Papers

Achieving cost efficiency while meeting strict user-facing SLOs (e.g., time-to-first-token) remains a fundamental challenge for cloud GPU clusters serving large language models (LLMs). Autoscaling is the key mechanism for cluster resource management, yet a basic system design question is open for serving LLMs: what should be the unit of scaling? Existing approaches primarily treat the entire model as a monolithic scaling unit–simple but unable to capture the fine-grained dynamics of inference workloads. As a result, such coarse-grained scaling often leads to either SLO violations under bursty demand or significant GPU under-utilization. Our characterization reveals substantial operator heterogeneity, exposing operator-level elasticity as a viable scaling primitive. We present OpScale, a practical operator-level orchestration framework of profiling, provisioning, placement, and runtime serving. OpScale is designed to tackle the high complexity and the space explosion problem, arising from operating at this finer granularity. Evaluated with production traces on up to 40 A100s and 24 GB200s, OpScale attains SLOs with up to 36.3% fewer GPUs and 28% less power, or achieves 44% higher throughput under fixed cost budgets.

Tendermint is among the most widely studied and deployed Byzantine fault-tolerant (BFT) consensus protocols, owing in part to its native leader-rotation mechanism that subsumes complex view changes. Like most partially-synchronous BFT protocols, Tendermint tolerates $f < n/3$ Byzantine processes and decides in three communication steps. Motivated by the push for lower-latency blockchains, a recent line of work shows that consensus can be solved in two communication steps when $f < n/5$. We present Fast Tendermint, an adaptation of Tendermint to the $n > 5f$ setting that decides in two communication steps in the good case, while preserving Tendermint's leader-rotation structure. Fast Tendermint collapses Tendermint's prevote and precommit steps into a single voting step and merges the $locked$ and $valid$ state. We give proofs of agreement, validity, and termination, and a formal specification in Quint, a modern surface syntax for TLA+, used to model-check the protocol.

The Lovász Local Lemma (LLL) is a probabilistic tool that has been shown to be of central importance in the study of distributed algorithms. For example, the constructive LLL is known to be complete for the class of locally-checkable labeling problems with $o(\log n)$ randomized complexities in the LOCAL model. One classic application of the LLL is in coloring graphs with some sparse structure, such as triangle-free graphs. Triangle-free coloring therefore serves as a benchmark problem for techniques for sublogarithmic randomized distributed algorithms. The state-of-the-art distributed triangle-free coloring algorithm of Pettie and Su [ICALP 2013, Information and Computation 2015] uses $\frac{\Delta}{k}$ colors (where $k$ can be up to $(\frac14 - \varepsilon)\ln \Delta$) and consists of $O(k+\log^* n)$ applications of the distributed LLL. However, the distributed LLL is itself a difficult problem; despite significant study, the fastest algorithms known require $O(\log_\Delta n)$ or $O(\frac{\Delta}{\log\Delta})+\log^{O(1)}\log n$ rounds. In this work, we adapt the Pettie-Su's algorithm so that the resulting LLL instances can be solved in $\log^{O(1)}\log n$ rounds, by employing the 'resilience' definition of Davies [SODA 2023]. This gives an $O(k)+ \log^{O(1)}\log n$ complexity (since the LLL is not needed when $k= \log^{\omega(1)}\log n$), essentially causing the LLL steps to no longer be the bottleneck of the algorithm. As a corollary we obtain the first $\log^{O(1)}\log n$-round algorithms for coloring triangle-free graphs with $o(\Delta)$ colors. The same framework also yields a companion girth-$5$ algorithm, using $(1+\varepsilon)\Delta/\ln \Delta$ colors in $O(k)+ \log^{O(1)}\log n$ rounds, matching the best known existential upper bound for the number of colors.

Large language model serving faces a critical memory bottleneck: the KV cache grows with sequence length and batch size. PagedAttention uses fixed-size memory blocks to reduce allocator-level fragmentation, but recent KV eviction algorithms operate at a token granularity finer than block-level management. This mismatch causes intra-block fragmentation, leaving a large fraction of allocated KV memory unreclaimable. We present vToken, a lightweight token-level virtualization layer that decouples logical token liveness from physical block placement. vToken maintains a stable logical token view through token-table indirection and realizes physical reclamation by repacking live tokens asynchronously. The design preserves PagedAttention kernels and CUDA Graph compatibility. We implement vToken in vLLM and evaluate it with H2O, Random, and Scissorhands across models. Compared with a paired Naive-Evict baseline, vToken reduces retained KV blocks per request by 27.2%–72.3% and improves SLA-constrained throughput by up to 1.37$\times$. Under a constrained active-KV budget, it extends the maximum feasible concurrency by up to 2$\times$, while reducing the per-policy integration footprint from 500+ lines to under 50.

As edge-side vision services continue to expand toward low-latency, high-throughput scenarios, reducing the inference cost of vision models without sacrificing reliability has become a central concern. Existing semantic caching methods largely rely on empirical similarity thresholds; while such thresholds improve hit rates, they tend to introduce silent misclassifications near decision boundaries. To address this, we propose LipCache, a certified semantic caching framework for image classification. Without modifying the existing deployed main model, MainNet, the framework introduces a lightweight network, GuardNet, that maps inputs into a low-dimensional feature space subject to a Lipschitz constraint. It then computes a per-sample certified reuse radius from the local classification margin and the spectral norm of the classification head. At runtime, a cached result is reused only when the query feature falls inside the certified reuse ball; otherwise, the query falls back to MainNet. Thus, cache hits are transformed from empirical threshold tests into geometric certification decisions with explicit theoretical boundaries. Across standard image classification tasks like CIFAR, Tiny-ImageNet, and SVHN, LipCache achieves a measured speedup of up to $1.65\times$ with limited end-to-end accuracy degradation, while all accepted cache hits satisfy the GuardNet-side certified-consistency condition. Furthermore, an enhanced GuardNet training recipe substantially improves cache hit rates in the Tiny-ImageNet multi-class extension while maintaining a certified-consistency rate of $100\%$. These results demonstrate that per-sample certified reuse can reduce main-model fallback while preserving theoretical consistency, providing a feasible approach to reliable cache-assisted inference at the edge.

Recent advances in large language models have made CLI-based AI agents a practical tool for accelerating GPU porting of large legacy scientific applications. Such applications, however, are not merely old code bases; they are scientific assets whose credibility has been accumulated through long-term development, comparison with observations, and use in domain studies. GPU porting must therefore preserve this scientific validity while adapting the implementation to GPU-centric HPC systems. This paper presents a validation-centric AI-assisted GPU porting workflow through a case study of CReSS, a legacy Fortran weather simulation code with more than 250,000 lines. The workflow uses an AI agent to extract OpenMP regions, generate dump-based kernel benchmarks from physically meaningful simulation states, apply OpenACC transformations, and validate results through element-wise comparison with dumped reference data and application-level validation. Using a real typhoon simulation, the workflow produced numerically validated GPU implementations for 162 target kernels and achieved a 5.1x application-level speedup within practical wall-clock development cost. In particular, it detected numerical discrepancies in five kernels caused by floating-point and intrinsic-function differences, including threshold-sensitive branch divergence and cancellation effects, enabling feedback to the application developers. The case study suggests that, for large legacy scientific applications requiring dump-based validation, practical AI-assisted GPU porting must manage session-spanning context, runtime-state reconstruction, and costly recovery from small static-analysis omissions. These findings demonstrate that AI-assisted GPU porting requires not only code generation, but validation-centric workflow design.

Service meshes facilitate service-to-service communication and enforce security policies in microservice architectures. However, they often depend on per-pod sidecar proxies, which introduce significant latency and resource overhead due to redundant application-layer parsing on every request. Eliminating sidecars without compromising security guarantees remains a central challenge. To address this, we introduce Meshlib, a sidecar-less service mesh extension built on Cilium as a control-plane extension. Meshlib incorporates a non-intrusive application-bound library that enforces Layer-7 policies within the application process, while delegating transport-level identity and routing to Cilium's eBPF-based data plane. This separation of responsibilities removes sidecar-induced latency and maintains the security semantics of the service mesh. The architecture remains fully interoperable with unmodified services, enabling incremental adoption within existing deployments. We evaluate Meshlib against Istio, Linkerd, and unmodified Cilium on the TrainTicket benchmark, enforcing 126 security policies across 37 services and show that it achieves the lowest end-to-end latency of all evaluated configurations with comparable resource overhead.

In expert-parallel (EP) MoE serving, every layer synchronizes at the slowest GPU. Dispatchers balance token counts (EPLB, LPLB, UltraEP) or activated-expert counts (METRO), assuming expert time is linear in one. Measurements on two datacenter GPU generations show it is neither: below $n^* \approx 156$–$168$ tokens, HBM weight streaming dominates—cost attaches to $activated replicas$, not tokens; above it, grouped GEMM rounds tokens to 128-tile $M$-tiles, so $splitting$ an expert adds padded compute. A max-affine profile $t=\max(a+bG,\,c+\beta N)$ captures both regimes. Realistic decode batches hold hot experts in the linear regime and cold in the flat $simultaneously$; recorded batches show proxy dispatches differ by $1.4$–$1.6\times$ in modeled block time (p95 up to $1.7\times$), and $which$ proxy wins flips with the regime. We formalize per-batch dispatch as a fixed-charge makespan problem—NP-hard on two fully replicated GPUs, polynomial in degenerate limits—and present TEMPO, a makespan-aware dispatcher solving it in milliseconds off the critical path; its SGLang integration runs out-of-process and fuses dispatch with count collection into one in-graph kernel. Anchored by an 8-GPU Testbed A microbenchmark, TEMPO stays within $1\%$ of the best fixed baseline everywhere and wins by up to $15.5\%$ where regimes mix. End-to-end on Testbed B, Qwen3-235B (inside the win region) gains $4$–$6\%$ throughput and cuts p99 latency by $\sim 15.6\%$; DeepSeek-V3 (outside, communication-dominated) shows only mechanism cost. A phase diagram, not a universal win, is the claim: it predicts both outcomes before deployment.

The growing power demands and variability of AI workloads make electrical power delivery a critical constraint in data-center operation. Distributed energy storage can reduce the power capacity required to support stochastic loads, but its benefits depend fundamentally on the statistics and time scales of demand. This paper develops a probabilistic framework that jointly characterizes provisioned power, energy-storage capacity, and the probability of overdraw. We show that storage-assisted provisioning separates into two operating regimes. In the Small Battery Region, overdraw is dominated by short-lived demand excursions and storage provides nearly linear reductions in the required power margin. In the Large Battery Region, overdraw results from sustained demand fluctuations over longer spans of time, and the required margin exhibits diminishing returns with storage. For this regime we introduce effective power, an analogue of effective bandwidth that captures the temporal statistics of the demand and gives an asymptotically tight characterization of the required power. We further quantify how temporal correlation and spatial aggregation affect storage requirements and statistical multiplexing gains, and extend the analysis to loads with multiple demand time scales. Finally, we evaluate the framework using power-demand traces from three production data centers spanning HPC, GPU-training, and cloud-service workloads. Despite their heterogeneous, cyclo-stationary and multi-modal behavior, the measured workloads exhibit the predicted regimes, and a simple four-parameter two-state model captures the dynamics governing their storage-power tradeoffs. The resulting framework provides both a probabilistic foundation and practical dimensioning principles for storage-assisted power provisioning in next-generation AI data centers.

We study the fundamental problem of implementing $m$ linearizable LL/SC objects with constant expected step complexity in a system of $n$ processes, using bounded base objects commonly available in hardware. Assuming that each process may have at most $\tau$ outstanding LL() operations, the best known deterministic algorithm requires $\Omega(n^2\tau + m)$ base objects (CAS objects and registers) [Blelloch and Wei, DISC 2020]. Previously, no comparable randomized algorithm was known. By employing randomization and FADD objects in addition to CAS objects, we obtain a space bound of $O(n\tau+m)$ against the weak adaptive adversary. For $m=O(1)$ this matches a lower bound for algorithms using CAS objects and registers [Aghazadeh and Woelfel, PODC 2015]. In addition, our object can be employed by quiescently history-independent (QHI) algorithms: Whenever no operation on the object is pending and no process has an outstanding LL() operation, its internal memory state is uniquely determined by the values of the $m$ LL/SC objects. An important application is a recent QHI dynamic hashing algorithm, which uses $\Theta(m)$ hardware LL/SC objects to maintain a hash table of size $m$ [Attiya, Bender, Farach-Colton, Oshman, and Schiller, STOC 2025]. But LL/SC is not available in hardware, and prior to our work no wait-free or efficient lock-free software implementation of LL/SC with similar properties was known. Our work demonstrates that one can actually implement the hashing algorithm on available hardware, without an asymptotic increase in step and space complexity, under the reasonable assumption that $m=\Omega(n)$.

The rapid growth of LLM inference is shifting sustainability concerns from one-time training to continuous serving, where infrastructure decisions shape energy use, carbon emissions, water consumption, and service quality. Yet operators often need to compare deployment alternatives before large-scale infrastructure is built, making direct measurement costly, slow, and sometimes infeasible. We present InFactPlanner, a trace-driven decision-support framework for what-if analysis of sustainable AI data center deployment for LLM inference across single and geo-distributed sites. InFactPlanner combines query traces, hardware-model profiles, candidate site configurations, PUE/WUE parameters, renewable generation models, and time-varying grid carbon intensity to estimate power, energy, carbon emissions, water use, latency, and server utilization. The framework abstracts low-level serving effects into configurable hardware-model profiles, enabling rapid comparison of site selection, capacity placement, hardware, model, renewable integration, and routing choices. We validate the energy accounting pipeline by reproducing reference LLM inference energy estimates with less than 10% deviation, evaluate scalability across multiple data centers and server counts, and demonstrate scenario-driven decision analyses for hardware selection, renewable placement, geographic deployment, and carbon-aware routing. Our results show that sustainability-optimal choices can differ from latency-optimal ones, and that the carbon value of deployment depends strongly on the local grid mix.

Medical AI has demonstrated specialist-level diagnostic accuracy, yet these capabilities remain largely inaccessible in resource-constrained rural settings where bandwidth is scarce, compute is limited, and clinical decision-making requires integrating heterogeneous modalities. We introduce a cloud–edge collaborative architecture that addresses these constraints: lightweight, domain-specific models on the edge transform raw medical data into compact structured outputs, while a cloud LLM synthesizes these outputs into clinical summaries. An LLM-based orchestrator dynamically selects diagnostic tools based on patient context, promoting comprehensive modality coverage without processing irrelevant inputs. We evaluate on 20 multimodal clinical cases spanning cardiac, obstetric, trauma, and screening scenarios under three simulated network profiles (500,kbps–5,Mbps). The hybrid system achieves 98–99% diagnostic tool recall with 92–96% precision, matches or exceeds cloud-only baselines on clinical accuracy, and maintains bandwidth-invariant latency (25–35,s) at 4–15x lower token cost. These results highlight the role of architectural design in enabling efficient multimodal integration and improving factual grounding compared to cloud-only approaches under deployment constraints.

Systems that generate GPU kernels with language models report high correctness rates. Those rates come from a single loose test: run the kernel on a few random inputs at one fixed shape and accept it if the output is close to a reference. A kernel can pass that test and still be silently wrong. It can return an ordinary number where the true answer is a NaN or an infinity, differ from run to run, break when the shape changes, or accumulate in fp16 where the reference keeps an fp32 total. We build the instrument that checks correctness properly: a contract-grade verifier of twelve adversarial gates, each a property a correct kernel must satisfy, several of them tolerance-free, so no choice of threshold can explain a failure away. Aimed outward, the verifier audits 2,638 machine-generated kernels that a public system's own harness had already accepted as correct. It finds 39.5% broken beyond any tolerance argument and 62.1% carrying at least one violation. The field's standard test accepts 1,487 kernels the verifier rejects, against only 14 the other way. We defend the finding four independent ways: a 7/7 positive control, a threshold-calibration sweep, 98.5% agreement with the reference benchmark's own correctness code, and a stratified hand-audit. Aimed inward, the verifier judges a kernel of our own: the first native Blackwell tcgen05 training backward for the gated-linear-recurrence (GDN) family, including the reverse-state stage the field still runs on a fallback. We establish its correctness independently, against a double-precision oracle, and train five family members through it. The correctness signal behind reported progress in kernel generation is far weaker than the numbers suggest, and a set of tolerance-free contracts would close most of the gap.

AI-powered search products such as ChatGPT search, Google's AI Overviews, and Perplexity provide LLM-synthesized answers grounded in live web results. We developed OreoLook (formerly lixSearch), an open-source answer engine using automated browser agents and provider-routed LLM inference. Its local search, caching, session-management, and embedding stack runs on commodity CPU hardware; answer synthesis is performed by a remote inference provider. As usage grew, sessions lost context, equivalent queries triggered redundant work, and URLs were repeatedly embedded across sessions. We present a three-layer caching architecture: (1) a Session Context Window maintaining a rolling window of recent messages in Redis with automatic overflow to Huffman-compressed disk archives; (2) a Semantic Query Cache catches rephrasings via cosine similarity on embedding vectors, eliminating redundant LLM invocations; and (3) a URL Embedding Cache that deduplicates embedding computations across sessions. Deployed on a single 8-vCPU Intel Cascade Lake server (2 GHz, 32 GB RAM) running 30 Hypercorn worker processes across three containerized replicas, the evaluated system reported an 89.3% aggregate Redis keyspace hit rate with 0.1 ms read latency and just 1.38 MB of memory overhead. A background LRU eviction daemon migrates idle sessions from Redis to disk and re-hydrates them on demand, enabling conversations that can be resumed hours or days later under the configured retention policy.

Serverless platforms commonly colocate many diverse workloads, each in a fast-booting, memory-lean virtual machine (VM), to improve deployment density. Overprovisioning each VM for its peak protects tail latency during traffic bursts but hurts density; maintaining high density while effectively protecting tail latency requires the infrastructure to be able to shift physical cores, at a microsecond timescale, to whichever latency-sensitive VM is bursting and reclaim them as the burst subsides. No VM substrate delivers this: conventional VMs resize a guest's cores only through a millisecond-scale vCPU hot-plug path, Firecracker fixes a VM's core count at boot, and the ultralight VMs that boot fastest drop multicore execution entirely. We present HyperFlux, a commodity-KVM ultralight VM substrate that makes a VM's parallelism width (the number of physical cores backing it) elastic at runtime. We show that HyperFlux can move a core across VMs in merely 13$\mu$s, even when forcibly reclaiming it from a busy donor, orders of magnitude faster than vCPU hot-plug. A HyperFlux VM incurs only a 3.2MB memory footprint and can cold-boot in 1.37ms, on par with the fastest-booting ultralight VMs, while uniquely supporting multicore parallelism. Under colocation, it can reduce high-priority VMs' tail latency by up to 10x under high load compared to static core-sharing with Firecracker and Cloud Hypervisor, and deliver a lower and more stable tail latency compared to using cgroup and vCPU hot-plug under changing load bursts.

Training Mixture-of-Experts (MoE) models for reinforcement learning (RL) couples two load-balancing problems: sequence composition determines dense attention work in each data-parallel microbatch, while token routing determines sparse expert work on expert-parallel ranks. Optimizing either alone can shift the bottleneck to the other. In MoE RL, rollout-time routing replay exposes every sample's sequence length and layer-wise expert demand before its training step. We present RoutePack, a hierarchical planner that coordinates state-consistent, layer-wise expert rerouting with joint attention- and expert-aware data packing over an optimizer-step window. RoutePack first places experts independently at each MoE layer using aggregate routing demand. It then packs samples into the smallest certified, or best-known feasible, number of token-capped execution rows and optimizes their DP layout with a projected EDP-shard-aware objective. The objective combines a window-normalized linear-quadratic attention proxy with per-layer physical EP-rank peaks and minimizes the accumulated cost of the slowest EDP shard. Parallel population annealing searches fixed-row feasible layouts while preserving sample coverage, capacity, nonempty cells, equal microbatch counts, and communicator topology. State-consistent materialization preserves logical top-k routing and existing MoE kernels without microbatch-level expert replication. Across Ling-3.0-Tiny and Ling-3.0-Flash, expert rerouting improves mean trainer-measured token throughput by 3.80% and 10.50%, while routing-aware packing adds another 4.86% and 3.98%, respectively. Overall, RoutePack improves throughput by 8.85% and 14.89% over the baseline.

We introduce a mechanism that improves LLM-agent execution: keeping four-byte, GPU-computed control-path routing decisions on-device, avoiding a round-trip to host memory for redispatch. It is faster than host-dispatch in all 36 placements x mechanisms settings in our benchmark across four named GPUs with row-median speedup of 1.19x-2.39x, and outputs are correct, matching an independent host implementation, for all 14,557,440 calls in all configurations, in both tested mechanisms. Agent control paths can be deterministic. Their transitions between LLM calls and external tool calls can be GPU-accelerated by batching across agents and time. We define a cohort of transitions with launch times in a window, and study four shares of GPU-active time: a fixed-partition baseline F, an optimal P*, an upper bound U, and an online result A. Under assumptions of uniform relative launch deadlines, infinite capacity, and zero service times, we compute P* with a custom dynamic program and validate our simulator against it. We replay an anonymized 851-session trace, shown in full as a panel, stationary-ized with a Poisson process. With 100,000 target active agents and a launch deadline of 50 ms, where the headline results have K=256, we find F=30.19%, P*=43.00%, U=45.85%; our method for computing P* reveals that 81.83% of this opportunity is recoverable from cohort packing at fixed window ends. Our method for U relies on a novel notion of route-key, used only as a proxy (via conditional statistics) and does not purport to establish route-equivalence. We find a nested graph that retains some host-side decisions slower in all 60 configurations (five placements).

Currently, most supercomputers are equipped with GPUs from manufacturers such as NVIDIA, AMD, or Intel, which provide substantial parallelism and high throughput. It is common for a single compute node (intranode) to host multiple GPUs, typically four or more. Therefore, effectively leveraging all these GPUs within a single compute node is essential for applications in scientific and engineering domains. However, several factors must be considered before utilizing these GPUs for scientific computing, including the implementation of data communication, the programming models available for use across these GPUs, and the level of performance that can be achieved with a single codebase across different GPU architectures and configurations within a single compute node. OpenMP Offloading is a prominent directive-based programming model that can be executed on all three GPU types: NVIDIA, AMD, and Intel. In this research, we present an analysis of the benefits and performance challenges of using OpenMP Offloading to address the 3D heat equations, which involve both primary computation, as well as halo computation and communication. For additional comparison and scalability study, we also consider the Conjugate Gradient method. We investigate how performance varies in relation to native GPU programming models-CUDA for NVIDIA, HIP for AMD, and SYCL for Intel. Furthermore, we demonstrate that OpenMP Offloading can achieve performance improvements of approximately 2x for 2 GPUs and around 4x for 4 GPUs when compared to single-GPU OpenMP Offloading implementations across all three GPU types. This analysis is conducted systematically through various OpenMP Offloading implementations that utilize different low-level APIs for memory allocation, memory transfer options (synchronous, asynchronous, and peer-to-peer), and other native GPU programming models such as CUDA(NVIDIA),HIP(AMD),and SYCL(Intel).

Growing demand for artificial intelligence (AI) inference services requires scalable infrastructure, yet centralized serving costs rise with demand. We propose a collaborative distributed inference system combining dedicated infrastructure with resources contributed by service users. Dedicated resources provide baseline capacity for maintaining quality of service (QoS), while volunteered resources absorb increasing demand without proportional growth in centralized infrastructure. To capture stochastic and dynamic interactions among users, resources, tasks, and policies, we develop a high-dimensional generative Markov model with structured temporal factorization. The model supports simulation and provides a foundation for task scheduling and QoS-aware resource allocation optimization. We evaluate the system across user populations, resource capacities, and centralized and distributed scheduling policies. Simulations show that distributed scheduling becomes increasingly advantageous as the user population grows, improving request completion and P99 latency while substantially reducing dedicated resource consumption. These results demonstrate the feasibility of user-assisted collaborative inference for infrastructure-efficient autoscaling.

Elasticity is commonly presented as the default response to capacity loss after failures, since replacement replicas can compensate for failed nodes and restore pre-incident service levels. Replacement capacity entails both delay and additional resource commitment, as replicas must be provisioned and synchronized before they can serve traffic. Under fixed budgets or constrained operating conditions, capacity restoration cannot be treated as the immediate recovery path. Failure handling must instead define how the service continues while capacity remains reduced. When the service exposes differentiated service levels, capacity loss cannot be handled uniformly. Degradation becomes part of the service behavior, requiring explicit control over how reduced capacity affects each class without erasing the intended differentiation. We study differentiated QoS degradation in replicated database services with service-class-aware sessions. We present a repair-to-target policy, implemented in PLB, a PostgreSQL JDBC middleware load balancer for service-class-aware routing. When fail-stop failures remove part of the available capacity, PLB updates the role assignment of healthy replicas into Premium, Mixed, and Freemium roles. This keeps the replica pool shared while ensuring that new session assignments continue to reflect the service class. We evaluate PLB under single and cascading replica failures across two deployment strategies: isolated perclass replica pools and shared, priority-agnostic routing. The results show that PLB improves median Premium goodput retention by 26-28 percentage points under a Premium-side fault, achieves more than 2x higher Premium goodput in the most severe cascading-failure phase, and reduces Premium p95 latency by 18.2% relative to shared round-robin.

Edge-deployed vision systems in target recognition, surveillance, autonomous vehicles, and drone domains require hierarchical inference pipelines where a detection model identifies objects of interest and downstream classifiers provide fine-grained attribute analysis. Running all models on the GPU creates a serial bottleneck that limits real-time throughput as pipeline stages grow. Modern edge SoCs pair GPUs with dedicated neural accelerators (NPUs, DLAs) capable of concurrent execution, yet deploying custom models on these accelerators remains impractical due to strict operator constraints, quantization incompatibilities, and an undocumented end-to-end pipeline. We target NVIDIA Jetson DLA cores as the representative platform. We present a five-step methodology for zero GPU fallback DLA INT8 deployment of classification backbones, comprising architecture adaptation, manual dynamic range workaround to rescue TensorRT's implicit quantization (recovering 94.0% accuracy from implicit quantization's 75%) for rapid pipeline validation before explicit quantization, quantization-aware training, ONNX graph surgery for DLA compilation, and a concurrent GPU-detection/DLA-classification inference pipeline. We document nine engineering constraints with root-cause analysis and generalizable solutions. Validation on a dual-head person attribute classifier running on DLA alongside a GPU object detector on a Jetson Orin NX demonstrates near-zero pipeline overhead (12.5 vs. 13.3~FPS detector-only at 1080p), with dual-DLA scaling at no additional cost. The methodology is backbone-agnostic and generalizes to any detection-classification edge pipeline.

Agents powered by AI/ML are becoming ingrained in orchestration. Dispatch of work is the task of receiving a request, transforming it for a workload manager, and successfully submitting it. Running scientific workflows across multi-cluster environments introduces substantial challenges of dynamic job transformation, dispatch, and submission to heterogeneous clusters. These tasks are well-suited to agents, which can receive textual instructions for work, prepare job specifications, and dispatch. In this work, we assess the reliability of a dispatch agent across 432 runs, testing all possible combinations of five feature dimensions across four prompt styles. The agent is highly reliable (97.9% success). We test a full orchestration to submit, queue, match, score, select, transform, and dispatch in a multi-cluster experiment. We find that descriptive metadata increases successful execution from 48% to 87% of 220 submitted jobs, eliminating architecture mismatch, and improving performance for five of ten measurable applications by up to 3.3x.

Leader-based consensus datastores (etcd, ZooKeeper) face two competing routing goals: spread load evenly across members, and route operations to the member whose protocol role matches the operation. Writes must commit through the leader, so sending them elsewhere adds a forwarding hop. Linearizable reads need only a lightweight leader confirmation before any member can serve them locally. The upstream etcd client uses gRPC's round_robin balancer, distributing reads and writes uniformly across cluster members. An operation-aware client resolves this by pinning writes to the leader and distributing reads across the healthy read pool. In steady state on a 3-node etcd cluster (80/20 read/write mix, 5 trials), this lowers write P50 by 29% and raises throughput by 9%. When a follower degrades silently, the operation-aware client detects the latency shift and removes it from the read pool, cutting read P99 by 64%, write P99 by 74%, and raising throughput by 89%. The same routing rule applied to ZooKeeper (ZAB protocol, different implementation) points in the same direction, showing that the result follows from leader-based consensus structure rather than one system's implementation. The key obstacle to discovering this policy adaptively is that the leader confirmation round-trip occurs between cluster members, so the client sees only a blended latency signal rather than the decisive coordination cost directly.

Root cause analysis (RCA) in large datacenter networks is challenging because telemetry is noisy, partial, and asynchronous. Score-based approaches degrade under these conditions, often yielding unstable or incorrect attributions. We present CoreSec, a production RCA system that replaces weighted fusion with a PAM-style abstention algebra. Telemetry agents are composed using control flags that yield deterministic decisions and explicit abstention when evidence is ambiguous. CoreSec combines this algebra with topology-aware configurations that capture failure surfaces across Clos fabrics and converge monotonically as evidence accumulates. Deployed at hyperscale, CoreSec provides stable and explainable RCA behavior across diverse environments without retuning. Our experience shows that structured composition with abstention forms a practical foundation for automated RCA in real-world cloud networks.

Energy-aware LLM serving requires comparing configurations under realistic request shapes, yet exhaustive target-GPU profiling is costly and a cheap predictor can be dangerously confident outside its measured scope. We present TokenPowerSandbox, an evidence-gated workflow that combines an interpretable CPU-resident projector, short target-GPU probes, full-workload verification, and tamper-evident freeze-before-measurement provenance. On one NVIDIA H100 80GB serving Qwen2.5-7B-Instruct with vLLM, three anchor repeats and six development workloads calibrate workload transfer. The same frozen model is evaluated on a blind holdout and a separately predeclared no-refit confirmation totaling 51 post-freeze runs. Energy MAPE is 6.23% and 7.35%, with Spearman rank correlations of 0.976 and 0.933. However, a predeclared TTFT gate passes at concurrency four (9.27% MAPE) and triggers abstention below four (64.80%), showing why energy accuracy cannot certify latency.

Portability in high-performance quantum-circuit simulation need not begin at the kernel. We present QSimAdv, which makes late binding, rather than a common kernel, the basis of vendor independence. Representation, operator lowering, and data placement are bound only when their required inputs become available. Before full-state allocation, circuit, noise, and output inspection can route eligible generic sampled-count requests to a stabiliser tableau; explicitly requested representations remain fixed. For full-state execution, backend constraints shape fusion; an ordered fused operator binds to a native lowering only after its physical targets are known. A first-class logical-to-physical layout map records non-canonical order across local and rank-address bits, so the dispatcher moves nonlocal targets only on demand. GPU, CPU, and Message Passing Interface (MPI) backends share these semantics while retaining native execution paths. We realize this design on NVIDIA GH200 and AMD MI250X/EPYC systems across local and distributed execution. With matched complex 32-bit floating-point state storage, QSimAdv leads both Aer Hopper configurations at $N=32$ and Aer's HIP backend at four shared MI250X sizes from $N=24$ to 30. Strong scaling exposes platform dependence: on setonix, QSimAdv leads both GPU and CPU comparisons at every measured rank, achieving $3.4\times$ and $2.8\times$ speedups, respectively, from one to eight ranks; neither the GH200 path nor the CPU path speeds up at eight ranks. Weak scaling reaches 256 ranks with 2 TiB GPU and 1 TiB CPU states. Together, these results support that portability can reside above the kernel boundary while execution remains native and extends across distributed memory.

Continuous physiological monitoring using consumer-grade wearables offers a transformative opportunity for clinical care and research, yet integration remains hindered by device heterogeneity, proprietary data formats, and strict regulatory requirements. We present an event-driven, cloud-native system designed to ingest, normalize, and analyze high-frequency vital signs from wearables at scale and without vendor lock-in. The system design proposes a multi-layered microservice architecture using cluster orchestration. Data acquisition is handled via a cross-platform mobile application that leverages native health frameworks, ensuring compatibility across fragmented device ecosystems. To address interoperability, we implement an event-driven transformation pipeline using stream processing engines and specialized services to map raw measurements to the FHIR standard for medical interoperability. Our novel dependency-aware FHIR minimization scheme reduces storage overhead while maintaining lossless resource reconstruction. Furthermore, the platform integrates a modular data analytics and machine learning layer based on a medallion lakehouse architecture, supporting the full machine learning lifecycle from real-time stream processing to model serving. Performance evaluation demonstrates that the ingestion pipeline sustains 50 full ingestion requests per second with median response times under 8 ms, satisfying the low-latency requirements for real-time patient monitoring. Our open-source implementation adheres to regulatory compliance standards through role-based access control and secure service-to-service communication, providing a robust foundation for deploying wearable-based monitoring in institutional healthcare settings for clinical decision support and research workloads.

Modern reinforcement learning (RL) post-training pipelines for large language models (LLMs) increasingly combine rollout workloads across multiple domains and feedback paradigms. Prefix-aware routing improves inference efficiency through cache reuse and load balancing, but it does not control how heterogeneous rollout sessions compete for KV-cache capacity. When reinforcement learning with verifiable rewards (RLVR), reinforcement learning from human feedback (RLHF), and agentic rollouts share an asynchronous inference service, their distinct sequence structures, interaction patterns, and KV-residency times create substantially different serving demands. Rollout scheduling must account for this heterogeneity without distorting the workload mixture specified by the trainer. We present MISA-T, a routing-layer admission policy for mixed rollout serving. MISA-T combines adaptive session admission, workload-aware KV-capacity allocation, and residency-time-aware KV accounting. In rollout-only ablations on Step3.7 and Qwen3.6-35B-A3B, MISA-T improves rollout throughput over a sweep-tuned cache-aware vLLM Router by 53.3% and 43.6%, respectively, while maintaining high prefix-cache hit rates. In a matched 50-iteration Step3.7 experiment, it increases rollout throughput by 35.6% and reduces mean iteration time by 22.8%, while keeping the consumed workload mixture close to the trainer target and achieving comparable task scores.

In LLM pre-training, synchronization propagates rank-local stalls, slowdowns, and numerical errors into job-wide symptoms, obscuring their origin. Existing diagnosis often relies on in-process monitors that cannot report after the trainer blocks or terminates, or on post-mortem logs that preserve only synchronized symptoms; offline health tests lose the workload and operating conditions that triggered the failure. We present SCOUT, a unified runtime failure-localization framework built on one design principle: identify outliers through strict-majority consensus among equivalent replicas. SCOUT aligns replica progress, timing, and numerical evidence, then uses its Consensus Collective Communication (C3) abstraction to identify ranks whose compact signatures disagree with their peers. An out-of-band CPU observer remains responsive when training hangs, whereas in-situ replay exercises recurring stragglers and silent data corruption (SDC) beside the live job with its model state, kernels, allocations, communication path, and thermal and memory pressure present. Collective fingerprints expose rank-local protocol divergence. Clean replay coverage certifies checkpoint numerical integrity, preventing recovery from selecting state corrupted by SDC. SCOUT integrates with PyTorch, TorchTitan, Megatron-Core, and DeepSpeed without training-loop or framework-source modifications. SCOUT is open source at https://github.com/LMResiliency/lm-resiliency.

Model Cards and Data Cards have demonstrated the value of structured, human-readable documentation for machine learning artifacts, capturing their context, parameters, limitations, and intended use. However, these practices remain focused on static artifacts (the datasets and trained models themselves) while overlooking the workflow executions that produce, transform, and evaluate them. Such executions hold critical details about data preparation, parameter choice, runtime behavior, resource use, and intermediate transformations, precisely where bias, performance variation, and reproducibility gaps tend to originate. To close this gap, we introduce Workflow Cards: structured summaries that condense the machine-readable provenance data of a workflow execution into a form both humans and large language models (LLMs) can read and analyze. This paper has two main parts. First, it defines a Workflow Card template informed by a representative set of provenance questions that surface from the execution-level data missing from Model and Data Cards. Second, it evaluates how effectively LLMs use Workflow Cards to understand workflow executions compared with querying provenance databases through a schema-based interface. Results show that Workflow Cards provide execution-level information absent from existing card types, such as Model Cards and Data Cards, thereby filling an important documentation gap; and that Workflow Cards nearly double answer quality compared with schema-based querying, consistently across LLM-as-a-Judge and human assessments.