In several queueing systems, arriving tasks have both service and timing requirements, and depart whenever one of them is attained. In several applications, the partial work performed by the system on tasks during their stay is still useful, even if the task is not completed. We analyze these partial service queues with abandonment under a many-server setting and characterize the equilibrium performance of several policies through fluid models. In particular, we show that deadline-oblivious policies like Least-Attained-Service and Last-Come-First-Served exhibit the same equilibrium performance as Earliest-Deadline-First. This striking property means that designers may avoid the difficult job of estimating deadlines without performance penalties. The conclusions are validated by numerical experiments.
Quantization reduces memory, computation, and communication costs by converting high-precision values into compact representations. Recent adaptive stochastic quantization (ASQ) methods optimize quantization values for a specific input vector, but they do not exploit shared randomness available to both the quantizer and the dequantizer. We introduce adaptive unbiased quantization (AUQ), which extends ASQ with shared randomness and can asymptotically improve achievable accuracy without increasing the encoded size. We then present Simba, a practical AUQ method that uses multiple sets of quantization values and selects among them using shared randomness while preserving unbiasedness. Simba starts by computing a fast approximate ASQ solution and refines it through a structured shared-randomness search. Across common synthetic distributions, Simba is up to about 24× more accurate than the state-of-the-art optimal ASQ method and can be up to two orders of magnitude faster. We also show that Simba can improve QUIC-FL by replacing its solver-generated lookup tables with tables computed by Simba.
We propose using Fleming-Viot (FV) particle systems to estimate small exceedance probabilities of the electricity grid frequency for the dynamic dimensioning of frequency containment reserves (FCR). The methodology is proposed as a simpler alternative to importance sampling for diffusion processes, which generally requires the optimization of a functional. Under equal sample sizes, the FV estimator can reliably estimate probabilities that are smaller than those estimated with vanilla Monte Carlo by a factor of 10 -4 .
Traditional enterprise networks rely on security perimeters for defense. These perimeters follow a ''castle-and-moat'' approach, where secure boundaries are established both at the edge of the network and around critical points within the network. However, such an approach assumes implicit trust for all assets within the boundaries. This lack of fine-grained access control enables modern attacks, such as Advanced Persistent Threat (APT), to exploit vulnerabilities in these boundaries and progress toward critical targets within the network, leading to significant financial losses [3].
Federated Learning (FL) enables collaborative model training across distributed clients while preserving data privacy. However, practical FL systems face challenges due to partial client participation and device heterogeneity, which introduce significant variance in training updates and hinder model convergence. Existing methods naïvely reuse stale updates from inactive clients to stabilize training without accounting for varying levels of staleness or optimizing the balance between fresh and stale updates to minimize training variance. To address these issues, we propose FedVarM, a novel method that optimally leverages stale updates by adjusting client aggregation coefficients to minimize training variance. Additionally, we introduce FedVarM-Approx, an efficient approximation without extra computation on the client-side to ensure its efficient deployment. Experiments demonstrate FedVarM achieves 17.4% higher accuracy compared to methods that naïvely reuse stable updates without optimal balancing.
To meet the growing demand for wireless services with diverse throughput, latency, and reliability requirements, network providers employ virtualization to create multiple virtual networks over a shared physical infrastructure. Network slicing (NS) builds on this by partitioning the infrastructure into logically isolated, end-to-end slices, each tailored for specific use cases. Leveraging software-defined networking (SDN) and network function virtualization (NFV), these slices can dynamically allocate computing, storage, and networking resources across the radio access, transport, and core networks. As systems evolve toward multi-domain, multi-technology architectures integrating terrestrial, aerial, and satellite components, NS must intelligently allocate radio, computing, and storage resources to diverse applications- including virtual, augmented, and mixed reality, autonomous vehicles, smart cities, Industry 4.0, digital twins, telemedicine, immersive education, and intelligent transportation- across heterogeneous infrastructures, while addressing device mobility and dynamic channel conditions. This work evaluates a novel NS framework that dynamically manages resources across fog, edge, and cloud layers within terrestrial domains and adapts allocation between terrestrial and nonterrestrial domains to improve reliability and reduce outages. The framework incorporates resilience and covert communication mechanisms for enhanced security in heteroge-neous IoT scenarios. Simulations show improved reliability, lower latency, and strong security, enabling effective support for diverse use cases in integrated multi-domain networks.
We study a dynamic game where a planner, under cost constraints, repeatedly allocates a resource to a group of agents who may strategically hide their true utilities. Via lazy dual updates, dual-adjusted payment rules, and uniform exploration rounds, we design an incentive-aware primal allocation framework. Further plugging in a novel Online Learning algorithm called O-FTRL-FP to learn dual variables, we show that one can achieve allocation efficiency, incentivecompatibility, and constraint-compliance at the same time.
Maintaining low delay is an aim of virtually every computer system, but any system with limited resources inevitably runs into queueing delays when load is high. Queueing delays occur at every level of abstraction, e.g. packet flows wait at network switches, queries wait in databases, and large-scale simulations wait for supercomputer time. One of the main tools we have to combat delay is load balancing, i.e. dispatching jobs (e.g. packet flows, queries, simulations) across servers (e.g. network switches, databases, supercomputers) to reduce queueing. The main metrics one evaluates load balances with are tail delays, or P[T > x] for large x, where T is the response time. This formalizes the goal of having few jobs experience large delays, which is a realistic goal for many computer systems practitioners.
Datacenter networks form the foundation of today's large-scale computing infrastructure, supporting a wide range of online services and data-intensive applications. As these systems continue to evolve, the pressure on the network to deliver higher bandwidth and lower latency has intensified. Measurements from production environments already show dramatic growth trends. For example, server-generated traffic in Google datacenters increased by a factor of fifty over a six-year period between 2008 and 2014. This growth is expected to accelerate further with the widespread deployment of GPU clusters for large-scale distributed training.
Fundamentally, utility defines the functional capacity of a blockchain system to execute its intended tasks, from propagating data to finalizing transactions. However, this overarching metric manifests differently across the technology stack. At the network layer, utility equates to connection speed and message propagation; at the consensus layer, it translates to agreement efficiency and fault tolerance. This work addresses four distinct utility problems across the stack, proposing algorithmic solutions that enhance performance at the network, consensus, Layer 2, and application layers.
Queueing theory has proved to be a successful tool for studying several service systems like analyzing data centers [12]. A popular approach is to study such systems in the steady-state owing to the tractability of analysis. However, these systems rarely operate in a steady state, so the theoretical analysis does not directly translate into real-life performance guarantees. Furthermore, a typical goal in the performance analysis of service systems is to understand queue length/waiting time behavior. However, it is not possible to obtain the exact queue length distribution except in special cases such as the M/M/1 queue [15]. So typically, one uses the steady-state behavior as a proxy for the finite-time behavior. However, this approximation is useful only when the system mixes quickly, i.e. approaches the steady-state behavior quickly. These discrepancies motivate us to analyze the rate of convergence to the steady state. Such a result allows one to better understand the applicability of the steady-state results. Moreover, sometimes it is not even possible to obtain the close form steady-state distribution of the system in many queuing systems, and therefore one results to asymptotic analysis such as heavy traffic analysis [8]. Recent works have characterized the rate of convergence in these regimes, thus obtaining a handle on the pre-asymptotic queue length behavior [9, 2]. Since these approximation errors are larger when one is further away from the asymptotic regime, and therefore one has a better handle on the stationary distribution when one is closer to the asymptotic regime. However, prior works suggest that the queuing systems mix slower as one gets closer to the asymptotic regimes, for example, in the Heavy Traffic regime [14, 15]. And so, it is important to characterize the goodness of the steady-state analysis and asymptotic analysis to the finite-time queue length behavior. Therefore, our objective is to establish the finite-time convergence to stationarity bounds of queuing systems, particularly the Join-the-Shortest-Queue policy.
There has been a growing interest in strategies to optimize the cost-performance tradeoff of LLMs, in particular LLM selection, where the aim is to understand which language models perform better than others for a given task. However, there arises some applications where simply finding the LLM which performs the best on the task is not sufficient as the task may be too specialized and difficult for an LLM to handle alone. In such cases considering a pipeline of LLMs seems to be more suitable to decrease the difficulty of the task by breaking it down into smaller subtasks where an LLM's output is fed as the input to another LLM. The task of suggesting diagnosis for patients based on provided medical reports of patients which are lengthy and include medical language is such an application. We believe that for such a task, the LLMs can benefit from being given a summarized version of the medical report (summary obtained through another LLM) which highlights the key points and necessary information which can guide diagnosis decisions.
Every year, the Association for Computing Machinery (ACM) spearheads a series of Student Research Competitions (SRCs) at ACM-sponsored or co-sponsored conferences. These SRCs provide graduate (Masters or PhD program) and undergraduate students an opportunity to present one of their research projects to the conference. At ACM SIGMETRICS, the competition was held as follows. First, the students submitted an extended abstract to a program committee that selects a subset of the abstracts that will be presented at the conference. A few of these posters were selected for an oral presentation. Participating students have the opportunity to win cash prizes. In each category (graduate and undergraduate), the 1st, 2nd, and 3rd place winners receive prizes of 500, 300, and 200, respectively. Additionally, first-place undergraduate and graduate student winners from the SRCs held during the year are invited to present their work at the SRC Grand Finals, where a different panel of judges evaluates these winners against each other.
My research focuses on designing resource-efficient, intelligent, and reliable systems that span the cloud-to-edge continuum and built environments. My primary methodology is to leverage the flexibility of computing workloads to design efficient and dynamic systems that adapt to the real world's dynamics. By integrating insights from systems, machine learning, theory, and optimization, my work aims to reimagine and establish foundational design principles to build efficient, agile, and reliable systems. Utilizing this methodology enabled me to make fundamental contributions to system design and resource management algorithms, optimizing the carbon emissions of AI training and inference workloads, as well as AI servicing systems for resource-constrained edge environments.
We frame community detection as a hedonic game where nodes face "frustrated choices" between maximizing internal links and minimizing internal non-links. The resolution parameter balances these objectives, and the framework ensures any sequence of selfish moves converges to a stable equilibrium, viewing community structure as the outcome of a strategic game.
Networked systems typically operate under limited resources, under strict stability and latency requirements, and with incomplete knowledge of their environments. The control of networked systems may therefore benefit greatly from learning-based decision-making in order to maximize overall network utility. However, this utility maximization goal is complicated by four notable challenges that recur across networked systems: constraints in the form of service guarantees and budgeted resource consumption, operational costs due to reconfiguration and service changes, i.e., switching costs, feedback delays that both shape user experience and influence what the learner observes, and risk aversion to changing system performance in user-centric systems. These four aspects are even more challenging when the performance-relevant random processes, e.g., job arrivals, service, user motion, and resource availability, are unknown.
Large Language Models (LLMs) offer unprecedented language understanding and generation capabilities. However, these advancements come at a cost. LLMs rely heavily on high-performance computing, not only for training but also for inference. This dependence translates into substantial energy consumption, with inference potentially exceeding the already significant environmental impact of training, which can generate thousands of tons of CO2eq. Therefore, quantifying the energy consumption of LLM inference is crucial. This research focuses on measuring this energy use across a wide range of transformer models, from smaller architectures to cutting-edge models like DeepSeek V3/R1. We utilize state-of-the-art inference frameworks and high-end GPUs to ensure accurate assessments.
The ACM SIGMETRICS Doctoral Dissertation Award recognizes outstanding doctoral thesis research in the field of performance evaluation of computer systems. Nominations for the 2025 award were solicited from faculty supervising graduating Ph.D. students whose work contributes to performance evaluation, broadly understood to include theoretical, methodological, and applied research across a wide range of systems and application domains.
His research is centered on sequential decision-making under uncertainty, aiming both to deepen theoretical understanding of decision-making with realistic feedback and to enhance practical performance in advanced application domains-especially multi-agent systems and quantum networks. His work was recognized as a Best Paper Finalist at SIGMETRICS 2025.
Modern sequential decision-making systems, such as datacenter demand response and online ad allocation, operate under significant uncertainty while requiring strict adherence to safety and resource constraints. While machine learning (ML) offers transformative potential for improving average-case efficiency, its inherent brittleness to distributional shifts poses a critical barrier to deployment in mission-critical applications. This work presents a comprehensive framework for robust learning-augmented algorithms that bridge this gap. We introduce a unified methodology for calibrating untrusted ML predictions using provable expert safeguards, such as differentiable calibrators and reservation costs, to guarantee bounded worst-case costs. We extend these foundations to structured and decentralized decision spaces, developing algorithms like LOMAR for online bipartite matching and LADO for networked control that balance robust guarantees with ML-driven gains. Finally, we translate these theoretical advances into practice for socially responsible AI, demonstrating how equity-aware geographical load balancing (eGLB) and fair schedulers (LCF) can mitigate regional environmental disparities in AI operations.
