arXiv subjects

Haisheng Tan

Publications and source records attributed to Haisheng Tan.

18 recordsLinked to original sources

arXiv2026

TOPAS: Workflow-Aware Prefix-State Scheduling for Multi-Agent LLM Serving

Prefix caching introduces a fundamental tradeoff in multi-agent large language model (LLM) serving: retaining a long system-prompt key-value (KV) cache for an agent accelerates future calls, yet it reduces the GPU memory available for batching concurrent requests. In multi-stage workflows, existing schedulers tend to prioritize either immediate prefix locality or overall workflow progress. However, under a shared KV cache budget, optimizing either objective in isolation can prolong tasklevel job completion time (JCT) through downstream delays or frequent prefix replacement. To strike a balance, we here propose TOPAS, a Task-Oriented Prefix-Aware Scheduler that jointly decides which agent prefixes to keep in the cache and which requests to schedule for execution. TOPAS scores candidate post-decision states by trading off the expected reduction in each task's longest remaining service path against the near-term benefit of downstream prefix reuse, accounting for the costs of prefix movement and preemption. A task-level aging mechanism is also incorporated to prevent starvation. We implement TOPAS within the SGLang framework and assess its performance on three synthetic DAGs and two MetaGPT software-development workflows. Compared with the best performing baseline for each workload and metric, TOPAS reduces the mean/p99 JCT by up to 39.8%/49.4% on the synthetic workloads, while lowering mean JCT by 9.8% on MetaGPT-SOP and mean/p99 JCT by 22.0%/26.6% on MetaGPT-TL.

cs.CL↗

arXiv2025

A Plan Reuse Mechanism for LLM-Driven Agent

Integrating large language models (LLMs) into personal assistants, like Xiao Ai and Blue Heart V, effectively enhances their ability to interact with humans, solve complex tasks, and manage IoT devices. Such assistants are also termed LLM-driven agents. Upon receiving user requests, the LLM-driven agent generates plans using an LLM, executes these plans through various tools, and then returns the response to the user. During this process, the latency for generating a plan with an LLM can reach tens of seconds, significantly degrading user experience. Real-world dataset analysis shows that about 30% of the requests received by LLM-driven agents are identical or similar, which allows the reuse of previously generated plans to reduce latency. However, it is difficult to accurately define the similarity between the request texts received by the LLM-driven agent through directly evaluating the original request texts. Moreover, the diverse expressions of natural language and the unstructured format of plan texts make implementing plan reuse challenging. To address these issues, we present and implement a plan reuse mechanism for LLM-driven agents called AgentReuse. AgentReuse leverages the similarities and differences among requests' semantics and uses intent classification to evaluate the similarities between requests and enable the reuse of plans. Experimental results based on a real-world dataset demonstrate that AgentReuse achieves a 93% effective plan reuse rate, an F1 score of 0.9718, and an accuracy of 0.9459 in evaluating request similarities, reducing latency by 93.12% compared with baselines without using the reuse mechanism.

cs.MA↗

arXiv2025

Embedding Samples Dispatching for Recommendation Model Training in Edge Environments

Training deep learning recommendation models (DLRMs) on edge workers brings several benefits, particularly in terms of data privacy protection, low latency and personalization. However, due to the huge size of embedding tables, typical DLRM training frameworks adopt one or more parameter servers to maintain global embedding tables, while leveraging the edge workers cache part of them. This incurs significant transmission cost for embedding transmissions between workers and parameter servers, which can dominate the training cycle. In this paper, we investigate how to dispatch input embedding samples to appropriate edge workers to minimize the total embedding transmission cost when facing edge-specific challenges such as heterogeneous networks and limited resources. We develop ESD, a novel mechanism that optimizes the dispatch of input embedding samples to edge workers based on expected embedding transmission cost. We propose HybridDis as the dispatch decision method within ESD, which combines a resource-intensive optimal algorithm and a heuristic algorithm to balance decision quality and resource consumption. We implement a prototype of ESD and compare it with state-of-the-art mechanisms on real-world workloads. Extensive experimental results show that ESD reduces the embedding transmission cost by up to 36.76% and achieves up to 1.74 times speedup in end-to-end DLRM training.

cs.DC↗

arXiv2025

Astraea: A State-Aware Scheduling Engine for LLM-Powered Agents

Large Language Models (LLMs) are increasingly being deployed as intelligent agents. Their multi-stage workflows, which alternate between local computation and calls to external network services like Web APIs, introduce a mismatch in their execution pattern and the scheduling granularity of existing inference systems such as vLLM. Existing systems typically focus on per-segment optimization which prevents them from minimizing the end-to-end latency of the complete agentic workflow, i.e., the global Job Completion Time (JCT) over the entire request lifecycle. To address this limitation, we propose Astraea, a service engine designed to shift the optimization from local segments to the global request lifecycle. Astraea employs a state-aware, hierarchical scheduling algorithm that integrates a request's historical state with future predictions. It dynamically classifies requests by their I/O and compute intensive nature and uses an enhanced HRRN policy to balance efficiency and fairness. Astraea also implements an adaptive KV cache manager that intelligently handles the agent state during I/O waits based on the system memory pressure. Extensive experiments show that Astraea reduces average JCT by up to 25.5\% compared to baseline methods. Moreover, our approach demonstrates strong robustness and stability under high load across various model scales.

cs.CL↗

arXiv2025

PICE: A Semantic-Driven Progressive Inference System for LLM Serving in Cloud-Edge Networks

Large language models (LLMs), while driving a new wave of interactive AI applications across numerous domains, suffer from high inference costs and heavy cloud dependency. Motivated by the redundancy phenomenon in linguistics, we propose a progressive inference paradigm over cloud and edge, i.e., firstly generating the sketch of the answer by LLMs at cloud, and then conducting parallel extension to fill in details by small models (SLMs) at edge. Progressive inference offers potential benefits to improve throughput and reduce inference latency while facing key implementation challenges, including decreased response quality from SLMs, a tradeoff between the brevity and comprehensiveness of sketches, as well as increased latency caused by network transmission and edge inference. In this work, we propose and implement PICE, an LLM serving system with semantic-level cloud-edge collaboration, enhancing inference throughput and quality through dynamic inference task scheduling, ensemble learning, and parallel edge inference. Extensive testbed experiments illustrate that our approach achieves $1.5-2\times$ throughput enhancement and up to 43% latency reduction, while also potentially enhancing the quality compared to SOTA systems.

cs.DC↗

arXiv2025

Real-Time Neural-Enhancement for Online Cloud Gaming

Online Cloud gaming demands real-time, high-quality video transmission across variable wide-area networks (WANs). Neural-enhanced video transmission algorithms employing super-resolution (SR) for video quality enhancement have effectively challenged WAN environments. However, these SR-based methods require intensive fine-tuning for the whole video, making it infeasible in diverse online cloud gaming. To address this, we introduce River, a cloud gaming delivery framework designed based on the observation that video segment features in cloud gaming are typically repetitive and redundant. This permits a significant opportunity to reuse fine-tuned SR models, reducing the fine-tuning latency of minutes to query latency of milliseconds. To enable the idea, we design a practical system that addresses several challenges, such as model organization, online model scheduler, and transfer strategy. River first builds a content-aware encoder that fine-tunes SR models for diverse video segments and stores them in a lookup table. When delivering cloud gaming video streams online, River checks the video features and retrieves the most relevant SR models to enhance the frame quality. Meanwhile, if no existing SR model performs well enough for some video segments, River will further fine-tune new models and update the lookup table. Finally, to avoid the overhead of streaming model weight to the clients, River designs a prefetching strategy that predicts the models with the highest possibility of being retrieved. Our evaluation based on real video game streaming demonstrates River can reduce redundant training overhead by 44% and improve the Peak-Signal-to-Noise-Ratio by 1.81dB compared to the SOTA solutions. Practical deployment shows River meets real-time requirements, achieving approximately 720p 20fps on mobile devices.

cs.NI↗

arXiv2024

An Experimental Study of Passive UAV Tracking with Digital Arrays and Cellular Downlink Signals

Given the prospects of the low-altitude economy (LAE) and the popularity of unmanned aerial vehicles (UAVs), there are increasing demands on monitoring flying objects at low altitude in wide urban areas. In this work, the widely deployed long-term evolution (LTE) base station (BS) is exploited to illuminate UAVs in bistatic trajectory tracking. Specifically, a passive sensing receiver with two digital antenna arrays is proposed and developed to capture both the line-of-sight (LoS) signal and the scattered signal off a target UAV. From their cross ambiguity function, the bistatic range, Doppler shift and angle-of-arrival (AoA) of the target UAV can be detected in a sequence of time slots. In order to address missed detections and false alarms of passive sensing, a multi-target tracking framework is adopted to track the trajectory of the target UAV. It is demonstrated by experiments that the proposed UAV tracking system can achieve a meter-level accuracy.

eess.SP↗

arXiv2023

Dynamic Uploading Scheduling in mmWave-Based Sensor Networks via Mobile Blocker Detection

The freshness of information, measured as Age of Information (AoI), is critical for many applications in next-generation wireless sensor networks (WSNs). Due to its high bandwidth, millimeter wave (mmWave) communication is seen to be frequently exploited in WSNs to facilitate the deployment of bandwidth-demanding applications. However, the vulnerability of mmWave to user mobility typically results in link blockage and thus postponed real-time communications. In this paper, joint sampling and uploading scheduling in an AoI-oriented WSN working in mmWave band is considered, where a single human blocker is moving randomly and signal propagation paths may be blocked. The locations of signal reflectors and the real-time position of the blocker can be detected via wireless sensing technologies. With the knowledge of blocker motion pattern, the statistics of future wireless channels can be predicted. As a result, the AoI degradation arising from link blockage can be forecast and mitigated. Specifically, we formulate the long-term sampling, uplink transmission time and power allocation as an infinite-horizon Markov decision process (MDP) with discounted cost. Due to the curse of dimensionality, the optimal solution is infeasible. A novel low-complexity solution framework with guaranteed performance in the worst case is proposed where the forecast of link blockage is exploited in a value function approximation. Simulations show that compared with several heuristic benchmarks, our proposed policy, benefiting from the awareness of link blockage, can reduce average cost up to 49.6%.

eess.SY↗

arXiv2023

Predictive Resource Allocation in mmWave Systems with Rotation Detection

Millimeter wave (MmWave) has been regarded as a promising technology to support high-capacity communications in 5G era. However, its high-layer performance such as latency and packet drop rate in the long term highly depends on resource allocation because mmWave channel suffers significant fluctuation with rotating users due to mmWave sparse channel property and limited field-of-view (FoV) of antenna arrays. In this paper, downlink transmission scheduling considering rotation of user equipments (UE) and limited antenna FoV in an mmWave system is optimized via a novel approximate Markov decision process (MDP) method. Specifically, we consider the joint downlink UE selection and power allocation in a number of frames where future orientations of rotating UEs can be predicted via embedded motion sensors. The problem is formulated as a finite-horizon MDP with non-stationary state transition probabilities. A novel low-complexity solution framework is proposed via one iteration step over a base policy whose average future cost can be predicted with analytical expressions. It is demonstrated by simulations that compared with existing benchmarks, the proposed scheme can schedule the downlink transmission and suppress the packet drop rate efficiently in non-stationary mmWave links.

eess.SY↗

arXiv2023

An Indoor Environment Sensing and Localization System via mmWave Phased Array

An indoor layout sensing and localization system in 60GHz millimeter wave (mmWave) band, named mmReality, is elaborated in this paper. The mmReality system consists of one transmitter and one mobile receiver, each with a phased array and a single radio frequency (RF) chain. To reconstruct the room layout, the pilot signal is delivered from the transmitter to the receiver via different pairs of transmission and receiving beams, so that the signals at all antenna elements can be resolved. Then, the spatial smoothing and two-dimensional multiple signal classification (MUSIC) algorithm is applied to detect the angle-of-arrival (AoAs) and angle-of-departure (AoDs) of the rays from the transmitter to the receiver. Moreover, the technique of multi-carrier ranging is adopted to measure the distance of each propagation path. Synthesizing the above geometrical parameters, the location of receiver relative to the transmitter can be pinpointed, both line-of-sight (LoS) and non-line-of-sight (NLoS) paths can also be determined. Therefore, the room layout can be reconstructed by moving the receiver and repeating the above measurement in different locations of the room. At the end, we show that the reconstructed room layout can be utilized to locate a mobile device according to its AoA spectrum, even with single access point.

eess.SP↗

arXiv2020

Online Distributed Job Dispatching with Outdated and Partially-Observable Information

In this paper, we investigate online distributed job dispatching in an edge computing system residing in a Metropolitan Area Network (MAN). Specifically, job dispatchers are implemented on access points (APs) which collect jobs from mobile users and distribute each job to a server at the edge or the cloud. A signaling mechanism with periodic broadcast is introduced to facilitate cooperation among APs. The transmission latency is non-negligible in MAN, which leads to outdated information sharing among APs. Moreover, the fully-observed system state is discouraged as reception of all broadcast is time consuming. Therefore, we formulate the distributed optimization of job dispatching strategies among the APs as a Markov decision process with partial and outdated system state, i.e., partially observable Markov Decision Process (POMDP). The conventional solution for POMDP is impractical due to huge time complexity. We propose a novel low-complexity solution framework for distributed job dispatching, based on which the optimization of job dispatching policy can be decoupled via an alternative policy iteration algorithm, so that the distributed policy iteration of each AP can be made according to partial and outdated observation. A theoretical performance lower bound is proved for our approximate MDP solution. Furthermore, we conduct extensive simulations based on the Google Cluster trace. The evaluation results show that our policy can achieve as high as $20.67\%$ reduction in average job response time compared with heuristic baselines, and our algorithm consistently performs well under various parameter settings.

cs.NI↗

arXiv2020

Entanglement as upper bounded for the nonlocality of a general two-qubit system

Nonlocality and entanglement are not only the fundamental characteristics of quantum mechanics but also important resources for quantum information and computation applications. Exploiting the quantitative relationship between the two different resources is of both theoretical and practical significance. The common choice for quantifying the nonlocality of a two-qubit state is the maximal violation of the Clauser-Horne-Shimony-Holt inequality. That for entanglement is entanglement of formation, which is a function of the concurrence. In this paper, we systematically investigate the quantitative relationship between the entanglement and nonlocality of a general two-qubit system. We rederive a known upper bound on the nonlocality of a general two-qubit state, which depends on the state's entanglement. We investigate the condition that the nonlocality of two different two-qubit states can be optimally stimulated by the same nonlocality test setting and find the class of two-qubit state pairs that have this property. Finally, we obtain the necessary and sufficient condition that the upper bound can be reached.

quant-ph↗

arXiv2020

Enhance the performance of navigation: A two-stage machine learning approach

Real time traffic navigation is an important capability in smart transportation technologies, which has been extensively studied these years. Due to the vast development of edge devices, collecting real time traffic data is no longer a problem. However, real traffic navigation is still considered to be a particularly challenging problem because of the time-varying patterns of the traffic flow and unpredictable accidents/congestion. To give accurate and reliable navigation results, predicting the future traffic flow(speed,congestion,volume,etc) in a fast and accurate way is of great importance. In this paper, we adopt the ideas of ensemble learning and develop a two-stage machine learning model to give accurate navigation results. We model the traffic flow as a time series and apply XGBoost algorithm to get accurate predictions on future traffic conditions(1st stage). We then apply the Top K Dijkstra algorithm to find a set of shortest paths from the give start point to the destination as the candidates of the output optimal path. With the prediction results in the 1st stage, we find one optimal path from the candidates as the output of the navigation algorithm. We show that our navigation algorithm can be greatly improved via EOPF(Enhanced Optimal Path Finding), which is based on neural network(2nd stage). We show that our method can be over 7% better than the method without EOPF in many situations, which indicates the effectiveness of our model.

eess.SP↗

arXiv2020

Joint Optimization of File Placement and Delivery in Cache-Assisted Wireless Networks with Limited Lifetime and Cache Space

In this paper, the scheduling of downlink file transmission in one cell with the assistance of cache nodes with finite cache space is studied. Specifically, requesting users arrive randomly and the base station (BS) reactively multicasts files to the requesting users and selected cache nodes. The latter can offload the traffic in their coverage areas from the BS. We consider the joint optimization of the abovementioned file placement and delivery within a finite lifetime subject to the cache space constraint. Within the lifetime, the allocation of multicast power and symbol number for each file transmission at the BS is formulated as a dynamic programming problem with a random stage number. Note that there are no existing solutions to this problem. We develop an asymptotically optimal solution framework by transforming the original problem to an equivalent finite-horizon Markov decision process (MDP) with a fixed stage number. A novel approximation approach is then proposed to address the curse of dimensionality, where the analytical expressions of approximate value functions are provided. We also derive analytical bounds on the exact value function and approximation error. The approximate value functions depend on some system statistics, e.g., requesting users' distribution. One reinforcement learning algorithm is proposed for the scenario where these statistics are unknown.

cs.IT↗

arXiv2019

Cooperative Job Dispatching in Edge Computing Network with Unpredictable Uploading Delay

In this paper, the cooperative jobs dispatching problem in an edge computing network with multiple access points (APs) and edge servers is considered. Due to the uncertain traffic in the network between APs and edge servers, the job uploading delay can not be predicted accurately. Specifically, the job arrivals at the APs, the job uploading delay from APs to edge servers and the job computation time at the edge servers are all modeled as random variables. Since each job dispatching decision will affect the system state in the future, we formulate the joint optimization of jobs dispatching at all the APs and all the scheduling time slots as an infinite-horizon Markov decision process (MDP). The minimization objective is a discounted measurement of the average processing time per job, including the uploading delay, the waiting time and the computation time at the edge servers. In this problem, the approximate MDP should be adopted to address the curse of dimensionality. Conventional low-complexity approximate solution of MDP is usually hard to predict the performance analytically. In this paper, a novel approximate MDP solution framework is proposed via one-step policy iteration over a baseline policy, where the analytical performance bound can be obtained. Moreover, since the expression of the approximate value function is derived, the value iteration in conventional methods can be eliminated, which can essentially reduce the computation complexity. It is shown by simulations that proposed low-complexity algorithm has significantly better performance than various benchmark schemes.

cs.NI↗

arXiv2016

Dynamic Virtual Machine Management via Approximate Markov Decision Process

Efficient virtual machine (VM) management can dramatically reduce energy consumption in data centers. Existing VM management algorithms fall into two categories based on whether the VMs' resource demands are assumed to be static or dynamic. The former category fails to maximize the resource utilization as they cannot adapt to the dynamic nature of VMs' resource demands. Most approaches in the latter category are heuristical and lack theoretical performance guarantees. In this work, we formulate dynamic VM management as a large-scale Markov Decision Process (MDP) problem and derive an optimal solution. Our analysis of real-world data traces supports our choice of the modeling approach. However, solving the large-scale MDP problem suffers from the curse of dimensionality. Therefore, we further exploit the special structure of the problem and propose an approximate MDP-based dynamic VM management method, called MadVM. We prove the convergence of MadVM and analyze the bound of its approximation error. Moreover, MadVM can be implemented in a distributed system, which should suit the needs of real data centers. Extensive simulations based on two real-world workload traces show that MadVM achieves significant performance gains over two existing baseline approaches in power consumption, resource shortage and the number of VM migrations. Specifically, the more intensely the resource demands fluctuate, the more MadVM outperforms.

cs.NI↗

arXiv2014

Optimal Rendezvous Strategies for Different Environments in Cognitive Radio Networks

In Cognitive Radio Networks (CRNs), the secondary users (SUs) are allowed to access the licensed channels opportunistically. A fundamental and essential operation for SUs is to establish communication through choosing a common channel at the same time slot, which is referred to as rendezvous problem. In this paper, we study strategies to achieve fast rendezvous for two secondary users. The channel availability for secondary nodes is subject to temporal and spatial variation. Moreover, in a distributed system, one user is oblivious of the other user's channel status. Therefore, a fast rendezvous is not trivial. Recently, a number of rendezvous strategies have been proposed for different system settings, but rarely have they taken the temporal variation of the channels into account. In this work, we first derive a time-adaptive strategy with optimal expected time-to-rendezvous (TTR) for synchronous systems in stable environments, where channel availability is assumed to be static over time. Next, in dynamic environments, which better represent temporally dynamic channel availability in CRNs, we first derive optimal strategies for two special cases, and then prove that our strategy is still asymptotically optimal in general dynamic cases. Numerous simulations are conducted to demonstrate the performance of our strategies, and validate the theoretical analysis. The impacts of different parameters on the TTR are also investigated, such as the number of channels, the channel open possibilities, the extent of the environment being dynamic, and the existence of an intruder.

cs.NI↗

arXiv2013

On the Complexity of Connectivity in Cognitive Radio Networks Through Spectrum Assignment

Cognitive Radio Networks (CRNs) are considered as a promising solution to the spectrum shortage problem in wireless communication. In this paper, we initiate the first systematic study on the algorithmic complexity of the connectivity problem in CRNs through spectrum assignments. We model the network of secondary users (SUs) as a potential graph, where two nodes having an edge between them are connected as long as they choose a common available channel. In the general case, where the potential graph is arbitrary and the SUs may have different number of antennae, we prove that it is NP-complete to determine whether the network is connectable even if there are only two channels. For the special case where the number of channels is constant and all the SUs have the same number of antennae, which is more than one but less than the number of channels, the problem is also NP-complete. For the special cases in which the potential graph is complete, a tree, or a graph with bounded treewidth, we prove the problem is NP-complete and fixed-parameter tractable (FPT) when parameterized by the number of channels. Exact algorithms are also derived to determine the connectability of a given cognitive radio network.

cs.DS↗