Caches at CPU nodes in disaggregated memory architectures amortize the high data access latency over the network. However, such caches are fundamentally unable to improve performance for workloads requiring pointer traversals across linked data structures. We argue for accelerating these pointer traversals closer to disaggregated memory in a manner that preserves expressiveness for supporting various linked structures, ensures energy efficiency and performance, and supports distributed execution. We design pulse, a distributed pointer-traversal framework for rack-scale disaggregated memory to meet all the above requirements. Our evaluation of pulse shows that it enables low-latency, high-throughput, and energy-efficient execution for a wide range of pointer traversal workloads on disaggregated memory that fare poorly with caching alone.

Emerging large language model (LLM) applications involve diverse reasoning strategies and agentic workflows, straining the capabilities of existing serving systems built on a monolithic token generation loop. This paper introduces Pie, a programmable LLM serving system designed for flexibility and efficiency. Pie decomposes the traditional generation loop into fine-grained service handlers exposed via an API and delegates control of the generation process to user-provided programs, called inferlets. This enables applications to implement new KV cache strategies, bespoke generation logic, and seamlessly integrate computation and I/O—entirely within the application, without requiring modifications to the serving system. Pie executes inferlets using WebAssembly, benefiting from its lightweight sandboxing. Our evaluation shows Pie matches state-of-the-art performance on standard tasks (3-12% latency overhead) while significantly improving latency and throughput (1.3×-3.4× higher) on agentic workflows by enabling application-specific optimizations.

We address the problem of fair resource allocation in multiuser remote memory systems. Allocating local memory (used as cache) and network bandwidth to remote memory in such systems is challenging due to the complex interdependence between the two resources and application performance. A larger cache may reduce the need for fetching data over the network, while a larger bandwidth may permit more concurrent network requests, avoiding the need for large caches. As a result, applications can achieve the same data access throughput for a wide range of cache and bandwidth allocations. Such interdependence is unique to each application and hard to capture offline.

Page-based far memory systems transparently expand an application's memory capacity beyond a single machine without modifying application code. However, existing systems are tailored to scenarios with low application thread counts, and fail to scale on today's multi-core machines. This makes them unsuitable for data-intensive applications that both rely on far memory support and scale with increasing thread count. Our analysis reveals that this poor scalability stems from inefficient holistic coordination between page fault-in and eviction operations. As thread count increases, current systems encounter scalability bottlenecks in TLB shootdowns, page accounting, and memory allocation.

Owing to remarkable improvements in deep neural networks (DNNs), various computation-intensive and delay-sensitive DNN services have been developed for smart IoT devices. However, employing these services on the devices is challenging due to their limited battery capacity and computational constraints. Although edge computing is proposed as a solution, edge devices cannot meet the performance requirements of DNN services because the majority of IoT applications require simultaneous inference services, and DNN models grow larger. To address this problem, we propose a framework that enables parallel execution of partitioned and offloaded DNN inference services over multiple distributed edge devices. Noteworthy, edge devices are reluctant to process tasks due to their energy consumption. Thus, to provide an incentive mechanism for edge devices, we model the interaction between the edge devices and DNN inference service users as a two-level Stackelberg game. Based on this model, we design the proposed framework to determine the optimal scheduling with a partitioning strategy, aiming to maximize user satisfaction while incentivizing the participation of edge devices. We further derive the Nash equilibrium points in the two levels. The simulation results show that the proposed scheme outperforms other benchmark methods in terms of user satisfaction and profits of edge devices.

The evolution of the Internet of Things (IoT) has been driving the explosive growth of deep neural network (DNN)-based applications and processing demands. Hence, edge computing has emerged as a potential solution to meet these processing requirements. However, emerging IoT applications have increasingly demanded to run multiple DNNs to extract multifaceted knowledge, requiring more computational resources and increasing response time. Consequently, edge nodes cannot act as a complete substitute for the previous cloud paradigm, owing to their relatively limited resources. To address this problem, we propose to incorporate nearby IoT devices when allocating resources to multiple DNN models. Furthermore, the optimization of resource allocation can be hindered by the heterogeneity of IoT devices, which affects the delay performance of DNN-based computing. In this context, we propose a DNN partition placement and resource allocation strategy that considers different processing powers, memory, and battery levels for heterogeneous IoT devices. We evaluate the performance of the proposed strategy through extensive simulations. Simulation results reveal that the proposed strategy outperforms other existing solutions in terms of end-to-end delay, service probability, and energy consumption. The proposed solution was further simulated in a Kubernetes testbed consisting of actual devices to assess its feasibility.

Memory disaggregation promises transparent elasticity, high resource utilization and hardware heterogeneity in data centers by physically separating memory and compute into network-attached resource "blades". However, existing designs achieve performance at the cost of resource elasticity, restricting memory sharing to a single compute blade to avoid costly memory coherence traffic over the network.

At present, wireless home routers are becoming increasingly smart. While these smart routers provide rich functionalities to users, they also raise security concerns. Although the existing end-to-end encryption techniques can be applied to protect personal data, such rich functionalities become unavailable due to the encrypted payloads. On the other hand, if the smart home routers are allowed to process and store the personal data of users, once compromised, the users' sensitive data will be exposed. As a consequence, users face a difficult trade-off between the benefits of the rich functionalities and potential privacy risks. To deal with this dilemma, we propose a novel system named Secure and Smart Network (S2Net) for home routers. For S2Net, we propose a secure OS that can distinguish and manage multiple sessions belonging to different users. The secure OS and all the router applications are placed in the secure world using the ARM TrustZone technology. In S2Net, we also confine the router applications in sandboxes provided by the proposed secure OS to prevent data leakage. As a result, S2Net can provide rich functionalities for users while preserving strong privacy for home routers. In addition, we develop a crypto-worker model that provides an abstraction layer of cryptographic tasks performed by a heterogeneous multi-core system. The other important role of crypto-worker is to parallelize the computations in order to resolve the high computation cost of cryptographic functions. We report the system design of S2Net and the details of our implementation. Experimental results with benchmarks and real applications demonstrate that our implementation is capable of achieving high performance in terms of throughput while mitigating the overhead of S2Net design.

Internet of Vehicles (IoV) has emerged as a key component of smart cities. Connected vehicles are increasingly processing real-time data to respond immediately to user requests. However, the data must be sent to a remote cloud for processing. To resolve this issue, vehicular fog computing (VFC) has emerged as a promising paradigm that improves the quality of computation experiences for vehicles by offloading computation tasks from the cloud to network edges. Nevertheless, due to the resource restrictions of fog computing, only a limited number of vehicles are able to use it while it is still challenging to provide real-time responses for vehicular applications, such as traffic and accident warnings in the highly dynamic IoV environment. Therefore, in this article, we formulate the problem of allocating the limited fog resources to vehicular applications such that the service latency is minimized, by utilizing parked vehicles. We then propose a heuristic algorithm to efficiently find the solutions of the problem formulation. In addition, the proposed algorithm is combined with reinforcement learning to make more efficient resource allocation decisions, leveraging the vehicles' movement and parking status collected from the smart environment of the city. Our simulation results show that our VFC resource allocation algorithm can achieve higher service satisfaction compared to conventional resource allocation algorithms.

Since Internet of Vehicles (IoV) generate and consume huge amount of data, caching becomes indispensable technique to provide better Internet services in IoV. However, deployment and operation of infrastructure to cache the massive vehicular data is very costly. To tackle this problem, we propose a vehicular caching scheme that reduces data delivery delay and cost using parked vehicles.

To meet the rising demand of mobile data, IEEE 802.11ac was recently developed to offer better network performance than previous IEEE 802.11 standards, representing the 5G Wi-Fi technology. One of the new technologies added to IEEE 802.11ac is a dynamic bandwidth channel access scheme that supports channel bonding (CB) with an up to 160 MHz bandwidth. However, the use of CB increases the number of stations sharing the same medium, resulting in an increased collision rate and severe interference caused by overlapping sub-channels. This problem is further aggravated in densely deployed wireless networks where the service ranges of access points (APs) are likely to be highly overlapping. To resolve these issues, we present an analytical throughput model that considers the effect of both collisions and interference. With this analytical model, we formulate the CB problem to optimize the throughput along with the traffic demand and propose a genetic algorithm for solving the problem. Further, a dynamic channel bonding algorithm (named DCB) is proposed to solve the problem with lower complexity. Extensive simulation results demonstrate that our proposed algorithm provides a higher performance than other recent dynamic CB algorithms with respect to the throughput achieved for the traffic demand.

Machine-to-machine (M2M) communication is an up-and-coming technology for Internet of Things. However, current wireless cellular network such as LTE-A that is originally designed for human-to-human (H2H) communication falls short in supporting massive number of bursty random access (RA) requests from machine type communication (MTC) devices. Specifically, random access network (RAN) overload due to the huge amount of MTC devices is a critical issue. In this paper, we propose a Dynamic Extended Access Barring (DEAB) scheme to regulate the number of RA requests to mitigate the RAN overload. By utilizing the direct communication between MTC devices in M2M, we exploit RA success rates of neighbour devices in choosing the optimal Extended Access Barring (EAB) parameters. The DEAB scheme is evaluated against EAB scheme and the simulation results show that the proposed scheme significantly outperforms the EAB.

Skiing is one of the most popular winter sports in the world. Even though the equipment has been improved for preventing injuries during skiing, injury risks of the lower extremity are still high. It is necessary to investigate injury risk by motion analysis during skiing. The wearable motion capture system consisting of inertial sensors and insole pressure sensors can be utilized due to restrictions of conventional optical motion capture system. In this study, the motions for short-and middle-turns during skiing were analyzed using a wearable motion capture system and the multi-scale computer simulation technology. Seven male certified ski coaches participated in this study and their full body motion and foot pressure data were simultaneously recorded by the wearable motion capture system. Joint kinematics and kinetics in the hip, knee and ankle of the right lower extremity were analyzed during short- and middle-turns. Even though a slight difference in the joint kinematics between two turns was predicted, the joint forces and moments for middle-turn were higher than those for short-turn. Because these higher joint forces and moments can result in osteoarthritis or ligament injury at the joint, injury risk of the joint for middle-turn may be higher than that for short-turn. This study confirms that the wearable motion capture system is useful for measuring the motion and plantar pressure data in outdoor sports such as skiing.

The development of telemonitoring via wireless body area networks (WBANs) is an evolving direction in personalized medicine and home-based mobile health. A WBAN consists of small, intelligent medical sensors which collect physiological parameters such as electrocardiogram, electroencephalography, and blood pressure. The recorded physiological signals are sent to a coordinator via wireless technologies, and are then transmitted to a healthcare monitoring center. One of the most widely used wireless technologies in WBANs is ZigBee because it is targeted at applications that require a low data rate and long battery life. However, ZigBee-based WBANs face severe interference problems in the presence of WiFi networks. This problem is caused by the fact that most ZigBee channels overlap with WiFi channels, severely affecting the ability of healthcare monitoring systems to guarantee reliable delivery of physiological signals. To solve this problem, we have developed an algorithm that controls the load in WiFi networks to guarantee the delay requirement for physiological signals, especially for emergency messages, in environments with coexistence of ZigBee-based WBAN and WiFi. Since WiFi applications generate traffic with different delay requirements, we focus only on WiFi traffic that does not have stringent timing requirements. In this paper, therefore, we propose an adaptive load control algorithm for ZigBee-based WBAN/WiFi coexistence environments, with the aim of guaranteeing that the delay experienced by ZigBee sensors does not exceed a maximally tolerable period of time. Simulation results show that our proposed algorithm guarantees the delay performance of ZigBee-based WBANs by mitigating the effects of WiFi interference in various scenarios.