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.

Server replacement scheduling decisions can significantly impact carbon emissions in data centers. However, commonly used periodic strategies are often sub-optimal in terms of total carbon emissions. In this article, we formulate the carbon-aware server replacements problem mathematically and investigate the structure of the optimal replacement policy. Based on our theoretical analysis, we then propose a simple adaptive online policy that yields significantly lower carbon emissions than periodic strategies and achieves near-optimal results across a variety of simulated scenarios.

We investigate how to allocate resources when the capacity of the channel fluctuates over time, and when the scheduler does not have full information regarding the actual state of the system. The main motivation comes from communication networks in which the available capacity might evolve over time. A nice introduction to the role of information in scheduling, mostly in the context of stability, is provided in [1]. This paper shows how the problem can be cast as a Partially Observable MDP (POMDP), and observes that in full generality there is very little that can be said. We thus set out to study a simple, yet interesting, instance of the problem in which the channel capacity can only take good values, ON, or bad values OFF. When the channel is ON, data can be transmitted through the channel, but not when it is OFF. Even though the scheduler does not have access to the state of every channel, it does observe successful transmissions.

Network tomography [5] is an essential approach for inferring the internal network (e.g., link) parameters, such as the loss rate, delay, and bandwidth, via end-to-end (external) measurements. These external measurements are practically more accessible than the network's internal components (e.g., routers), which may be owned by different internet service providers (ISPs) [2] and are difficult to access.

Traditional approaches to network management have been accessible only to a handful of highly-trained network operators with significant expert knowledge. This creates barriers for lay users to easily manage their networks without resorting to experts. With recent development of powerful large language models (LLMs) for language comprehension, we design a system to make network management accessible to a broader audience of non-experts by allowing users to converse with networks in natural language. To e!ectively leverage advancements in LLMs, we propose an agentic framework that uses an intermediate representation to streamline configuration across diverse vendor equipment, retrieves the network state from memory in real-time, and provides an interface for external feedback. We also conduct pilot studies to collect real user data of natural language utterances for network control, and present a visualization interface to facilitate dialogue-driven user interaction and enable large-scale data collection for future development. Preliminary experiments validate the e!ectiveness of our proposed system components with LLM integration on both synthetic and real user utterances. Through our data collection and visualization e!orts, we pave the way for more e!ective use of LLMs and democratize network control for everyday users.1

We are delighted to host the inaugural AI Crossroads workshop, colocated with ACMSIGMETRICS 2025 on June 13, 2025 at Stony Brook, NY. Our aim is to unite researchers at the intersection of artificial intelligence, scalable computer and networked systems, and sustainable energy technologies. We hope the papers and discussions captured here will spark new ideas on how AI can both drive and benefit from efficient, environmentally responsible infrastructures.

We reinterpret the Leiden algorithm for community detection through a game-theoretic lens, modeling it as a hedonic game where nodes act as rational agents. This perspective uncovers an equivalence between Constant Potts Model (CPM) optimization and Nash equilibria. We introduce an agent-level notion of frustration, which arises when agents face a trade-off between maximizing intra-community links and minimizing non-links. The resolution parameter resolves this conflict by acting as a decision threshold against a proposed social metric. We prove convergence guarantees and demonstrate empirically that robustness (the absence of frustration) correlates with higher accuracy in synthetic networks with known, symmetric ground-truth communities.

Real-time personalization has advanced significantly in recent years, with platforms utilizing machine learning models to predict user preferences based on rich behavioral data on each individual user. Traditional approaches usually rely on embedding-based machine learning models to capture user preferences, and then reduce the final real-time optimization task to one of nearest-neighbors, which can be performed extremely fast both theoretically and practically. However, these models struggle to capture some complex user behaviors, which are essential for making accurate recommendations. Transformer-based models, on the other hand, are known for their practical ability to model sequential behaviors, and hence have been intensively used in personalization recently to overcome these limitations. However, optimizing recommendations under transformer-based models is challenging due to their complicated architectures. In this paper, we address this challenge by considering a specific class of transformers, showing its ability to represent complex user preferences, and developing efficient algorithms for real-time personalization.

This paper examines a continuous-time routing system with general interarrival and service time distributions, operating under either the join-the-shortest-queue policy or the power-of-two-choices policy. Under a weaker set of assumptions than those commonly established in the literature, we prove that the scaled steady-state queue length at each station converges weakly to a common exponential random variable in heavy traffic. Specifically, our results hold under the assumption of the (2 + ε)th moment for the interarrival and service distributions with some ε > 0. The proof leverages the Palm version of the basic adjoint relationship (BAR) as a key technique.

Today's distributed machine learning (DML) introduces heavy traffic load, making the network one of the primary bottlenecks. To mitigate this bottleneck, existing state-of-the-art network optimization methods, such as traffic or topology engineering, are proposed to adapt to real-time traffic. However, current traffic measurement and prediction methods struggle to collect sufficiently fine-grained and accurate traffic patterns. This limitation impedes the ability of cutting-edge network optimization techniques to react agilely to the ever-changing traffic demands of DML jobs. This paper proposes NetJIT, a novel program-behavior-aware toolkit for foreseeing the traffic pattern of DML. To the best of our knowledge, this is the first work proposing the use of just-in-time (JIT) program analysis for real-time traffic measurement. In DML applications, communication behavior is primarily determined by the previously computed results. NetJIT leverages this characteristic to anticipate communication details by tracing and analyzing the data relations in the computation process. This capability enables the deployment of optimization strategies in advance. We deploy NetJIT in real-world network optimization for traffic preknowledge. Evaluation with the self-built testbed demonstrates that NetJIT can achieve up to about 97% less error of detecting communication events compared with other methods. Simulations with real-world DML workloads further illustrate that NetJIT enables more precise network optimization, leading to approximately 50% better network performance w.r.t the metrics including average iteration time, throughput, and average packet delay.