 16 x NVIDIA B300 (two nodes, FSDP / ZeRO-3) – among the first published field accounts on this accelerator. We claim no new algorithm. The individual mechanisms we use are established practice; our contribution is the integrated field experience and a set of calibrated measurements on new hardware. Concretely we offer four practitioner artifacts. (1) A B300-calibrated power-draw triage table that distinguishes compute / communication / data-starvation / checkpoint-or-deadlock / idle by board wattage (utilization% reads 100% during an NCCL hang). (2) A set of honest negative results that dispel common optimization folklore at this scale: a controlled A/B in which per-step NFS reading matches a pretokenized local cache (~53k tok/s) because the corpus fits in page cache and the job is compute-bound; and a reconstruction of an earlier "throughput collapse" as NFS/CPU contention rather than a storage-medium limit. (3) Calibrated 4/8/16-GPU strong-scaling and GPU-hour numbers on B300 (near-linear, as expected in this regime; we report absolute values as reference data). (4) A worked failure case – an epoch-end NCCL deadlock from per-rank token-packing imbalance – together with a 2.7-second pre-run invariant gate and an external watcher that turn multi-hour silent failures into instant rejections. This deadlock and its remedy correspond to PyTorch's documented Join / equalize-to-minimum practice; we position our instantiation against that prior art and report the GPU-hours the failure cost and the gate saves. The transferable takeaway is operational, not algorithmic: for data-dependent data-parallel jobs, watch power rather than utilization, and verify invariants before launch – a passing smoke test is not evidence of a safe full run.

OpenMP parallelization of hotspots in mature repositories remains difficult because loop safety and optimization payoff often depend on non-local evidence. Rule-based tools under-parallelize when legality is not locally provable, while agent-based approaches become unstable when retrieval misses decisive dependencies or includes irrelevant code. We present RepoOMP, a hybrid framework that recovers parallelization-relevant evidence before generation. RepoOMP builds a Multi-granularity Attributes Performance graph (MAP), routes hotspots between deterministic rules and an LLM agent, and constructs a Structured Transformation Context (STC) that exposes dependency facts without flooding the model with unrelated repository text. We evaluate RepoOMP on 951 profiled hotspots from NPB, BOTS, FFmpeg, NCNN, and GROMACS. Under compilation, workload-specific checks, and positive speedup, 372 hotspots are accepted, including 330 real-world repository hotspots. RepoOMP achieves average speedups of $8.23\times$ on NPB and $8.96\times$ on BOTS. For the nine detailed real-world kernels used in matched-backbone and robustness analyses, RepoOMP reaches a cross-backbone mean of $5.25\times$, improves speedup by 18–28%, and reduces agent-side token cost by 47–68% relative to the unstructured Claude Code baseline. Across 330 accepted real-world hotspots, median speedup is $2.25\times$. Overall, RepoOMP provides an evidence-guided workflow for hotspot parallelization in repository settings. The open-source repository is available at https://github.com/Qlalq/RepoOMP_Simplified.

Classical simulation remains essential for developing and validating quantum algorithms, but its cost grows rapidly with circuit size. Tensor-network contraction can reduce this cost by exploiting circuit structure, although its efficiency depends strongly on the chosen contraction plan. On GPUs, plans with similar theoretical complexity may perform very differently because execution also depends on parallelism, reduction structure, memory traffic, and contraction geometry. We present a learning-to-rank framework for selecting efficient contraction plans before executing them. Each plan is represented by structural features derived directly from its sequence of pairwise contractions, and gradient-boosted rankers are trained from GPU measurements using listwise and pairwise objectives. We evaluate the resulting models on diverse circuit families, using separate in-distribution and circuit-family-shift test sets, and compare them with random and MinFill-based baselines. The learned rankers generally identify better plans, with the listwise model providing the strongest overall decision quality. We also study backend shift by comparing empirical plan orderings on two GPU architectures and evaluating the source-trained models on the second device without retraining. The rankings remain substantially, though not perfectly, stable across GPUs, and the models retain useful decision quality. These results support Learning to Rank as a practical way to reduce contraction-plan search, while showing that performance remains partly backend dependent.

A linear computation broadcast (LCBC) problem comprises $K$ users (receivers) and a transmitter. The users wish to compute various (vector) linear functions of a common dataset, and possess in advance heterogeneous side information corresponding to various other linear computations. The goal for the broadcast transmitter, who knows the dataset and all desired and side-information functions, is to satisfy all demands as efficiently as possible. Prior work has explored the information-theoretic capacity of LCBC for zero-error finite-field computation over an ideal (noiseless) broadcast channel (BC). This paper develops a wireless LCBC (WLCBC) framework for the Gaussian MIMO BC with noisy receiver-side information under a transmit power constraint and arbitrary antenna configurations. In the WLCBC setting, a Gaussian source is linearly precoded for broadcast; each user requests a linear function of the source and forms an estimate by linearly combining its channel observation with its noisy side information. Assuming perfect channel state information, we cast the centralized joint linear transceiver design as a weighted sum-MSE minimization problem and propose an efficient alternating optimization algorithm. For a fixed precoder, the optimal decoders are the linear minimum mean-square error (LMMSE) estimators. For fixed decoders, the precoder update reduces to a convex quadratically constrained quadratic program which leads to a semi-closed-form solution parameterized by a single dual variable. The resulting algorithm guarantees a monotonic decrease in the objective and convergence of the weighted sum-MSE (WSMSE) objective sequence. Simulations demonstrate pronounced robustness gains over natural baselines obtained from prior works.

Serverless platforms maintain a global view of function invocations and resource utilization, yet existing systems largely restrict CPU scheduling decisions to the operating system scheduler. This paper presents a serverless platform-driven CPU load balancing framework that enables the control plane to directly influence CPU scheduling through a custom Linux scheduler built on SchedExt(SCX). The proposed scheduler introduces configurable scheduling domains and a shared interface that allows the control plane to assign functions to domains based on their historical inter-arrival times. Within each domain, a single-queue load-balancing strategy combined with a virtual-time prioritization policy improves task placement while reducing interference from busy-polling tasks. Results show that an eight-domain configuration achieves the best trade-off, reducing system energy consumption by approximately 15% while increasing invocation cost by only 5%. Under heavily loaded workloads, the proposed scheduler also reduces request latency by up to 50% compared to the default Linux Completely Fair Scheduler (CFS). These results demonstrate that exposing CPU scheduling decisions to the serverless control plane can improve both energy efficiency and workload performance while preserving scheduling flexibility.

Circuit cutting lets a large quantum neural network (QNN) run as independent subcircuits on small devices, but rebuilding its outputs by reconstruction carries a classical sampling overhead exponential in the number of cuts - the dominant runtime cost in prior work. We ask whether, for machine-learning tasks, this step is necessary, and replace it with late fusion: each subcircuit is trained and measured independently, and a small classical head combines their outputs - a linear-cost, decision-level combination borrowed from multimodal learning. To characterize the trade-off we introduce a quantumness dial $Q$, a tunable reconstruction budget interpolating from pure fusion to full reconstruction, and a cut-entanglement diagnostic that indicates how much reconstruction a task needs (Spearman $\rho=0.59$ over $104$ runs). Across synthetic and standard datasets, independently trained late fusion matches full reconstruction accuracy within $0.04$ at every point of the controlled sweep and on every classical benchmark, at exponentially lower cost; it is also markedly more robust to shot and device noise. Controlled entangled-data experiments locate the boundary where fusion must fail. We do not claim advantage over classical machine learning - consistent with recent benchmarking, quantum offers no accuracy edge on these datasets. Late fusion is thus an efficient, noise-robust, self-characterizing alternative to reconstruction for circuit-cutting QML.

The proliferation of multi-sensor Internet of Things (IoT) systems, from Body Sensor Networks (BSNs) to industrial monitoring, is increasingly constrained by strict energy budgets and limited on-device storage. Continuous high-fidelity sensing leads to rapid battery depletion and data gaps that compromise application reliability. Existing strategies address this through sensor selection or adaptive sampling in isolation, or rely on computationally expensive agents for joint optimization. They lack context granularity or introduce significant overhead, and critically, they do not account for the risk that an aggressive, context-specific sensing policy applied to a misidentified context degrades accuracy. In this paper, we formulate joint sensor and sampling-rate selection as an NP-hard energy-minimization problem and propose Viveka, a lightweight, context-aware framework. Viveka couples a cheap, always-on controller that estimates context and how much to trust that estimate with a stability and confidence gated policy that applies an aggressive per-context configuration only when context is certain, and falls back safely otherwise. Per-context configurations are instantiated using permutation feature importance and spectral energy analysis. Evaluation on the MHEALTH and PAMAP2 datasets shows that Viveka achieves up to 75% energy savings and 78% data reduction over standard baselines in a best-case configuration, while maintaining classification accuracy within 3-5% of the baselines.

Multi-LoRA serving is how one base model becomes thousands of specialized variants, one adapter per user, task, or agent, and the deployments can hold 1000-plus adapters. Serving them is hard because the workload inverts what GPUs provide: terabytes of memory against only tens of TFLOPS, and because every published system stages its adapters from CPU DRAM over PCIe, where each access pays a kernel stop and a host-run copy and capacity ends at the motherboard's DIMM slots. Meanwhile, memory-semantic fabrics such as CXL and NVLink are converging on pooled memory that an accelerator addresses with its own loads and stores, and near-data processing (NDP) can place compute beside the pooled data. How to serve multi-LoRA workloads on such hardware remains unexplored. This paper introduces PLoRA, an NDP-enhanced pooled-memory system for cost-efficient multi-LoRA serving. PLoRA keeps adapters and KV cache in the pool and returns only reduced results over the link, through a read-compute interface the GPU drives with its own loads and stores. Above this architecture, a GPU memory management system picks among four LoRA and two attention execution strategies for each adapter and caches the most performance-critical bytes in GPU memory, guided by a link-parameterized cost model. On one H100 serving 1000 adapters, PLoRA attains the lowest decode latency on every model and workload we measure, averaging 6.6x below a real-machine S-LoRA at under 3.4% added device area. The link itself stops mattering: throughput saturates at 32 GB/s on short contexts, a quarter of CXL 3.1, and the verdict survives scale: per-GPU demand falls from 7B to a modeled 1.2T deployment once adapter traffic shards with the tensor parallelism. The design runs unchanged from CXL-class to NVLink-class fabrics, and surplus bandwidth buys pooled capacity rather than speed.

Approximate nearest-neighbor (ANN) search increasingly runs alongside structured data - "find the 10 nearest documents where tenant='acme' AND lang='en'" - yet similarity and filtering are usually bolted together: a specialized vector index for one, a separate filter step for the other. We ask what happens when both live inside an open lakehouse table (Apache Iceberg over Parquet on object storage), where the engine already owns a mature file-pruning stack (partition pruning, zone-maps, a bitmap index). We embed an IVF index in place in each Parquet file's footer and make filtered vector queries fast not with a new filtering algorithm but by composing the table's existing file pruning with per-file ANN: the planner prunes data files by the predicate first, then runs IVF only over the survivors. The index is built distributed and non-destructively - a metadata-only Iceberg replace that every other engine still reads - and a rendezvous-hashed per-file cache keeps object-store read latency from swamping the algorithmic win. The payoff comes entirely from file pruning. On an 11.5M x 768 table, warm IVF search is ~32x faster than brute force at recall@10 >= 0.90, a selective predicate having pruned 355 of 444 data files before ANN runs; on 5M real IBM Granite embeddings, a filter arriving across a join prunes four of five region partitions and runs nearly two orders of magnitude (~94x: 14.7 s -> 157 ms) faster than the query-time join at identical top-k, once the reduction is materialized into a region-partitioned layout. We characterize when the composition pays off - it requires file-level locality on the filter column, and the residual predicate is only safe to push into the search over a provably pure (partitioned) column, not a merely sorted one - and report the failure modes we hit bolting ANN onto a lakehouse engine.

We present a framework for evaluating uptime efficiency of Exascale-class scientific computers when application failure rates are appreciable. This is the situation that confronts current leadership-class scientific computing platforms and large AI training installations. What distinguishes scientific computing platforms is the heterogeneity of their applications. We argue that this diversity requires that failure rates and mean intervals between failures should be specified in terms of usage (e.g. node-hours) rather than time, as is currently customary. We consider the usage loss terms due to failures, to checkpointing, and to restart costs, and update the framework of Daly (2006) allowing users to specify optimal checkpointing usage intervals that minimize such losses. We derive the machine computational efficiency, which specifies the expected fractional resource allocation that is available for scientific computation. We illustrate the methodology using one year of production runtime data from the Frontier supercomputer at Oak Ridge National Laboratory.

Federated learning repeatedly incurs local optimization and model-update transmission. We study DG-FedReuse, a simulator-level mechanism that allows selected clients to contribute age-decayed cached updates when a stochastic head-gradient discrepancy proxy remains below a round-dependent threshold. A hard cache-age limit and minimum fresh-client quota constrain reuse, while fresh updates use an adaptive per-tensor Top-K numerical-field representation. Experiments cover six image-classification datasets, 50 virtual clients, Dirichlet label heterogeneity ({\alpha}=0.5), and three seeds. At a common 90-round budget, DG-FedReuse yields 83.36-85.42% modeled update-data-field uplink saving, compared with 76.88% for matched Top-K FedAvg; the seed-aligned accuracy differences range from -5.29 to -0.14 percentage points. Best-observed test accuracies obtained under test-controlled checkpointing are retained only as exploratory archival evidence and range from -2.38 to +0.45 percentage points relative to matched FedAvg. A symmetric dense-model-downlink sensitivity reduces the headline saving to 41.68-F42.71% and the incremental gain over Top-K FedAvg to 3.24-4.27 percentage points, demonstrating the dependence of communication conclusions on the accounting boundary. The study characterizes the proposed reuse rule in the implemented simulator; it does not establish unbiased generalization, end-to-end bandwidth reduction, runtime or energy savings, faster convergence, or superiority over existing stale-update and lazy-aggregation methods.

Agentic science is transforming the landscape of computational work, and is applied to scientific pipelines and workload managers. Scientific workloads require specialized hardware within and between institutions. Automated resource discovery is an essential step for scheduling workloads with specific hardware and environmental requirements. In this paper, we present a hierarchical, dynamic architecture and accompanying software to discover resources across diverse cloud, edge, and HPC systems. The design enables concurrent, asynchronous negotiation, selection, and dispatch of requests for work using secretary agents. The agents probe and discover 51 real and simulated providers across 7 categories. We perform 19,973 negotiation and 6,952 selection simulations to assess reliability of decisions, demonstrating high (87.71%) negotiation accuracy and selection costs comparable to more traditional strategies. Designed for extensibility and currently supporting the US DOE Genesis Mission, this architecture exemplifies the importance of careful coordination between agents, discovery tools, and infrastructure for agentic science.

Modern cloud applications often comprise thousands of microservices whose interactions form complex request paths. Traditional inter-service access control restricts individual service-to-service requests, but fails to prevent multi-hop attacks, where each hop appears legitimate yet the overall path violates security intent. This gap leaves systems exposed to unauthorized access and data exfiltration. In this paper, we present eMicro, a path-aware defense system for microservices that prevents such attacks while remaining efficient and deployable. eMicro enforces real-time multi-hop access control through three key techniques: (1) history-based access control extended to capture service invocation sequences; (2) security policies encoded as efficient deterministic finite automaton (DFA), supporting constant-time lookups and compact label propagation; (3) eBPF-based in-kernel request tracing for transparent, low-overhead enforcement without code changes. Evaluations on DeathStarBench and production cloud traces from Uber, Alibaba, and ByteDance, covering 12 million request workflows and thousands of services, demonstrate the scalability of eMicro. eMicro performs policy checks in 1 microsecond, stores 50 million policies in only 100 MB, and reduces propagation overhead by 90% with negligible runtime impact. These results show that eMicro delivers scalable and efficient protection against multi-hop attacks, making it practical for deployment in large-scale microservice environments.

Sparse matrix computation performance on GPU depends on how representation and execution schedule match the input structure and target hardware. No single implementation consistently dominates across sparsity patterns, operators, and hardwares. Existing sparse compilers and specialized systems cannot cover all of them simultaneously. We present SparseDitto, an agentic sparse compilation framework for sparse matrix computation on GPUs. It jointly synthesizes representation, execution schedule, and hardware mapping in a unified compilation plan. Structural analysis and a learned template-ranking prior guide architecture-aware synthesis. LLM-guided lowering realizes each plan as CUDA code, while target-GPU profiling drives plan refinement. SparseDitto covers multiple operators, e.g., SpMV, SpMM, and SpGEMM, and various representations within one framework. It can also automatically adapt to different hardwares. Across various SuiteSparse matrices, SparseDitto achieves geometric-mean speedups over cuSPARSE of $2.68\times$ on an NVIDIA RTX PRO 6000 and $2.79\times$ on an NVIDIA H200 (up to 146.61$\times$). Its generated SpMM kernels accelerate full-batch GCN training by up to $3.39\times$.

Large language model (LLM) agents repeatedly process long, privacy-sensitive contexts, while cloud-only deployment exposes user data beyond the trusted endpoint and fully local deployment often requires costly hardware. Split inference offers a middle ground by executing the model head, tail, and tools locally and the middle layers in the cloud, but its local-cloud-local path transfers boundary hidden states at every invocation and creates a critical communication bottleneck. We present \system, a reference-aware codec that retrieves exact-token historical spans for prefill uplinks, reuses the reconstructed uplink state for same-round prefill downlinks, and generates boundary-specific decode references with lightweight causal predictors. RAC applies grouped affine alignment and calibrated residual quantization with optional prefill outliers, while sender-side wire-format reconstruction synchronizes subsequent references and offline calibration accounts for quality and packed representation costs. Across three models and nine evaluated model-link pairs, Raw-to-RAC mean time to first token (TTFT) and time per output token (TPOT) ratios are 1.24-2.72$\times$ and 1.01-2.79$\times$, while the 12 non-perplexity task-score changes range from $-0.40$ to $+2.50$ points.