Graph Neural Networks (GNNs) have emerged as powerful tools for learning from graph-structured data and are widely used in various applications such as traffic prediction, and Electronic Design Automation, among others. However, training GNNs on large-scale graphs with hundreds of millions of nodes and billions of edges is time-consuming, often requiring days or weeks on a single machine. While distributed GNN training across multiple machines has been explored to leverage greater computation and memory resources, existing approaches primarily focus on inter-machine scalability while overlooking intra-machine scalability across multiple cores. As a result, state-of-the-art distributed GNN training frameworks lead to severe resource underutilization and limited training performance in terms of epoch time. This work introduces ARGO+, a novel GNN system designed to achieve multi-level scalability for distributed GNN training by efficiently scaling across both inter- and intra-machine levels. ARGO+ features a two-level graph partitioning strategy, exploiting parallelisms across both graph topology and feature dimensions while minimizing communication overhead and ensuring balanced workloads. In addition, ARGO+ integrates NUMA-aware optimizations to enhance intra-node scalability, addressing inefficiencies in remote-socket data accessing. During runtime, ARGO+ adopts a two-stage parallel training scheme to further reduce communication overheads and instantiates multiple training processes to exploit computation-communication overlapping. We evaluate ARGO+ on a distributed multi-core CPU cluster consisting of 8 machines, each with a dual-socket 80-core Intel Xeon processor. Our results demonstrate that ARGO+ achieves up to 1.89× speedup compared with DistDGL, and 1.53-1.56× speedup compared with state-of-the-art distributed GNN training systems. In addition, ARGO+ is compatible with the Deep Graph Library (DGL), a widely used GNN framework, enabling seamless integration with existing GNN programs. Finally, while ARGO+ adopts various optimizations to improve scalability and training performance, these optimizations do not alter the semantics of the training algorithm; thus, the model accuracy and convergence remain consistent with the original implementation of the algorithm.

Recently, Graph Neural Networks (GNN) have been integrated into various local applications, such as local community detection and local code assistant, making edge inference increasingly important. To support diverse workloads, state-of-the-art edge devices have evolved into heterogeneous platforms, integrating components like CPU, GPU, and NPU. To this end, we propose GNX, a novel GNN system that accelerates GNN inference on edge heterogeneous platforms by leveraging all the heterogeneous processing units. Given a GNN model and a heterogeneous platform, GNX automatically generates parallel execution plans, consisting of both data and pipeline parallelism. To reduce the complexity of the design space, GNX converts GNN models into coarse-grained blocks and performs the search at the block level. By leveraging the APIs provided by state-of-the-art heterogeneous frameworks, GNX can flexibly schedule various parallel execution plans and seamlessly adjust the workload across the heterogeneous processing units for load-balanced execution. Our study shows that GNX effectively accelerates three widely-used GNN models on two state-of-the-art edge heterogeneous platforms. Compared with the baseline approach that uses only a single processing unit, GNX achieves up to a 2.57× speedup. Compared with adopting data parallelism and a state-of-the-art scheduler, GNX achieves up to 1.90× and 1.79× speedup, respectively. We also discuss the applicability of and extensions to GNX to support other GNN models.

As Graph Neural Networks (GNNs) become popular, libraries like PyTorch-Geometric (PyG) and Deep Graph Library (DGL) are proposed; these libraries have emerged as the de facto standard for implementing GNNs because they provide graph-oriented APIs and are purposefully designed to manage the inherent sparsity and irregularity in graph structures. However, these libraries show poor scalability on multi-core processors, which under-utilizes the available platform resources and limits the performance. This is because GNN training is a resource-intensive workload with high volume of irregular data accessing, and existing libraries fail to utilize the memory bandwidth efficiently. To address this challenge, we propose ARGO, a novel runtime system for GNN training that offers scalable performance. ARGO exploits multi-processing and core-binding techniques to improve platform resource utilization. We further develop an auto-tuner that searches for the optimal configuration for multi-processing and core-binding. The auto-tuner works automatically, making it completely transparent from the user. Furthermore, the auto-tuner allows ARGO to adapt to various platforms, GNN models, datasets, etc. We evaluate ARGO on two representative GNN models and four widely-used datasets on two platforms. With the proposed autotuner, ARGO is able to select a near-optimal configuration by exploring only 5% of the design space. ARGO speeds up state-of-the-art GNN libraries by up to 5.06× and 4.54× on a four-socket Ice Lake machine with 112 cores and a two-socket Sapphire Rapids machine with 64 cores, respectively. Finally, ARGO can seamlessly integrate into widely-used GNN libraries (e.g., DGL, PyG) with few lines of code and speed up GNN training.

Software switch has emerged as a critical component in software defined networking and network virtualization areas. Open vSwitch (OvS) is a widely used software switch which uses tuple space search algorithm for packet classification, and an exact match cache (EMC) for caching most frequently used flows. In this paper, we propose two new optimizations for OvS to further improve its performance and scalability. First aims to completely remove the sequential search overhead of the tuple space search layer of OvS, and second is a dynamic cache insertion optimization for the EMC to improve EMC effectiveness. We show that the optimizations can improve OvS's throughput by up to 3.5X for millions of active flows.

Network function virtualization (NFV) promises significant cost savings, flexibility and ease of deployment. However, potential challenges in implementing virtualized network elements that can support real-world performance requirements are still an open question. For example, traditional telecom networks have a lot of complex interdependencies that can affect performance. In this paper, we study the potential bottlenecks in virtualizing cellular core network functions. Using a combination of analysis and experimentation, we quantify the impact of software-based EPC elements on various metrics including physical processing, memory, IO, and bandwidth resource requirements. We use production grade, software-based cellular network elements running on general purpose Linux servers, driven by a variety of realistic workloads derived from a realworld cellular network, to examine the combined effects of control and data planes on an LTE enhanced packet core (EPC). In particular, we discover that the SGW handles about 33% of the control plane transactions and is a potential source for performance bottlenecks as a result of the interdependencies between control and data plane processing. Our results indicate that simply replacing existing EPC elements with virtualized equivalents can have severe performance bottlenecks and that virtualized EPC elements need to be carefully designed.

The current trend to move from homogeneous to heterogeneous multicore systems provides compelling opportunities for achieving performance and energy efficiency goals. Running multiple threads in multicore systems poses challenges on meeting limited shared resources, such as memory bandwidth. We propose an optimization approach that includes an Integer Linear Programming (ILP) optimization model and a scheme to dynamically determine thread-to-core assignment. We present simulation analysis that shows energy savings and performance gains for a variety of workloads compared to state-of-the-art schemes. We implemented and evaluated a prototype of our thread assignment approach at user level, leveraging Linux scheduling and performance-monitoring capabilities.

The current trend to move from homogeneous to heterogeneous multi-core systems promises further performance and energy-efficiency benefits. A typical future heterogeneous multi-core system includes two distinct types of cores, such as high performance sophisticated ("large'') cores and simple low-power ("small'') cores. In those heterogeneous platforms, execution phases of application threads that are CPU-intensive can take best advantage of large cores, whereas I/O or memory intensive execution phases are best suited and assigned to small cores. However, it is crucial that the assignment of threads to cores satisfy both the computational and memory bandwidth constraints of the threads. We propose an optimization approach to determine and apply the most energy efficient assignment of threads with soft real-time performance and memory bandwidth constraints in a multi-core system. Our approach includes an ILP (Integer Linear Programming) optimization model and a scheme to dynamically change thread-to-core assignment, since thread execution phases may change over time. In comparison to state-of-art dynamic thread assignment schemes, we show energy savings and performance gains for a variety of workloads, while respecting thread performance and memory bandwidth requirements.

Today's platforms offer ubiquitous network connectivity through one or more communication interfaces and while the communication devices consume a small portion of the overall platform power the impact of network connectivity and individual packet processing on the overall platform energy consumption is significant, due to the non-deterministic characteristics of network traffic. In this paper we show that significant energy is unnecessarily wasted at the platform level while it is idle and connected to one or more networks, especially unlicensed networks like WiFi or Ethernet, because the system is constantly processing what we term "background/noise" traffic. We quantify the negative impact of network connectivity on platform energy-efficiency and provide novel techniques to mitigate this impact. We implemented these mitigation techniques in our wireless network interface cards and our results indicate that when our scheme is used the total platform sleep time increases by up to 40% with no performance degradation and no user experience impact.

The goal of Wireless Sensor Networks is to extract useful global information from individual sensor readings, which are typically collected and aggregated over a spanning tree. However, the spanning tree structure is not robust against communication errors; a low-quality (i.e., high-error-rate) wireless link close to the tree root may result in a high rate of global information loss. Therefore, many schemes have been proposed to achieve fault-tolerant aggregation. Intuitively, using timely link-quality information, which is gathered by continuous monitoring and error-rate measurement of network links, improves the performance of fault-tolerant aggregation schemes. In this paper, we show that this intuition is not always true. In particular, we show that using link-quality information in an intuitive but wrong way results in degraded performance in some schemes, and therefore, care should be taken in using link-quality information. We also show that some schemes make better usage of link-quality information than others, and some schemes are more robust to errors in link-quality estimation than others. We support our findings by an extensive simulation study, and we focus on the (more general) class of fault-tolerant duplicate-sensitive aggregation schemes.

TDMA has been proposed as a MAC protocol for wireless sensor networks (WSNs) due to its efficiency in high WSN load. However, TDMA is plagued with shortcomings; we present modifications to TDMA that will allow for the same efficiency of TDMA, while allowing the network to conserve energy during times of low load (when there is no activity being detected). Recognizing that aggregation plays an essential role in WSNs, TDMA-ASAP adds to TDMA: (a) transmission parallelism based on a level-by-level localized graph-coloring, (b) appropriate sleeping between transmissions ("napping"), (c) judicious and controlled TDMA slot stealing to avoid empty slots to be unused and (d) intelligent scheduling/ordering transmissions. Our results show that TDMA-ASAP's unique combination of TDMA, slot-stealing, napping, and message aggregation significantly outperforms other hybrid WSN MAC algorithms and has a performance that is close to optimal in terms of energy consumption and overall delay.

Denial-of-service (DoS) attacks remain a challenging problem in the Internet. By making resources unavailable to intended legitimate clients, DoS attacks have resulted in significant loss of time and money for many organizations, thus, many DoS defense mechanisms have been proposed. In this paper we propose live baiting, a novel approach for detecting the identities of DoS attackers. Live baiting leverages group-testing theory, which aims at discovering defective members in a population using the minimum number of dasiadasiatestspsilapsila. This leverage allows live baiting to detect attackers using low state overhead without requiring models of legitimate requests nor anomalous behavior. The amount of state needed by live baiting is in the order of number of attackers not number of clients. This saving allows live baiting to scale to large services with millions of clients. We analyzed the coverage, effectiveness (detection time, false positive and false negative probabilities), and efficiency (memory, message overhead, and computational complexity) of our approach. We validated our analysis using NS-2 simulations modeled after real Web traces.

In wireless sensor networks (WSN) node failures are typically detected using a heartbeat application, where a neighbor detects a failed node when it misses successive short messages (ldquoheartbeatsrdquo) that should have been sent by the failed node. However, wireless links are usually lossy, hence, to distinguish node failures from intermittent link failures, the threshold on the number of missed heartbeats is usually set to a large number, incurring in a long delay for declaring a node dead. In this paper we present ldquoGroupBeatrdquo an accurate node failure detection system for WSN and propose the ldquoCommunication By Signalingrdquo scheme as an energy-efficient low-overhead implementation of GroupBeat.

In this paper, the authors motivate the interesting characteristics of small-to-medium enterprises and study the problem of wireless IP telephony for such deployments. While small enterprises are characterized by few network nodes and small deployment areas, they also demand good IP telephony quality, ease of deployment, and low management and configuration overhead. The authors evaluate the feasibility of peer-to-peer telephony over a multi-hop wireless ad-hoc network for such small enterprises. The authors designed a system and built a prototype using off-the-shelf components. In this paper, the authors describe the implementation and deployment of this system. Several fundamental routing protocol issues needed to be resolved to develop a stable system and the authors present them here. The authors also report the results of a simulation study they conducted to establish the configuration and performance parameters for such systems, and identify the limits to the use of multi-hop ad-hoc wireless for peer-to-peer IP telephony for small enterprises.

In wireless mobile adhoc networks (MANETs) channel and energy capacities are scarce resources, a lot of energy-efficient routing protocols for MANETs have been previously proposed to take into consideration the nodes’ residual energies when establishing routes between source-destination pairs. In this paper we are not trying to introduce a new routing algorithm to be added to the already proposed stack of energy-efficient protocols, but rather, we identify a problem in cost-based energy-efficient routing for MANETs, we call this problem "Flooding Waves". We show that the "Flooding Waves" is a serious problem in dense networks, to the extent that the excessive energy overhead consumed in these waves can outweigh the gain achieved by energy-efficient path selection. We propose the "Delayed-Forwarding" as a solution for this problem. We provide both a simulation analysis and a simple theoretical framework to validate and support this solution.

In wireless sensor networks (WSNs), the users' objective is to extract useful global information by collecting individual sensor readings. Conventionally, this is done using in-network aggregation on a spanning tree from sensors to data sink. However, the spanning tree structure is not robust against communication errors; when a packet is lost, so is a complete subtree of values. Multipath routing can mask some of these errors, but on the other hand, may aggregate individual sensor values multiple times. This may produce erroneous results when dealing with duplicate-sensitive aggregates, such as SUM, COUNT, and AVERAGE. In this paper, we present and analyze two new fault tolerant schemes for duplicate-sensitive aggregation in WSNs: (1) cascaded ridesharing and (2) diffused ridesharing. These schemes use the available path redundancy in the WSN to deliver a correct aggregate result to the data sink. Compared to state-of-the-art, our schemes deliver results with lower root mean square (RMS) error and consume much less energy and bandwidth. RideSharing can consume as much as 50% less resources than hash-based schemes, such as SKETCHES and synopsis diffusion, while achieving lower RMS for reasonable link error rates

In wireless adhoc networks, channel and energy capacities are scarce resources. However, the design of the IEEE 802.11 DCF protocol leads to an inefficient utilization of these resources. We introduce BLAM, a new battery level aware MAC protocol, which is developed from an energy-efficiency point of view to extend the useful lifetime of an adhoc network. We modify the IEEE 802.11 DCF protocol to enable BLAM to tune the random deferring time for fresh and collided data packets dynamically, based on the node's energy. We show that BLAM can achieve an increase of 15% in network lifetime and an increase of about 35% in the total number of received packets.

In this paper we address the issue of controlling transmission power in power-aware ad hoc networks. Previous work that minimizes the transmission power does not consider both the energy consumed in collision resolution and the energy disbursed to overcome the interference resulting from neighboring nodes. We investigate the basic transmission power control for the 802.11 MAC protocols in which the control frames and the data frames can be transmitted at different power levels. A collision model together with an interference model of a uniformly distributed network is constructed. Based on these models, the end-to-end network throughput and the total energy consumption of the network are examined. For a network with a given node density, our results show the optimal transmission power for control messages and for data messages that will yield maximum throughput and minimum energy consumption per message.

Wireless networks have a constraint on their functional lifetime. This is due to the limited energy capacity of batteries powering the wireless nodes. For extending the lifetime of such battery-operated networks, we present a scheme for dynamically selecting the transmission rate for each node in the network. The transmission rate is based on the available energy budget in each node's battery. The goal is to increase the network capability of delivering more packets. The rate selection for each node is subject to satisfying a QoS timing constraint on the packet delivery time. Through adaptively varying each node's rate, we extended the lifetime 10 times on average more transmitting at a maximum rate and delivered on average 7.5 times more data packets. When compared with a scheme that transmits data at a lower rates independent of the battery levels, our scheme delivers up to 12% more packets for the same available total energy.