Proceedings of the ACM on Measurement and Analysis of Computing Systems

Robot teleoperation with extended reality (XR teleoperation) enables intuitive interaction by mapping user motions to remote robots with real-time 3D feedback. However, existing systems suffer from large completion delays and trajectory deviations under prolonged network latency, rooted in their exclusive reliance on network communication and strict synchronous execution architecture. Moreover, network fluctuations destabilize teleoperation accuracy, while dynamic user motions amplify teleoperation errors. We present MATER, an end-to-end XR teleoperation framework that introduces a mutually-aware architecture in which each side reconstructs its counterpart's delayed or missing state to decouple the execution from network dependency. MATER includes latency-adaptive input window and user motion gap interpolation techniques to handle unstable network communication. It also proposes motion-driven robot state rollback and robot trajectory coordination to handle complex motions. The key idea behind these techniques is to adapt on the fly by reshaping reconstruction and filling gaps as network conditions fluctuate, and by realigning states when motions become fast or complex. Together with lightweight local synchronization and bandwidth optimizations, these system-level advances make MATER resilient to both network and motion dynamics. We implement MATER across three hardware settings, including simulated and physical robots, and evaluate it on 9,500 real-world teleoperation trials from the RoboSet dataset, covering single- and multi-step missions. Compared to state-of-the-art XR teleoperation frameworks, MATER reduces teleoperation error by up to 69.8% on WLAN and 73.1% on cellular networks with only 6.7% maximum runtime overhead. It also shortens mission completion time by up to 47.7%, enabling smoother teleoperation. A real-world case study on ten stationary and mobile missions further shows MATER achieves up to 37.7% faster completion while lowering average teleoperation error by up to 57.2%. MATER code is available at: https://github.com/rtenlab/mater

This paper studies an online resource allocation problem, where resources are reusable, allocated in combinatorial bundles, and subject to endogenous performance deterioration. Unlike existing models that assume singleton decisions or stationary or exogenous dynamics, we consider a combinatorial decision structure where decisions involve selecting among resource-bundled options, while capturing a self-induced feedback loop in which system performance, like rewards, activation probabilities, and service times, degrade as a function of cumulative resource usage. We propose a unified framework that subsumes a broad class of existing models as special cases, and develop a ranking-like online policy, termed IPRANK, which employs an indicator-perturbation mechanism to induce a randomized ranking over configurations based on their size-penalized rewards. Our analysis is based on a queue-coupling argument and shows that under adversarial arrivals, the greedy policy inevitably suffers a competitive ratio with linear dependence on reward heterogeneity, whereas IPRANK achieves a competitive ratio with only logarithmic dependence, making it far more robust in heterogeneous environments.

The Proceedings of the ACM on Measurement and Analysis of Computing Systems (POMACS) focuses on the measurement and performance evaluation of computer systems and operates in close collaboration with the ACM Special Interest Group SIGMETRICS. All papers in this issue of POMACS will be presented at the ACM SIGMETRICS 2026 conference on June 8-12, 2026 in Ann Arbor, Michigan, USA.

We study the idea of information design for inducing prosocial behavior. As an example, we ground our study in the context of electricity consumption. Electricity utility providers would like to reduce the total power consumed by their residential consumer segment. Supply to this segment is often subsidized, and the saved power can be diverted to more profitable segments. Alternatively, the provider may be keen on earning carbon credits by inducing reduced consumption in this segment. The societal network in which the consumers reside may have a prevailing norm, for example, saving power is environment friendly and is considered good. Those that consume less are considered more prosocial and may derive a larger reputational benefit. How can the service provider, who is familiar with the prevailing norm and the consumption of all users, design suitable feedback signals that exploit reputation benefits to reduce net consumption? We call this a problem of information design and address this question in this paper. We consider a continuum of agents. Each agent has a different intrinsic motivation to reduce her power consumption. Each agent models the power consumption of the others via a distribution. Using this distribution, agents will anticipate their reputational benefit and choose a power consumption by trading off their own intrinsic motivation to do a prosocial action, the cost of this prosocial action and their reputation. We assume that the service provider can provide two types of quantized feedbacks of the power consumption. We study their advantages and disadvantages. For each feedback, we characterize the corresponding mean field equilibrium, using a fixed point equation. Besides computing the mean field equilibrium, we highlight that, in some situations, revealing less information can lead to more prosociality. We illustrate the result on London smart-meter data. We also introduce the notion of privacy and provide a new quantized feedback respecting agents' privacy concerns yet improving prosociality. The results of this study are not restricted to the framework of energy efficiency but are also applicable to congestion problems in road traffic and other resource sharing problems.

Continuously updated corpora make freshness a first-class systems objective, yet periodic TTL refresh introduces unavoidable stale windows. We formalize a staleness-latency-cost tradeoff and show that any periodic refresh policy has violation rate at least p dep (τ)/2, where p dep (τ) is the probability that a query depends on an update committed within the TTL window τ. We present EviDex, a provenance-weighted evidence-path index that continuously compacts a commit stream into intent-partitioned path buckets and supports auditable multi-hop retrieval with bounded visibility delay. On clinical and Wikipedia update streams, EviDex reduces evidence-set violation to 1.3% and 1.1% at T=15 minutes, versus 2.4% and 2.1% for the strongest streaming-aware baseline, while reducing clinical cost by 42%. In a physician-rated clinical QA study, EviDex reaches 0.884 correctness with 0.84% macro-SCER.

Virtual memory is essential for modern GPUs, yet address translation remains a major bottleneck. In irregular workloads, frequent TLB misses and costly page table walks (PTWs) dominate memory latency. We propose C ompute U nit P age T able W alk (cuPTW), which repurposes idle GPU compute resources to accelerate PTWs. By offloading requests to idle functional units, cuPTW transforms address translation into a massively parallel task. We further optimize this with cuPTW-SW, which caches walks in local data store (LDS) memory, and cuPTW-MT, which parallelizes walks across SIMD lanes. Evaluation shows that cuPTW-Full achieves a 4.43× average speedup (up to 76.09×) and improves PTW throughput by 9.92×. Compared to state-of-the-art designs, Marching Page Walks and SnakeByte, cuPTW-Full delivers a 2.08× and 1.97× performance gain, respectively.

Traditional tools for verifying network configurations: (1) require high manual effort or lack semantic depth, and (2) lack interpretable, human-readable explanations to support diagnosis. Large Language Models (LLMs) present a promising alternative, but simple approaches like full-file, partition-based, or chain-of-thought prompting fail to handle the large, interconnected nature of configuration files. We introduce the Context-Aware Prompting (CAP) framework, which enables an LLM to reason about network configurations like a human expert. CAP first identifies all potentially relevant neighboring, similar, and referenced configuration segments. CAP then engages the LLM in a structured dialogue, where the model requests the specific context it needs before performing a focused analysis. This on-demand approach provides the necessary context for deep reasoning while avoiding overload. Our evaluation on production network configurations demonstrates that CAP reliably detects a wide range of known and previously unknown configuration errors. CAP outperforms existing LLM-based baselines, and achieves performance comparable to state-of-the-art non-LLM tools.

Sparsity is crucial for deep neural networks, reducing storage and bandwidth while improving computational efficiency. With the rapid development of deep neural networks, sparse computation accelerators have become a key research focus. NVIDIA introduced 2:4 fine-grained structured sparsity in the Ampere architecture, doubling Tensor Core peak throughput by retaining two out of four elements, but this may compromise model accuracy. In contrast, unstructured sparsity achieves maximum sparsity without accuracy loss. However, the irregular distribution of non-zero elements (NNZs) in unstructured sparse matrices causes complex memory access and load imbalance, preventing direct use of GPU Tensor Cores. Thus, computation relies on CUDA cores via sparse libraries like cuSPARSE, underutilizing Tensor Cores' capabilities. To address this, we extend the Tensor Core architecture based on Gustavson's dataflow and propose a re-structured row offline compression format for inference. This format integrates with the extended architecture to enable efficient unstructured sparse-dense matrix multiplication. Experiments show negligible area and power overheads, with an average speedup of 3.54x over cuSPARSE and 1.2x over state-of-the-art Tensor Core extensions.

Multipath TCP (MPTCP) enables a transport connection to utilize multiple network paths simultaneously and is expected to achieve higher throughput than single-path TCP. To ensure compatibility with the existing Internet, MPTCP congestion control follows three classical design principles: (P1) an MPTCP flow should perform at least as well as TCP on its best path, (P2) it should remain fair to TCP flows, and (P3) it should pool resources across paths. However, extensive experiments consistently show that MPTCP often delivers only modest throughput improvement over best-path TCP. In this paper, we show that this phenomenon reflects a structural limitation imposed by the joint enforcement of P1–P3 under end-to-end transport. Our analysis characterizes the achievable throughput envelope of MPTCP and proves that, even under favorable conditions, the gain over best-path TCP is fundamentally bounded. Using a network utility maximization (NUM) framework, we further show that relaxing P1 enlarges the feasible allocation region and admits significantly higher-throughput operating points. Under a restricted symmetric setting, the achievable gain approaches 4-2√2 ≈ 17.2% with logarithmic utilities. Guided by this insight, we derive a NUM-optimal congestion control family, N-MPTCP(ω → ), and design a practical instance with adaptive weights that realizes these gains in packet-level simulations.

Personalized decentralized federated learning (PDFL) aims to learn client-specific models from heterogeneous data without a central coordinator. However, push-based gossip over large peer graphs is graph-oblivious: it mixes semantically mismatched updates, attenuates minority signals, and assumes broad peer reachability. We present SemanticDFL, a fully decentralized, pull-based PDFL framework built on a hierarchical semantic overlay network (SON). Each client exports a compact top- P model signature; bounded peer discovery forms zones; affinity propagation clusters similar peers; and replica-backed super-peers route lightweight bounded-fanout similarity queries. Each client then pulls only its top- K most similar peers for personalized aggregation. Experiments on multiple benchmarks with dozens to hundreds of peers on the EU SLICES testbed show that SemanticDFL improves final accuracy, reaches target accuracy in substantially fewer rounds than FedAvg and the strongest PDFL baseline, while incurring only modest per-round overhead.

Starlink has deployed over 7,800 satellites serving millions of subscribers, yet predicting its performance remains an open challenge. Rapid orbital dynamics, frequent handovers, and weather-induced signal attenuation create variability that existing models, built on a handful of instrumented terminals in limited regions, cannot capture at global scale. We present Horizon, the first global-scale machine learning system for predicting LEO satellite Internet performance. Our key insight is that crowdsourced measurement platforms, while noisier than controlled experiments, provide the geographic diversity necessary to build globally generalizable models. Horizon integrates 11 months of measurements from M-Lab and Cloudflare spanning 90+ countries with meteorological data and satellite orbital propagation features. On a fully held-out one-week temporal window, Horizon achieves mean absolute errors of 17.76 ms for latency and 25.63 Mbps for throughput; on a standard 80/20 split it outperforms all baselines, including adaptations of state-of-the-art architectures. Feature importance analysis reveals that geographic position dominates prediction, with latitude alone contributing 42-46%, while weather features account for 14-15%, quantifying the impact of atmospheric conditions on Ku/Ka-band links. Leave-one-location-out experiments confirm that Horizon generalizes to regions absent from training, enabling performance estimation where measurement infrastructure does not yet exist. Our dataset and pipeline are publicly available, providing a foundation for global LEO network performance visibility.

We study sequential allocation of a single reusable resource to stochastically arriving tasks in continuous time, where rewards and service durations are random with unknown statistics. When the resource is idle, the controller observes the arriving task's type and duration and must immediately accept or reject it; while the resource is busy, new arrivals are unobserved, which tightly couples learning and control. We first show that no non-anticipatory online policy can guarantee more than a 1/2 fraction of the full-information offline prophet benchmark, motivating the use of 1/2-approximate regret. To approach this limit, we derive an infinite-dimensional steady-state linear program that upper bounds the offline optimum and reveals a threshold structure: a task should be accepted when its profitability, defined as reward divided by duration plus expected waiting time, exceeds a critical threshold. Guided by this characterization, we develop online threshold-learning algorithms that achieve O (√ T ) 1/2-approximate regret when rewards are known and O (T 2/3 ) when reward functions are unknown. We further show that if durations are unobserved upon arrival, meaningful guarantees become impossible without additional flexibility, but can be recovered via task termination.

In designing an effective GPU for large-scale workloads, a trustworthy and fast simulator is required to evaluate performance and explore the design space. However, existing GPU simulators suffer from long execution times due to detailed component simulation, limiting their utility for evaluating the effects of architectural modifications. It is necessary to improve the performance of a GPU simulator such that quick architecture exploration and evaluation for large-scale workloads are available, at the expense of accuracy. This paper presents LPGSim, a trace-driven and cycle-level GPU simulator. LPGSim aims to provide fast and accurate GPU simulation. To this end, LPGSim first eliminates instruction metadata that has minimal impact on simulation accuracy. Next, LPGSim parallelizes GPU simulation. LPGSim partitions a GPU architecture into three parallelizable subsystems and introduces local-clock-based parallelization. LPGSim further employs parallelization methods to achieve scalability on NUMA systems with an acceptable trade-off in accuracy. LPGSim shows 21.4%, 23.3%, and 22.7% errors across three different GPU architectures, while achieving a total 197.4x speedup over the state-of-the-art simulator.

Enabling autonomous CubeSats requires shifting data processing from ground stations directly to on-board hardware to bypass severe downlink bottlenecks, using a paradigm known as Orbital Edge Computing (OEC). A critical, yet overlooked, OEC prerequisite is Image Registration (IR), whose pipelines are usually developed for unconstrained platforms, failing to meet strict on-board latency, accuracy, and energy trade-offs. For this reason, we introduce STARBench , a modular framework to standardize the evaluation of IR methods. Leveraging this framework, we proposed the Quality-Adjusted Cost (QAC) as a new system-level metric that quantifies the latency-accuracy-energy trade-off and reveals a fundamental ''specialization trade-off.'' Based on these insights we propose ETNA, a mission-aware solution that dynamically orchestrates heterogeneous IR pipelines. Through a hardware-software co-designed pyramidal Field Programmable Gate Array (FPGA) accelerator, ETNA achieves real-time multimodal registration, improving the QAC by up to 12.03× over Pareto-optimal baselines.

Persistent memory has emerged as a groundbreaking solution for byte-addressable storage-class memory (SCM). While persistent memory is highly desirable due to its advantages in high capacity, performance, and cost-effectiveness, novel system architecture solutions are required for its adoption in data center applications that demand high performance and reliability. The major challenges are issues related to persistent memory media, such as endurance, retention, and reliability, along with hardware and software system integration within the data center platform. To reduce the total service cost for end users by utilizing persistent memory media as the main memory, we have developed a comprehensive system comprising persistent memory chips, a persistent memory controller, firmware, and system software layers that effectively mitigate persistent memory issues and are compatible with data center applications. In this paper, we discuss the system design of a high-performance byte-addressable persistent memory with a Compute Express Link (CXL) interface. This design choice was based on real-world persistent memory limitations. It was implemented on Application-Specific Integrated Circuit (ASIC) platform, and tested on in-memory database (IMDB) applications. Bandwidth and latency results show that our design approach for CXL persistent memory offers comparable performance to CXL dynamic random access memory (DRAM) under certain test cases while providing increased memory capacity and reduced cost. Our main contributions are summarized as follows: (1) We have conducted a systematic study on persistent memory media issues and have proposed a hardware and software architecture that ensures reliable and efficient utilization of persistent memory media. Media management algorithms for endurance, disturbance, retention, and drifting have been developed and validated. (2) We have pioneered the first SCM-based data-center CXL memory. By utilizing the CXL interface, the barrier to adapting persistent memory to server platforms in data centers has been largely reduced, as it uses Peripheral Component Interconnect Express (PCIe) interconnect slots instead of a limited number of DIMM slots. Our persistent memory solution with the CXL interface offers enhanced compatibility across various platforms and unlocks the potential for memory pooling. (3) We have developed a software path that integrates persistent memory into our server systems, spanning from the application layer to the driver and firmware layers. To evaluate its performance, we have conducted benchmarking tests using in-memory databases as the application and compared the results with those of CXL DRAM devices and Intel's Optane product.

Prefix caching is a key performance optimization in Large Language Model (LLM) serving systems, enabling reuse of attention Key-Value (KV) states across requests with shared prompt prefixes. However, the size of GPU memory limits cache capacity, making the eviction policy a critical factor in overall system performance. Existing systems primarily rely on simple heuristics, such as LRU, and apply the same policy across task categories, implicitly assuming homogeneous workloads. In practice, however, modern LLMs serve heterogeneous workloads that mix multi-turn conversational traffic with diverse single-turn API requests, leading to fundamentally different prefix reuse patterns. In this work, we first design a trace-driven prefix cache simulator built on vLLM to systematically characterize prefix reuse across representative workloads. Our analysis reveals two dominant reuse patterns---session reuse and structural reuse---that vary significantly across task types. Motivated by these observations, we propose \sys, a unified eviction policy that jointly captures both reuse patterns and dynamically balances cache allocation across tasks. When implemented in vLLM, UniCache achieves substantial improvements under heterogeneous workloads, improving prefix cache hit rates by up to 17.32% and reducing inference latency by up to 3.63× compared to existing policies.

We study the Strategic Information Game (SIG), a distributed learning model over a fixed graph in which useful transfer on an edge requires bilateral commitment. Our central result is a sharp degree-based law: for every connected graph, the Price of Strategic Information (PoSI) satisfies PoSI( G ) ≤ 1 + d max , where d max = max i ∈ V d i is the maximum degree, and in the negligible-cost regime this becomes exact for regular graphs. We also prove equilibrium existence, an exact harmonic-mean formula for arbitrary graphs when communication and competition costs vanish, polynomial-time topology design under edge budgets, no-regret convergence to coarse correlated equilibria, and an impossibility theorem showing that transfer-free reciprocal protocols cannot eliminate the all-zero equilibrium.

This work proposes overlapping Random Number Generation (RNG), the main runtime contributor of Dropout, with preceding GEneral Matrix Multiplication (GEMM) layers to hide RNG latency during LLM training. The state-of-the-art optimization is to fuse Dropout into the Flash-Attention kernel; however, evaluating fine-grained architecture resource constraints beyond traditional compute or memory metrics reveals that fusion fails to fully hide RNG latency due to shared lower-level architecture bottlenecks. RNG and GEMM have distinct hardware bottlenecks, so they can run together without compromising each other's performance. Our analytical model, validated on GH100 GPUs, shows 1.26× speedup over sequential execution and 1.22× over state-of-the-art fusion on Llama3 for a single Transformer block. Our methodology generalizes to fusion-versus-overlap decisions for new LLM operators across various architectures, models, and configurations.

SIEVE and CLOCK have attracted significant interest as cache replacement algorithms due to their simplicity in implementation and, in the case of CLOCK, frugality in memory usage. Unfortunately, they are susceptible to pathologies such as performance cliffs, wherein increasing cache capacity has little effect over a range of sizes until a workload-dependent threshold is crossed, after which hit rate improves abruptly and disproportionately. Performance cliffs arise precisely because scan-based eviction paths create a synchronized, deterministic aging frontier in which many objects are admitted and removed in lockstep, causing catastrophic performance degradation under scan-like workloads. This observation motivates a design principle for cache replacement: randomize the eviction path to mitigate worst-case scans while keeping the per-object state minimal. Intuitively, randomization mitigates performance cliffs by dispersing eviction pressure across objects; from a queueing-theoretic perspective, this dispersion is akin to a per-object birth-death process. We thus introduce a randomized SIEVE/CLOCK variant that uses ⌈ log 2 (K+1) ⌉ ≥ 1 access bits per cached item; it is as cheap to implement as CLOCK using a single circular array, yet remains robust to scans and amenable to a natural CTMC description, allowing accurate performance approximation via a heterogeneous mean-field model. Compared with popular baselines and their generalizations, the randomized SIEVE/CLOCK achieves hit rates of 1.5× or more on production workloads with long scan sequences. Furthermore, though the cache hit rate improves with the number of access bits, we show that the majority of the gain is already achieved with as few as 4 access bits per cached item.

Large language model (LLM) inference consumes substantial datacenter energy and produces both carbon emissions and criteria pollutants (PM2.5, SO2, NOx). These criteria pollutants impact public health depending on the meteorological conditions of emission regions and population exposure. We apply health impact assessment to LLM inference, a use case of growing concern due to its rapidly rising energy consumption, and characterize health impacts from datacenter energy consumption as quantifiable health cost metrics. We then present HealthServe, a systems-level framework for health-aware LLM inference serving, which jointly minimizes carbon emissions and health costs across geo-distributed GPU clusters while meeting latency SLOs. HealthServe employs a hierarchical scheduling architecture with Social Cognitive Optimization (SCO) that maintains condition-indexed decision libraries for configuration reuse under recurring operating conditions. Evaluation on a heterogeneous GPU cluster across different U.S. datacenter locations demonstrates over 50% carbon footprint and over 25% health cost reductions over the state-of-the-art energy-aware LLM inference serving solutions. Additionally, we show that HealthServe can be augmented with existing carbon-aware serving approaches to provide additive sustainability benefits.