Papers

We consider a team of synchronous mobile agents operating in a port-labeled network. There is one node in the network, called a black hole, that permanently destroys any agent that visits the node. The team of agents must safely locate the black hole, i.e., at least one agent must survive, terminate its algorithm at a node adjacent to the black hole, and output the port number that leads to the black hole from its current position. In the setting where the network is a 1-bounded 1-interval connected dynamic graph, Kaur et al. (SSS 2025) showed that a team consisting of $2\delta_{BH}+17$ agents is sufficient to solve the task from a scattered configuration, where $\delta_{BH}$ denotes the degree of the black hole node. We show that $2\delta_{BH}+3$ agents are sufficient, nearly matching the $2\delta_{BH}+1$ lower bound provided in Kaur et al. (ICDCN 2025).

The deployment of Hierarchical Federated Learning (HFL) in resource-constrained Internet of Things (IoT) environments requires careful configuration to balance predictive performance with energy consumption and execution time. This challenge is particularly relevant to smart agriculture, where distributed IoT devices can support automated plant disease classification while operating under limited computational and communication resources. This paper presents a constrained Bayesian Optimization framework for the efficient configuration of HFL deployments. The proposed approach jointly explores the deep learning backbone architecture, aggregation strategy, and number of communication rounds, while the federation size is determined according to the spatial coverage requirements of the agricultural deployment. A weighted objective function captures user-defined trade-offs among energy consumption, execution time, and predictive performance, while explicit constraints ensure compliance with deployment-specific resource and accuracy requirements. The framework is evaluated on an IoT-based plant disease classification task considering multiple deep learning architectures, federated aggregation strategies, and communication-round settings. Experimental results across 30 independent optimization runs show that the proposed approach explores only 11.11% of the search space, while consistently identifying solutions within 1% of the exhaustive-search optimum, with a mean optimality gap of only 0.056%.

Industrial and scientific computing rests on a few core kernels, and the stencil is among the most widely used: weather and climate models, seismic imaging, fluid dynamics, and image processing all run on it. No single stencil implementation is fastest: the optimal kernel changes qualitatively with stencil shape, grid shape, precision, and host application. For two decades the field has answered with general methods (DSLs, code generators, autotuners), because specialized solutions were too expensive to build per case, so all reuse one human-authored recipe. That reuse costs performance; we call the cost the generality tax. This premise no longer holds: code-synthesis agents now build a correct, specialized solution per case at acceptable cost. ForgeStencil automates this. A Kernel Agent synthesizes CUDA and forges a per-configuration map of specialized operators, removing the tax case by case. On an A100 the map beats the strongest public baseline in 37 of 37 cases: geometric mean 2.35x against same-precision f32 baselines and 1.95x for fp16, each reported under its own precision. The same change reaches end-to-end application performance. A generic operator library is tuned once for its own general case and reused across applications, so its shapes, layouts, and launch boundaries are optimal for none of them: using it is the application-level form of the tax. An App Agent instead forges a specialized solution per application, locating hotspots, rewriting application structure, and validating and integrating each change. Across 100 real industrial and scientific codes the end-to-end median speedup is 1.41x against each application's own GPU baseline. To our knowledge this is the first demonstration that per-case synthesis carries from a kernel library to complete applications at this breadth, and evidence that reuse is no longer the default in a domain built on it for two decades.

Latency-critical cloud services place growing pressure on memory while requiring isolation, fairness, and predictable QoS. Near-data processing (NDP) reduces data movement by executing requests close to memory, and prior systems further improve locality through caching and replication. However, writes make replica maintenance expensive, while uneven compute and memory traffic can overload a few NDP units and increase tail latency. Existing throughput-oriented schedulers do not fully address these challenges for co-located cloud services. We present Gutenberg, a CPU+NDP for mutable, latency-critical cloud services. Gutenberg stages subpage updates in a CPU-resident delta buffer, allowing hot writable pages to remain replicated without eager full-page synchronization. It also adopts CPU helper cores to assist request execution when NDP execution or replica maintenance becomes costly. An online controller jointly decides page placement, replication, CPU/NDP execution, and routing using access patterns, queue pressure, and feedback from prior decisions. The system further enforces isolation and fair resource allocation across services. We also model-check CPU–NDP coordination protocol for correctness. We evaluate on TailBench using ZSim with Ramulator-calibrated memory timing. Across evaluated services, Gutenberg outperforms prior systems, reducing average and p99 latency by up to 80.4% and 85.8%. It also improves isolation and fairness while adapting to changing workload behaviors.

FBPIC (Fourier-Bessel particle-in-cell) is a high-performance simulation code for relativistic plasma and accelerator physics. Its original accelerator backend relies on Numba CUDA, which limits its direct deployment on accelerators using the HIP (Heterogeneous-Compute Interface for Portability) programming environment, such as DCU (Deep Computing Unit) accelerators. In this work, we develop an accelerator backend compatible with HIP that enables FBPIC to run efficiently on DCU platforms while preserving its Python user interface and high level simulation workflow. For the evaluated LWFA (laser-wakefield acceleration) workloads, the proposed backend achieves 1.32-1.54x speedups over the original FBPIC implementation on an NVIDIA V100 GPU and enables efficient execution on the DCU platform. We also summarize the key lessons learned from porting FBPIC to the DCU platform. Multi-DCU experiments achieve a 1.88x strong-scaling speedup on four accelerators and a 2.72x increase in aggregate throughput at approximately 68% weak-scaling efficiency, with communication analysis identifying inter-node communication and synchronization as the main scalability limitations. Beyond FBPIC, the proposed approach provides a practical reference for porting and optimizing other scientific computing applications developed with Python on heterogeneous accelerator platforms.

Mobile vendors and application developers increasingly deploy LLMs on smartphones for diverse prefill-only services. Yet current systems rely mainly on dense models whose regular computation maps efficiently to mobile NPUs, leaving more capable MoEs underused. MoE prefill does not fit mobile NPUs: NPU graphs are fixed at compile time, yet MoE picks experts at runtime; and one request touches most experts, more than a phone can hold in memory. We present EStream, which resolves both by separating what the NPU must fix from what MoE decides at runtime. A single compiled expert graph serves every expert, with each expert's routed tokens and weight address bound at call time, so dynamic MoE execution runs entirely on the NPU without padding or CPU/GPU fallback. Expert virtualization keeps the expert pool in UFS flash storage and pages it through a fixed-size NPU-addressable arena, group by group, with loading hidden behind computation, so memory is bounded by the arena rather than by the model. It further introduces a hardware-aware configuration algorithm that automatically configures the UFS–NPU pipeline and maximizes loading–computation overlap. Across 18 comparative settings covering three 7B–16B MoEs and 256–4,096-token prompts, we evaluate EStream on a commercial Snapdragon smartphone. Compared to the fastest baseline at each setting, EStream achieves a 2.25–27.57X pure-prefill TTFT speedup and reduces peak physical memory by 1.19–12.29X. EStream further scales to MoE models with up to 46.7B parameters.

Distributed DL training on HPC systems often shares one network fabric between NCCL/RCCL collective communication and parallel-filesystem I/O. Using a real GNN training workload on a Slingshot-11 system, we show that this sharing imposes two distinct costs. The primary cost is heavy-tailed DataLoader stalls: the typical DataLoader wait is just 15 ms at steady state, yet spikes to multiple seconds in 28% of Lustre iterations and 12% of VAST iterations. The secondary cost is traffic-class contention on collective communication: Lustre I/O stalls the all-reduce by up to 145$\times$ in an isolated benchmark. The two costs arise from different mechanisms. I/O stall latency affects any storage path that traverses the shared fabric, whereas all-reduce network contention occurs only when storage and collective communication share the same traffic class. Their common root cause is that storage I/O traverses the shared fabric. This work shows that node-local NVMe staging via DYAD (Our code is publicly available at https://github.com/flux-framework/dyad) eliminates both effects by keeping storage I/O off that path. Across a full training epoch, DYAD achieves a 7.4 times speedup over direct Lustre reads and a 1.06 times speedup over VAST. By the second epoch, once the local cache is fully warmed, DataLoader stalls are eliminated entirely, allowing DYAD to reach a 1.31 times speedup over VAST.

Fast Fourier transforms on parallel machines incur two forms of data movement: processor-local transfers through the memory hierarchy and global redistribution between processors. Four-step FFT organizations switch the active transform dimension with one global transpose-like exchange, but this alone does not yield cache-efficient local computation. Cache-oblivious FFTs achieve asymptotically optimal local memory traffic through recursive layout transformations, whose direct materialization can require extra data-rearrangement passes and global exchanges. We present PACO, a fully cache-oblivious parallel FFT framework that reconciles these objectives. PACO executes LocalFFT -> OneGlobalPermutation -> LocalFFT. Its local stages recursively partition both transform and batch dimensions without knowledge of the cache parameters. Rather than materializing the transpose-like layouts induced by this recursion, PACO defers them, showing that they compose into a base-b digit-reversal permutation that is fused with the redistribution already required to change the local transform dimension. The resulting middle stage is a perfectly balanced parallel cache-oblivious digit-reversal permutation in which each source-destination processor pair exchanges exactly $N/p^2$ elements. Under an exact base-b slab decomposition in a hybrid ideal-cache/BSP model, PACO computes an N-point DFT exactly with maximum per-processor work $\Theta((N/p) \log N)$ and maximum per-processor cache complexity $\Theta((N/(pB))(1 + \log_M N))$, using exactly one global redistribution round. PACO returns canonical logical DFT coefficients under a factor-swapped target slab ownership. The single redistribution is necessary under the stated no-replication ownership model, while its communication volume is optimal for the prescribed fused permutation and source-target slab distributions.

Randomized benchmarking can hide classical temporal correlations because its Clifford-twirled response is even in the noise phase. For a stationary symmetric telegraph fluctuator, we show that continuous evolution and independent stationary resets at slot boundaries yield identical mean responses for arbitrary fixed idle modulations. We construct an eight-setting phase-cycle measurement of the connected sine-phase covariance under ideal Clifford twirling and classical idle dephasing. This observable vanishes for independent slot noise and fixed detuning without a weak-phase or Gaussian approximation. A closed telegraph response, independent circuit calculations and 800 simulation trials validate the construction and quantify the empirical coverage of a paired bootstrap estimator. A separate conservative confidence set states its finite-sample assumptions. Two acquisitions on an IBM processor compare engineered shared-sign and independently reset phases with identical marginals. Their primary contrasts are 0.254 and 0.211, with empirical 95% intervals [0.177, 0.331] and [0.136, 0.285], respectively; all negative-control intervals include zero. At equal shot and sensingwindow budgets, an ideal Ramsey/echo estimator is more precise in every tested class. The result supplies an explicit connection between a benchmarking identifiability limitation and a controlled correlation measurement. No native or quantum-memory detection is claimed.

The Cops and Robber game is extensively studied in the sequential setting where the main goal is to capture the robber. Capturing the robber means at least one cop, and the robber will be at the same vertex together at some time. There are several variants, including variants where the goal of the cops is to surround the robber. Surrounding the robber means there is at least one cop in each of the neighboring vertices of the robber's position. In this paper, we introduce it in the distributed setting while empowering the robber by saying it can even kill cops. Specifically, in our model, the robber moves in odd rounds and has unbounded speed, cops move in even rounds, and if one or more cops move into a vertex where the robber is currently residing, all these cops get killed. We call this lethal robber. This also connects our work to the Intruder Capture and Black Hole Search problems by introducing an entity which is dynamic as well as lethal, a setting that, to the best of our knowledge, has not been studied. In this work, we introduce the lethal robber, define the cops and lethal robber problem in the distributed setting and study it on a static ring of size $n$. We prove $n$ cops are not enough, even if all start from the same vertex, and provide an algorithm starting from an arbitrary initial configuration that requires $n+\lfloor\log n \rfloor+4$ cops in the worst case.

Calibration, leak localisation and sensor placement on water distribution networks (WDNs) are decisions about continuous parameters, yet the hydraulic engine that defines the physics returns a solution and no derivatives, so practice falls back on derivative-free search or on surrogates whose error the answer inherits. We make the global gradient algorithm itself exactly differentiable: the forward pass reproduces the reference engine's discrete devices, status switching and low-flow linearisation included, and the backward pass solves the implicit adjoint by reusing the forward pass's terminal factorisation, so one extra sparse solve returns every parameter's gradient at once, batched over scenarios on one graphics processor. Across 52 public, synthetic and operational networks and 8,140 simulation frames, every network meets the acceptance criterion, the largest head deviation from EPANET 2.2 is 1.137e-13 ft and 25 agree exactly. One adjoint solve replaces the 906 simulations a finite-difference roughness Jacobian costs on the 905-pipe L-TOWN benchmark, and a leak-inversion training loop runs at 463-470 ms per optimiser step for 256 scenarios, 191 times the prior pipeline. Gradient calibration reaches its endpoint within a median 595 model calls, where the strongest of five tuned metaheuristics needs 8,060 to match it on the training loss and two never do within 20,000. On a 554-link operating network, one adjoint pass audits, pipe by pipe, which roughness parameters the installed sensors can constrain and which sensors to add, on the model the utility already operates.

Two centuries ago Pierre-Joseph Proudhon envisioned a social contract that is (1) an agreement of man with man; (2) reciprocal; (3) imposing no obligation upon the parties except that which results from their personal promise; (4) subject to no external authority; (5) freely accepted and signed by all the participants; (6) of the nature of a contract of exchange. He also envisioned a property of social contracts, analogous to a property of digital platforms we term grassroots: that one could "make a contract with all, as ... with some"; digitally, that a grassroots platform can have multiple instances, which may coalesce by mutual consent. We define grassroots social contracts as social contracts meeting these conditions and that (7) people are free to deal with each other; and (8) there is no external register of people. We show that a grassroots social contract can be transformed into a working grassroots platform through an abstraction cascade, from the contract text to formal act schemas, verified syntactically to be grassroots, and then to volition-guarded multiagent atomic transactions. Act schemas are a formal language for the acts the contract describes, each naming the parties' roles, which of them must will the act, and its precondition and effect at each role. Any contract written in this language meets the eight conditions, provided it is syntactically grassroots, satisfying three decidable conditions, Introduction, Provenance and Volition. We prove that the protocol realising a syntactically grassroots contract is grassroots, and volitionally so. Clauses of a grassroots social contract are of three legal types: a breach of an enforced clause is impossible with a correct implementation; a breach of an attested or undertaken clause can be taken to court, with signed evidence for an attested clause. We illustrate them with a grassroots social graph and grassroots currencies.

Efficiently aggregating and orchestrating computing power across heterogeneous clusters for HPC workflows faces four practical challenges: preserving workflow context across independently administered clusters, moving large datasets between sites, reasoning about site-specific environments and scheduler policies, and exploiting live queue and resource states for efficient task scheduling. To this end, we design Diamond Agent, an agentic system that enables intelligent execution of HPC workflows across heterogeneous clusters with typed skills as the interface. Diamond Agent provides an agent-facing workspace and skills that unify cross-site resource discovery, resource specification, data movement, task execution, and result retrieval. A centralized Diamond Agent instance can operate multiple supercomputers without being deployed separately on each login node. Diamond Agent translates high-level agent actions into valid site-specific executions, moves data through Globus Transfer, and uses live system capability and queue information to select feasible placements. Its event-driven continuation mechanism decouples agent actions from long-running batch jobs: persistent services monitor remote execution and resume the agent only when a result or decision-relevant event is available. We experiment with 27 hours of telemetry and 19 matched multi-site submission rounds comprising 83 jobs across four production supercomputers. Compared with a fixed-site baseline, Diamond Agent reduces the median additional completion time relative to the fastest observed placement from 42 seconds to 4 seconds, a 10.5x reduction.

Heterogeneous Federated Learning (HFL) aims to train models across devices with diverse resource budgets while preserving data privacy. Existing HFL methods typically bind training to a small predefined menu of model configurations, which limits architectural coverage. To address this bottleneck, we introduce Federated Adaptive Network Search (FANS), a hypernetwork-based framework that learns a shared architecture space rather than a fixed set of client models. To optimize this shared space efficiently, we propose the Federated Parallel Scaling (FPS) algorithm, which jointly trains multiple sampled subnetworks in parallel with self-distillation so that larger sampled subnetworks can supervise smaller ones during local updates. We evaluate FANS on CIFAR-10, CIFAR-100, and MNLI using ResNet-18, DenseNet-121, and BERT-base, respectively. Across all benchmarks, FANS expands the feasible subnetwork pool by orders of magnitude (e.g., 4,680 candidates for ResNet-18 vs. 4 in existing methods) and improves the average accuracy-efficiency trade-off relative to representative HFL baselines. Device heterogeneity is emulated through resource tiers, and evaluation covers accuracy, parameter count, and MACs.

With the rapid advancement of accelerator technologies, pre-training large language models (LLMs) on heterogeneous accelerator clusters has become increasingly crucial for maximizing hardware utilization. Existing systems, however, suffer from inaccurate training time modeling, which undermines the parallelization optimizations built upon it. Moreover, for current approaches, the vast configuration search space makes exhaustive exploration infeasible, forcing a trade-off between search time and training efficiency. To overcome these limitations, we introduce Poseidon, an efficient and scalable LLM training framework designed with heterogeneity awareness. Its core is an explicit training time model based on a directed acyclic graph. Building on this graph, Poseidon employs two efficient, theoretically grounded strategies: stage-level pruning via early stopping with partial estimation, and layer-to-stage mapping exploiting a ridge-like distribution pattern. These strategies reduce the search space without sacrificing optimal training efficiency. Experiments on heterogeneous clusters show that Poseidon improves training throughput by up to $2.76\times$ over state-of-the-art systems.

Low earth orbit (LEO) satellite constellations enable geographically distributed ground devices to collaboratively train a global model via federated learning (FL) without sharing raw data, with applications in environmental monitoring and disaster prediction. However, in satellite-assisted FL scenarios, intermittent satellite-ground links allow only a subset of devices to participate in global aggregation within each visibility window, leaving unscheduled devices idle and their local computational and data resources underutilized. Under partial device participation, data heterogeneity among devices may bias the global model toward certain devices, thereby deteriorating learning performance. In this regard, we propose a continual computing based federated learning framework, referred to as CoCoFL, in which scheduled devices participate in the global model aggregation, while unscheduled devices continue updating their local models taking into account model staleness. Guided by the convergence analysis of CoCoFL and subject to visible-window-related time constraints, we jointly optimize the device scheduling and the number of local epochs for scheduled and unscheduled devices. Experimental results demonstrate that CoCoFL achieves faster convergence, lower training loss, and higher test accuracy compared with baselines.

Mixed-integer programming (MIP) is a cornerstone in applied optimization, both in industry and academia. Recently, there has been increased attention to finding strong primal solutions quickly. This is reflected, for example, in the development of the NVIDIA cuOpt solver and, most recently, in the new MIPFEAS benchmark, which has a tight time limit of 600 seconds and evaluates solvers based on how quickly they find high-quality primal solutions. This article introduces a MIP portfolio parallelization scheme, focusing on efficiently exchanging information between its workers. We present two implementations of this scheme: one built directly into the open-source MIP solver SCIP, and an external one, which we call ReXi. ReXi is currently the fastest non-commercial solver in the MIPFEAS benchmark, followed by the SCIP-integrated implementation. Moreover, we present new versions of both implementations that considerably outperform their predecessors on the MIPFEAS benchmark.

Conventional federated learning relies on parameter averaging, which forces clients to be doubly homogeneous: all must run an identical architecture, and accuracy degrades when local data are non-IID. Decentralized federated distillation sidesteps both: each client runs its peers' model snapshots as teachers on its own local data and distills from their soft predictions, with no server, no public data, and no shared architecture. Under severe non-IID skew, however, the trustworthiness of the aggregated teacher target is a matter of degree, yet existing pipelines make hard, all-or-nothing decisions: outlier teachers are discarded by threshold, and whatever target survives is trusted in full. We propose CALM, which replaces every hard decision with a smooth trust gate at three levels: per class, teachers are weighted by agreement with the peer consensus; per sample, distillation is scaled by the teachers' divergence from that target; and a label gate scales it by how strongly the target supports the sample's true label. None of this adds communication or auxiliary data. On CIFAR-10, SVHN, OrganAMNIST, and Google Speech Commands with heterogeneous client architectures under Dirichlet label skew, CALM consistently outperforms uniform and hard-filtered distillation and matches or exceeds competing heterogeneous-FL methods.

We present the first polynomial lower bounds for several fundamental problems in the distributed graph sketching model in the tiny-error regime, which includes deterministic algorithms as a special case. In the graph sketching model, every node sends a single message to the referee who does not have any prior knowledge of the graph and must output the answer. While the work of Nelson and Yu (SODA 2019) and Yu (SODA 2021) showed that $\Theta( \log^3n )$ is optimal for constructing a spanning forest or deciding whether the graph is connected with error at most $\frac{1}{\text{poly}(n)}$ , their approach does not yield any stronger bounds for significantly smaller error probabilities. Our main result is to show that solving either connectivity or spanning tree construction with error at most $\delta$ requires messages of length $\Omega( \min\{n, \log_2 \frac{1}{\delta}\}^{1/3} )$, which implies that algorithms with exponentially small error must send messages of $\Omega( n^{1/3} )$ bits in the worst case. Our results significantly narrows the current gap between the Jelani-Yu threshold of $\Theta( \log^3n )$ and the trivial upper bound of sending $O(n)$ bits per node for deterministic graph sketching. We also extend our results to $k$-edge connectivity. For any $k=O(n^{1/7})$, we recover the same bound of $\Omega( k )$ on the message length for algorithms with exponentially small error that was shown by Robinson and Tan (PODS 2026) only for deterministic algorithms. Finally, for $k=n^{o(1)}$, our result implies a stronger lower bound of $\Omega_\epsilon( n^{\epsilon} )$ bits, for any constant $\epsilon<\tfrac{1}{3}$.

Instance-billed serverless platforms charge for CPU and memory over the lifetime of a warm instance, making idle inference state a direct serving cost. We present billing-aware neural text-to-speech (TTS) serving on serverless CPUs, optimizing CPU-seconds and GB-seconds rather than throughput or latency alone. Conventional runtimes are poorly suited to this setting: per-request parallelism causes CPU contention under concurrency, while warm instances retain gigabytes of billable inference and page-cache state. We address these costs with request-sized concurrent inference, which bounds per-request CPU parallelism, and a reclaimable instance lifecycle, which releases inference state and page-cache memory after idle periods while retaining the server process and compile cache. On Kokoro-82M, our system achieves 2.71 audio-seconds per CPU-second versus 0.89 with ONNX Runtime defaults and reduces cost per audio-hour from $0.0631 with PyTorch to $0.0153, a 4.1x reduction. Idle billed memory falls from 8.7 GB to 1.33 GB, while restoration reaches first audio in 2.2 s versus 7.7 s for a PyTorch cold start. Under bursty traffic, lifecycle reclamation is essential for translating inference efficiency into lower serverless cost.

GPU applications on unified-memory (UMA) edge platforms often inherit a discrete-GPU memory abstraction in which they allocate one buffer for the CPU, another for the GPU, and copy data between them before and after GPU execution. On UMA hardware these buffers reside in the same physical DRAM pool, so the copies consume bandwidth, time, and energy. Despite the growing adoption of UMA platforms, this pattern remains common because production software stacks, libraries, and samples were written for portability across discrete GPUs. However, removing these copies is not as simple as merging the two buffers, because the original program may rely on the two buffers being distinct, or on the copy itself ordering CPU and GPU accesses. Echo removes these copies only when it preserves the data values and access ordering on which the program depends. It does so along two complementary paths: source-level rewriting and binary deployment. Echo-SR is a compile-time LLVM transformation that proves safety and rewrites accepted pairs in place. Echo-DR is a profile-guided binary optimizer for existing applications that profiles and validates stable allocation/copy patterns and, at runtime, intercepts the matching calls to unify profile-matched pairs while preserving the ordering effects of removed copies. Across seven benchmarks on three NVIDIA Jetson platforms, Echo's benefit grows with the fraction of baseline time spent on copies. Copy-dominated workloads speed up by up to 7.05x, closed-source end-to-end applications speed up by up to 1.40x. On the five source-available Orin benchmarks, both automatic paths recover >=99% of the gain of a manually optimized reference.

Mysticeti and Mahi-Mahi represent leading approaches to consensus protocols, leveraging a novel uncertified Directed Acyclic Graph data structure to achieve substantial performance benefits compared to prior work. Mysticeti and Mahi-Mahi differ in their underlying network assumptions, where the former makes assumptions about timing bounds for communication while the latter does not. Mysticeti trades robustness for lower levels of latency in ideal conditions, whereas Mahi-Mahi favours increased latency in order to ensure stronger progress guarantees in asynchronous settings. Although dual-mode protocols exist (i.e., protocols that can adapt do different network assumptions), none have been implemented to take full advantage of the novel data structure seen in the aforementioned protocols. Thus, our research project aims to trace the progression of consensus protocols, analyse how they address network assumptions constraints, and explore dual-mode protocols. Then, we present Sharpedo, the first dual-mode uncertified Directed Acyclic Graph consensus protocols which aims to achieve the best of both worlds: the high-performance of partially-synchronous protocols, while providing guarantees of progress even in asynchronous settings. Furthermore, Sharpedo integrates a novel dynamic scheduling mechanism resulting in improved system efficiency and performance optimization.

Grid studies often represent data-center flexibility as a fixed percentage of load, although no public production trace has shown how much eligible load persists across event durations or co-moves across clusters. We reconstruct 4,439 hourly power observations from a 185-day trace of 155,410 GPUs and derive a workload-semantic flexibility envelope. The fleet's time-averaged Monte Carlo median facility demand is 55.8 MW, while immediate eligible curtailment averages 3.55 MW after retaining allocated-GPU idle power: 12.1% of workload power and 6.35% of median facility power. Under full realization of that eligibility, 95%-available relief falls from 2.51 MW for one hour to 2.32 MW for four hours and 1.95 MW for 24 hours; a common realizable fraction q scales every value exactly by q. A mean-calibrated scalar overstates these quantities by 17%, 25%, and 47%, while a scalar tail-calibrated at four hours understates the one-hour product by 6% and overstates the 24-hour product by 17%; the share that reproduces the surface varies by a factor of 1.6 across durations and reliability levels. Aggregating 13 clusters raises four-hour firmness from 0.38 to 0.66, but cross-cluster covariance limits the gain. The production scheduler exposes almost no additional delay-based capacity: newly deferrable arrivals average 0.008 MW and have zero 95%-available capacity. These results replace an assumed flexibility percentage with duration, reliability, portfolio, and realizability terms that can be written into interconnection and demand-response contracts.

Comparisons between GPU implementations are usually asymmetric: one side is tuned by its author, the other is run as found. I report a programme that tuned both a novel SOM algorithm (SparseBin) and the baseline algorithm it was being compared to (cuSPARSE). The best-matching-unit search that dominates self-organizing map training was tuned through four levers - tile size, tile-membership clustering, neuron-axis chunking and vectorised loads - reaching 5.6-10.1x per epoch over the previously published configuration at map sizes from 32x32 to 512x512, and lifting the margin over the CUDA implementation behind our earlier MEDLINE atlases from ~80x to ~385x. cuSPARSE, the implementation SparseBin is compared against, received every lever with an analogue on its side, and became 2-3x faster in the process. The tuned kernel pressed the L2 bandwidth roof at 77% of peak with every other unit at 40-65%, bounding any further lever at ~1.3x - a terminal result rather than a waypoint, and every untested lever was either capped by that bound by construction or measured null.

Federated scientific machine learning enables institutions to train neural surrogates without centralizing local physical data, yet studies of partial differential equations (PDEs) lack a transferable definition of non-independent and identically distributed data. Existing protocols partition coordinates, coefficients, boundary conditions, or geometries according to equation-specific rules. Here, we introduce solution-space PDE-Dirichlet, a protocol that converts continuous supervised responses into reusable solution bins and quantifies the realized separation between clients through optimal transport over the geometry of these bins. We derive an exact inverse relation between population allocation heterogeneity and the Dirichlet concentration, and we establish conditions under which response heterogeneity induces gradient disagreement, local-update dispersion, and parameter divergence. Across seven controlled and public PDE tasks, three neural-operator families, and five random seeds, a lower concentration consistently increases the realized solution distance and optimization heterogeneity. The degradation in final error is task dependent: the largest effect occurs for low-viscosity Burgers, reaching 4.157 percentage points under the most heterogeneous setting, whereas additional communication or smoother dynamics can reduce the final gap despite persistent parameter separation. These results distinguish a reproducible geometric mechanism from task-dependent generalization outcomes and provide a common basis for evaluating non-IID federated PDE learning.

Distributed DNN inference is increasingly deployed in containerized edge-cloud environments, where workloads run on-device or are exposed to remote clients over the network. Accurate online power estimation on resource-constrained ARM nodes without hardware power counters such as RAPL remains a challenge, and CPU-only models fail to capture multi-resource behavior. We present GreenPipe, an automated profiling-training-validation pipeline that builds multi-resource regression models from external power meter measurements and attributes power to containers proportionally. GreenPipe is evaluated on a Raspberry Pi 4 edge node in a K3s edge-cloud testbed, covering DNN inference with three vision models, multiple precisions, thread counts, and both local and serving scenarios. System-level MAPE is 6.3-9.4%, improving over CPU-stress and utilization-only baselines by 26.9% MAPE on average. We jointly report inference latency and energy per inference, exposing performance-energy trade-offs across workload configurations.

Due to resource constraints or external and internal uncertainties, clients in real-world federated learning systems are often intermittently available edge devices. In highly dynamic environments, the parameter server lacks prior real-time knowledge of clients' availability, making it challenging to adapt traditional federated learning algorithms to be resilient to uncertainties in client availability. If not carefully addressed, complex client availability can introduce significant bias, potentially harming the performance of the trained model. Most prior work either fails to account for non-stationary client availability dynamics or demands significant memory and computational overhead. This paper aims to develop efficient federated learning algorithms that are provably resilient to heterogeneous and non-stationary stochastic client availability. We propose FedSWE, which admits novel algorithmic structures to (i) compensate for missed computations, (ii) stabilize and diffuse the global updates over rounds, and (iii) evenly mix the local updates through implicit gossiping, despite being agnostic to non-stationary dynamics. Compared with the standard FedAvg, FedSWE introduces light additional memory and computation overhead. We show that FedSWE converges to a stationary point of non-convex objectives while achieving the desired linear speedup property in certain special cases. We corroborate our analysis with numerical experiments over diversified client unavailability dynamics on real-world data sets.

Major open-source serving stacks reuse the key-value tensors of a shared prefix by default. We measure what this prefix reuse costs in reproducibility. Holding model, decoding, seed and request order fixed at batch size one, we ran an eighty-episode multi-turn agentic tool-use workload with prefix reuse on and off across two engines and four weight formats. Without prefix reuse, repeated runs reproduced every token and log-probability in all ten configurations. With it, cache-hit and recompute paths diverged on both engines. On llama.cpp with GGUF weights this path divergence rises with quantization: for Qwen2.5-7B over ten independent server histories, it reached 37.4 percent of episodes at 16-bit and 82.5 percent at four-bit. Near-tie decisions became only about 1.2 times as common, while the cache-path change in top-two margins grew 4.9-fold from 16-bit to three-bit. Serving four-bit weights through 16-bit kernels left divergence not detectably different from 16-bit, placing the amplification mainly in the quantized execution path. At Q4_K_M, run-to-run divergence came mainly from llama.cpp's server-level prompt cache: under the canonical episode order it moved divergence by 37.5 points, and at its default, divergence spanned 20.0 to 28.7 percent under five random orders. With cache state restored, each path reproduced exactly while the two still differed. In the tested stacks, cached serving is irreproducible because the cache state is absent from the request and not reset by default. A single-turn bridge shows path divergence reaching task outcomes with no detectable net change in accuracy. We release the harness, logs, and analysis pipeline.

Modern CPUs increasingly integrate high-throughput matrix engines optimized for low-precision AI workloads, while many scientific computing applications still rely on FP32 and FP64 GEMM to meet their numerical accuracy requirements. This mismatch motivates an algorithmic bridge that uses low-precision matrix products to emulate higher-precision GEMM. This paper presents a CPU-oriented method based on Intel Advanced Matrix Extensions (AMX) and BF16 matrix products. For FP32, each operand is decomposed into three BF16 components and six selected component products are evaluated, targeting FP32-level accuracy relative to oneMKL SGEMM without claiming elementwise or bitwise identity. For FP64 inputs within the supported BF16 exponent range, the method uses a simplified fixed six-slice Ozaki decomposition. Each retained BF16 product is first produced in FP32, then widened and accumulated in FP64. Four product-count settings retain 6, 10, 15, or 21 component products, exposing the accuracy–performance tradeoff relative to oneMKL DGEMM. The implementation combines precomputed packed component buffers, VNNI-packed $B$ panels, and an FP32 tile-resident operand-reuse schedule. On the tested square matrices, AMX-FP32 exceeds oneMKL SGEMM throughput. For AMX-FP64, low-product-count variants can exceed DGEMM at sufficiently large orders, while retaining more products improves accuracy at additional cost.

In training Mixture-of-Experts (MoE) models, sharded data parallelism partitions each expert's parameters across GPUs, requiring an Allgather operation to reconstruct the full weight matrix before each layer executes. This communication often dominates iteration time. Prior work often reduces this overhead using lossy compression methods that sacrifice numerical fidelity, while existing lossless methods do not exploit cross-iteration exponent stability. In this paper, we propose Cross-Iteration Exponent Reuse Allgather (CIERA), a lossless, system-aware communication method for sharded MoE training. We observe that after a brief warmup phase, the exponent values of most weights remain unchanged across iterations. Based on this, we cache exponents locally and transmit only the sign and mantissa when exponents are unchanged. The receiver reconstructs the original weights exactly by combining the cached exponents with the received data. Since parameter matrices vary in shape across layers, compression is applied only when it yields net time savings. Moreover, the compression operation is overlapped with both Allgather communication and computation. Our real experiments and large-scale trace-driven simulator show that on OLMoE-1B-7B at 16 GPUs, CIERA achieves a 3.70x speedup over the lossless baseline and 3.68x over the lossy baseline, projected to reach 4.28x and 4.42x respectively at 128 GPUs, while preserving bitwise-exact parameter reconstruction in all evaluated runs.