Recent advancements in artificial intelligence (AI) have yielded sophisticated natural language interaction systems, largely driven by Large language models (LLMs). These models generate text auto-regressively by predicting continuations. However, the substantial computational demands of LLM inference typically necessitate reliance on cloud-based services. As demand escalates, organizations are increasingly exploring private cloud deployments to bolster information security and comply with data regulations. Nevertheless, deploying private LLM inference servers entails significant capital expenditure on high-performance Graphics Processing Units (GPUs) and substantial operational costs, including high power consumption, raising environmental concerns. While existing research has explored optimizing individual components of LLM processing, a holistic, system-level perspective on inference efficiency remains underexplored. This paper addresses this gap by formulating private cloud-based LLM inference as a stochastic model. We analyze its performance, focusing on the throughput and processing latency, and specifically investigate the impact of different inference methods and batch sizes on overall efficiency.

Aerial base stations (ABSs), i.e., unmanned aerial vehicle-mounted base stations, have emerged as an essential technology to expand the coverage and capacity in next-generation wireless networks. Determining the optimal ABS placement that maximizes the communication quality of users is a critical but challenging problem due to complex air-to-ground (A2G) channel characteristics and inter-cell interference. Although various ABS deployment methods have been proposed, existing methods do not consider the actual line-of-sight (LoS) and non-LoS (NLoS) conditions of A2G channels properly, assuming a spatially independent probabilistic LoS model to simplify the problem and analysis. In this study, we propose a distributed ABS deployment method considering the actual LoS/NLoS conditions. We model the ABS deployment space as a 3D grid space in which the LoS/NLoS condition between each grid point is given and formulate the ABS deployment problem as the maximization problem of the expected number of covered users. By applying a potential game to the problem, we develop a distributed ABS deployment algorithm that probabilistically converges to a Nash equilibrium that maximizes the objective function. Furthermore, we demonstrate through simulation that the proposed method can improve the actual communication quality of users.

Network operators in telecom carriers perform failure diagnosis by executing various commands (e.g., show interfaces) on network devices through a command line interface (CLI). The outputs of these CLI commands, which we call command outputs, contain more detailed information than typical one-line system logs but have complex characteristics; thus, existing parsers for system logs are not applicable. In this study, we propose COPA, a parsing method for command outputs for automating time-consuming and labor-intensive failure diagnosis.

An unmanned aerial vehicle (UAV)-mounted aerial base station (ABS) is considered as a promising technology to enhance the capacity and coverage in future mobile networks. However, determining the optimal placement of ABSs that maximizes the overall communication quality of users is a challenging problem because of complicated air-to-ground channel characteristics and inter-cell interference. Moreover, ABSs are faced with the problem of limited available power resources. In this study, we consider an ABS network in which ABSs perform radio frequency (RF) energy harvesting from terrestrial base stations and propose a novel distributed ABS deployment method that optimizes the performance of the energy harvesting and downlink transmission jointly. In our method, each ABS updates its position by using local information only and communicating with its neighboring ABSs; thus, the overhead of information gathering is smaller than that in the centralized method. Simulation results demonstrate that our method improves both the energy harvesting and communication quality of users.

We study wireless networked control systems (WNCSs), where controllers operate controlled objects through wireless networks. In WNCSs, controlled objects could become unstable due to bursty packet losses and delays on wireless networks. To solve the problem, we utilize the packetized predictive control (PPC) method, where multiple future control vectors are generated in the receding horizon manner and packed into packets for transmissions. In order to enhance the usefulness of control vectors to be packed, we propose a control vector selection scheme with state clustering (SC), where estimated states are clustered and control packets contain the selected control vectors and the corresponding estimated observation vectors. When the controlled object receives a control packet, it observes its own state and selects an appropriate control vector from among those in the control packet. With simulation experiments, we show that our proposed selection scheme with SC achieves the better performance in terms of the transient response of controlled objects.

This paper considers the stationary distribution of the age of information (AoI) in information update systems. We first derive a general formula for the stationary distribution of the AoI, which holds for a wide class of information update systems. The formula indicates that the stationary distribution of the AoI is given in terms of the stationary distributions of the system delay and the peak AoI. To demonstrate its applicability and usefulness, we analyze the AoI in single-server queues with four different service disciplines: first-come first-served (FCFS), preemptive last-come first-served (LCFS), and two variants of non-preemptive LCFS service disciplines. For the FCFS and the preemptive LCFS service disciplines, the GI/GI/1, M/GI/1, and GI/M/1 queues are considered, and for the non-preemptive LCFS service disciplines, the M/GI/1 and GI/M/1 queues are considered. With these results, we further show comparison results for the mean AoI’s in the M/GI/1 and GI/M/1 queues under those service disciplines.

Recently, network function virtualization has been focused on achieving a flexible deployment of networking services. Despite the fact that the heart of network function virtualization is its software implementation, there are fewer studies on how network functions are implemented as software. In this study, we investigate the evolution of functional connectivity in network functions using an implementation of Internet protocol suite in the Linux kernel. We constructed a call graph for the Linux kernel and analyzed the change of connectivity between the protocol components based on the directory structure of Linux kernel. Our results on the connectivity analysis show that new sub-directories appears for new emerging technologies, such as “bluetooth” and “sctp”, and they rely mostly on the function of sub-directories “core” and “ipv4”. Since the number of functions in sub-directories “core” and “ipv4” is increasing, we defined a functional core for the call graph to see whether there is a core part which has lower variability during the kernel development or not. The result shows that the functional core consists of mostly 50-70 functions and has lower variability comparing with the increase of a number of network-related functions during the kernel development.

We consider the stationary distributions of the age of information (AoI) and the peak AoI in information update systems. We first derive an invariant relation among the distributions of the AoI, the peak AoI, and the system delay, which holds for a wide class of information update systems. Based on this result, we next obtain several formulas for the stationary distributions of the AoI and the peak AoI in the first-come first-served (FCFS) GI/GI/1 queue, which is a general model of FCFS information update systems. Finally, we derive explicit formulas for the Laplace-Stieltjes transforms of the stationary distributions of the AoI and the peak AoI in FCFS M/GI/1 and GI/M/1 queues.

Wireless networked control systems (WNCSs) are control systems whose components are connected through wireless networks. In this paper, we consider a multihop WNCS, where controllers and controlled objects are connected through a multihop wireless network. Round-trip delays between controllers and controlled objects may cause unstable behavior of the WNCS due to inconsistency between states of the system at time when an observation packet is generated and when the corresponding control packet is received. In order to solve this problem, we propose a cross-layer design of an LQG (Linear Quadratic Gaussian) controller, which specializes in TDMA (Time Division Multiple Access)-based WNCS. In the proposed method, control packets are designed by using a property of TDMA that a sequence of round-trip delays is deterministic. When the controller receives an observation packet, it generates a control packet based on the estimated state of the controlled object at time when the control packet will reach the controlled object. Simulation results show that the proposed method achieves better performance in terms of RMSE (Root Mean Square Error) of the controlled object's state to their targets.

We propose a probabilistic store-carry-forward message delivery scheme based on node density estimation. In our scheme, when a node with a message copy encounters another node, the former forwards its copy to the latter with a certain probability. The forwarding probability is determined depending on a node density at the contact location where two nodes encounter. More specifically, when the node density is high, the forwarding probability is set to be low. This policy is designed to avoid excess message copy transmissions in a high node-density area. In general, nodes frequently encounter each other in high node-density areas and message copies rapidly spread over the nodes. In order to determine whether the node density is high or not, each node estimates the node density distribution over the whole network based on the contact location information. The information is collected by each node and exchanged among nodes. With simulation experiments, we evaluate the performance of our scheme in terms of the mean delivery delay and the number of forwarded message copies.

In sparsely populated mobile ad hoc networks (MANETs), mobile nodes are chronically isolated each other and they meet very occasionally. Global clock synchronization among nodes in such networks is a challenging problem because reference clock information cannot be disseminated promptly over nodes due to the lack of stable connections among nodes. In recent years, averaging-based algorithms for distributed global clock synchronization have been studied. In this paper, we conduct the continuous-time analysis of the simplest one, called the simple averaging scheme, where two mobile nodes exchange their local clock times when they meet and adjust their own clocks to the average of them. Through the analysis and simulation experiments, we reveal how the clock accuracy of nodes and meeting rates among them affect the rate of convergence to the steady state and the accuracy of clock synchronization in steady state.

In the IEEE 802.11 MAC protocol, access points (APs) are given the same authority as wireless terminals (WTs) in terms of acquiring the wireless link, even though they aggregate several downlink flows. This feature leads to a serious throughput degradation of downlink flows, compared with uplink flows. In this paper, we propose a dynamic contention window control scheme for the IEEE 802.11e EDCA-based WLANs, in order to achieve fairness between uplink and downlink TCP flows while guaranteeing QoS requirements for real-time traffic. The proposed scheme first determines the minimum contention window size in the best-effort access category at APs, based on the number of TCP flows. It then determines the minimum and maximum contention window sizes in higher priority access categories, such as voice and video, so as to guarantee QoS requirements for these real-time traffic.

In delay tolerant networks (DTNs), custody transfer mechanism provides reliable end-to-end data transfer in which special nodes (custodians) transfer data with custody in a hop-by-hop manner. As a result, storage congestion occurs when data with custody increases and/or the network is partitioned into multiple sub-networks for a long time. The storage congestion can be alleviated with the help of message ferries. In such a scenario, data should be aggregated to some custodians so that message ferries can effectively collect them. In this paper, we propose a scheme to aggregate data into selected custodians, called aggregators, in a fully distributed and autonomous manner by using evolutionary game theoretical approach where we can also control the number of aggregators to a desired value.

This paper considers (p, q )-Epidemic Routing, a class of store-carry-forward routing schemes, for sparsely populated mobile ad hoc networks. Our forwarding scheme includes Two-Hop Forwarding and the conventional Epidemic Routing as special cases. In such forwarding schemes, the original packet is copied many times and its packet copies spread over the network. Therefore those packet copies should be deleted after a packet reaches the destination. We analyze the performance of (p, q)-Epidemic Routing with VACCINE recovery scheme. Unlike most of the existing studies, we discuss the performance of (p, q)-Epidemic Routing in depth, taking account of the recovery process that deletes unnecessary packets from the network.

This paper considers a fairness issue between uplink and downlink flows in IEEE 802.11 wireless LANs, where uplink flows dominate over downlink flows in terms of wireless bandwidth usage. As a solution to this unfairness, we propose modifying the IEEE 802.11 MAC protocol at access points (APs). Our scheme dynamically controls the minimum contention window size at APs and provides an ample opportunity for them to acquire the transmission right. Note that the proposed scheme does not require any modifications in the MAC protocol at wireless terminals. Furthermore, the optimal minimum contention window size is given by an explicit function of the ratio of the total packet rate of downlink flows to the packet rate of an uplink flow. Through simulation experiments with UDP and TCP flows, we show that the proposed scheme can achieve fairness between uplink and downlink flows.

This paper considers a fairness issue between uplink and downlink flows in IEEE 802.11 wireless LANs, where uplink flows dominate over downlink flows in terms of wireless bandwidth usage. As a solution to this unfairness, the authors propose modifying the IEEE 802.11 MAC protocol at access points (APs). Our scheme dynamically controls the minimum contention window size at APs and provides an ample opportunity for them to acquire the transmission right. Because the control is based only on the number of downlink flows, APs can easily find the optimal window size to achieve per-flow fairness. Note that the proposed scheme does not require any modifications in the MAC protocol at wireless terminals. Through simulation experiments with UDP flows, the authors show that the proposed scheme can achieve fairness between uplink and downlink flows.

This paper proposes a contention-free burst scheduling scheme in optically burst-switched WDM networks. We construct contention-free wavelength planes (lambda-planes) by assigning dedicated wavelengths to each ingress node. Bursts are transmitted to their egress nodes on lambda-planes, along routes forming a spanning tree. As a result, contention at intermediate core nodes is completely eliminated, and contention at ingress nodes is resolved by means of electric buffer. This paper develops a spanning tree construction algorithm, aiming at balancing input loads among output ports at each ingress node. Further a wavelength assignment algorithm is proposed, which is based on the amount of traffic lost at ingress nodes. We show that the proposed scheme can decrease the burst loss probability drastically, even if traffic intensities at ingress nodes are different.

This paper proposes an analytical model for a delayed reservation scheme, which can be exactly analyzed. By applying it to a server-based QoS management network, we can obtain the related blocking probability and the waiting time distribution of requests, and discuss its performance by means of numerical results. The numerical results show that the delayed scheme can improve the blocking probability by about two orders of magnitude compared with the conventional reservation scheme even with a small acceptable waiting time (e.g. 20% of average flow duration). We can conclude that the delayed reservation scheme can lead to significant improvement in a server-based QoS management network performance.

In the Internet, because of huge scale and distributed administration, it is of practical importance to infer network-internal characteristics that cannot be measured directly. We propose a general method of determining characteristics of links from given characteristics of end-to-end paths. Our method can be applied to an arbitrary path-topology. Furthermore we show the general conditions that the link characteristics to be inferred must satisfy. Packet loss and queuing delay time are shown to satisfy them. Case studies which our method can treat are also provided.

We examine the delay performance of packets from constant-bit-rate (CBR) traffic whose delay is affected by non-real-time traffic. The delay performance is analyzed by solving the nD/D/1 queue with vacations. We obtain an exact and closed form solution, hence obviating the need of any approximations or numerical Laplace inversions. We then provide various numerical results for low-bit-rate transmission links, in which packets can experience large delay. From our quantitative evaluation, we conclude that there exists an optimum packet size for a given delay bound. In extremely slow links, such as modem links, transmission control protocol (TCP) packets should be segmented to reduce the CBR delay. We therefore investigate the delay impact of TCP packet sizes as well.

For ATM networks we derive the bounds formulas for the asymptotic tail distribution in a queue whose arrival process is a superposition of general periodic Markov sources. Note that we make no assumption for the structure of the general periodic Markov sources except that the underlying Markov chains are irreducible. The periodic source model in this paper is thus rather general. Taking initial state conditions of the periodic sources into account, we construct a superposed arrival process. Contrary to the previous works, this implies that the periodic sources are not independent. We also provide examples to investigate the efficiency and the accuracy of our bound formulas.

In transport-layer protocols such as TCP over ATM networks, a packet is discarded when one or more cells of that packet are lost, and the destination node then requires its source to retransmit the corrupted packet. Therefore, once one of the cells constituting a packet is lost, its subsequent cells of the corrupted packet waste network resources. Thus, discarding those cells will enable us to efficiently utilize network resources, and will improve the packet loss probability. We focus on tail dropping (TD) and early packet discard (EPD) as selective cell discard schemes which enforce the switches to discard some of the arriving cells instead of relaying them. We exactly analyze the packet loss probability in a system applying these schemes. Their advantages and limits are then discussed based on numerical results derived through the analysis.

In transport-layer protocols such as TCP over ATM networks, a packet is discarded when one or more cells are lost in that packet, and the destination node then requires its source to retransmit the corrupted packet. Therefore, once one of the cells constituting a packet is lost, the subsequent cells of the corrupted packet waste network resources. Thus, discarding those cells will enable us to efficiently utilize the network resources and improve the packet loss probability. We focus on tail dropping (TD) and early packet discard (EPD) as selective cell discard schemes which force the switches to discard some of the arriving cells instead of relaying them. We exactly analyze the packet loss probability in a system applying these schemes. Their advantage and limits are then discussed based on numerical results derived through the analysis.

The FRP (fast reservation protocol) utilizes a unique feature of ATM (asynchronous transfer mode) technology. The FRP is supposed mainly to be applied to LAN interconnection. We provide an exact analysis for a class of the FRP to obtain the burst (a protocol data unit of FRP) level performance of the FRP. One of the main features of the ATM is that the transmission rate can be adjusted according to the network congestion states. In our modeling, this feature is incorporated in such a way that the bandwidth a source requests is reduced if an attempt to reserve the bandwidth is rejected by the network. A rationale behind this is that rejection of the bandwidth request indicates network congestion. Therefore, the request with the smaller bandwidth after the reservation failure enables better sharing of network resources, which results in a performance improvement. To solve our model, which contains a very huge number of system states, we introduce a new numerical approach which is an extension of Sumita and Rieders's (1991) replacement process approach. Through numerical examples, the appropriate transmission rate control method is examined, and considerations on backoff time and overhead of RM cells are also presented.

This paper analyzes an ATM switching node with correlated cell arrivals. An ATM switching node is modeled as a discrete-time finite-buffer queue. Cells arrive according to a semi-Markovian process, where the number of cell arrivals in a given slot depends on the state of the underlying (M-state) Markov chain in both the given slot and the previous slot. This paper presents analysis for various cell loss characteristics such as cell loss probability, the consecutive loss probability, and the distribution of loss period lengths. It also obtains characteristics of the cell output process from an ATM switching node including the joint distribution of successive cell interdeparture times, and the distributions of busy and idle periods of the cell output process. It is shown through numerical examples that both the correlation and the variation of cell arrivals significantly affect the cell loss and the output process characteristics.>

In almost all conventional transport protocols, ARQs (automatic repeat request) have been used to recover lost packets, in which receivers would request sources to retransmit lost packets. In these retransmission based error recovery schemes, the transmission delay can reach an unacceptable level for traffic with stringent delay constraints in high speed networks because their propagation delay is very large relative to packet transmission times. On the other hand, FEC (forward error correction) adds redundant packets to data ones, all of which are organized into a group called a block, and hence allows the receivers to correct some of the lost packets without retransmissions. This capability is desirable for high speed networks. The authors exactly analyze the distribution of the number of cells lost within a block in ATM networks with correlated cell arrivals, and examine the benefits due to FEC by means of the analysis.>

One purpose of a traffic shaping mechanism (such as the leaky bucket) is to smooth out an input process into an output process. In so doing, a question of interest is the following: to what extent has the input process been modified, or, to what extent does the output process conform to the input process? To this end, the authors define and characterize a conformance measure. They apply this idea to a leaky bucket in a high-speed network in which this issue may arise naturally. They show how to find the conformance measure for a leaky bucket with the Markovian arrival process.>

In broadband ISDN, different classes of traffic expect to receive different quality of service. One way of providing service is to implement a priority structure among traffic classes. The authors analyze a single server queue in which video and voice traffic receive priority over data traffic. Corresponding to the fixed cell size of ATM, they assume that service times are deterministic. They further assume that the high priority traffic is correlated and the arrivals are governed by a Markov chain. With these assumptions, they characterize the queue length distributions and the waiting time distributions for this problem. They show by numerical examples that the delay distribution depends highly on the correlations and that a priority service discipline is capable of providing very good service for real time traffic.>

The performance of an asynchronous transfer mode (ATM) switching node is analyzed, taking cell arrival correlation into consideration. An ATM switching node is modeled as a discrete-time finite-buffer queue. Cell arrivals are assumed to follow a semi Markovian process. Various characteristics of the cell loss and the distribution function of the cell output process from an ATM switching node are analyzed. The cell loss probability, the consecutive loss probability, the distribution of loss period lengths, the joint distribution of successive cell interdeparture times, and the distributions of busy and idle periods are determined. It is shown that both the correlation and the variation of cell arrivals significantly affect the cell loss and the output process characteristics.>

The paper provides a performance analysis of automatic-repeat-request (ARQ) protocols in connection-oriented transmission. Each message arriving at a transmitter is divided into several packets, which are continuously transmitted to a receiver according to go-back-N ARQ or selective-repeat ARQ protocols. Because of connection-oriented transmission, messages are served on a first come first served basis, i.e., transmission of a message is commenced after all packets in the previous message are successfully transmitted. For the two ARQ protocols, the authors derive the probability generating functions of message waiting time and queue length at an arbitrary instant, from which the average performance measures are explicitly obtained.>

A continuous-time carrier-sense multiple access with collision detection (CSMA/CD) system with a finite number of homogeneous stations, each possessing an infinite buffer, is analyzed. The system and each station is treated approximately as an independent M/G/1 queuing system. Since stations are mutually dependent, the interference from other stations is taken into account, in terms of the service-time distribution. With this analysis, the mean message delay can be numerically obtained. The stability of the system is also discussed. In addition, the upper-bound condition for the transmission interval is presented that keeps the system stable. It is concluded that the stability of the system becomes more sensitive to the retransmission interval as the number of stations increases.>

A unified approach to general asymmetric polling systems with a single buffer at each station is presented. Two variations of single-buffer polling system are considered: the conventional system and the buffer-relaxation system. In the conventional system, a new massage is not allowed to queue until the previous message has been completely transmitted, while in the buffer-relaxation system, a newly arriving message can be stored in the buffer once the previous message's transmission has started. For each system, the Laplace-Stieltjies transform (LST) of the joint probability-distribution function (PDF) of station times is derived, from which the LST of the PDF of message delay is straightforwardly obtained.>

We first consider the interdeparture process of a symmetric polling system with single buffer at each station, where distributions of transmission time and switchover time are assumed to be general. Under the assumption that consecutive interdeparture times are independent, the probability distribution is analyzed and the LaplaceStieltjes transform is given. Numerical examples show that the coefficient of variation of interdeparture time is not monotonic with respect to the arrival rate. Based upon the analysis, we can obtain the characteristics of outbound traffic if we are concerned with unidirectional traffic flow. In general, however, there is inbound traffic from external networks in the interconnected communication system and its effect on performance characteristics should be taken into account. We next propose the model of a polling system having the traffic from outside the system. The model consists of homogeneous stations with single buffers and a depot station with infinite buffer where inbound traffic is dealt with. We analyze the model and obtain throughput and mean waiting time for the case of constant transmission time and constant switchover times. Numerical examples show that the throughput of the system is not sensitive to inbound traffic if the total load of the system is moderate.