Kubernetes has become the dominant platform for orchestrating containerized applications in modern cloud environments. However, its default scheduling mechanisms primarily focus on CPU and memory resources, overlooking network performance–a critical factor for latency-sensitive applications in domains like 5G and edge computing. This paper introduces NAS, a novel network-aware scheduling framework that integrates real-time network metrics such as latency and bandwidth using an eBPF-based Cilium service mesh. By incorporating these metrics into the scheduling process, NAS optimizes pod placement for performance-sensitive applications. Experimental results show that NAS reduces average latency by 52.66% compared to the default Kubernetes scheduler and 2.68% compared to Diktyo while minimizing maximum latency spikes by 85.61% and 7.23%, respectively. Further, NAS effectively distributes workloads and provides co-location benefits by considering microservices dependencies and network costs during pod placement.

Microservices have emerged as a popular pattern for developing large-scale applications in cloud environments for their flexibility, scalability, and agility benefits. Furthermore, orchestration services like Kubernetes have simplified the deployment of cloud-native applications. However, monitoring and debugging these complex networked applications has become increasingly challenging, creating additional overheads. Traditionally, observability or monitoring requires developers to instrument their applications to expose metrics, logs, and traces using language-restricted libraries. This approach does not work well in a multi-tenant cloud environment as it cannot monitor processes or containers that are not instrumented or hidden. A critical challenge is managing complexity by consistently instrumenting multiple microservices across application platforms and programming languages. Hence, there is a need for a low-overhead cloud-native solution that provides complete observability for distributed and containerized environments. eBPF is a Linux VM technology that can instrument the host kernel directly and provides out-of-the-box cloud-native observability with negligible performance overheads. This paper proposes an eBPF-based solution that offers complete observability for cloudnative deployments. Further, we compare the performance and effectiveness of our solution with popular observability agents like Node Exporter and cAdvisor. We show that our proposed solution reduces CPU overheads by up to 210 times while requiring up to 159 % less memory than the alternatives. Lastly, we deploy and test our solution on a Chameleon cloud bare metal testbed.

The landscape of cloud deployments has surged, evolving into a cornerstone for modern infrastructures. However, with this advancement, a key challenge remains: general-purpose CPUs perform infrastructure functions of the cloud deployments, i.e., storage, security, and networking. SmartNICs have emerged as a transformative solution, integrating specialized processing capabilities with the help of processing cores and accelerators directly on the NIC hardware. These intelligent adapters offload and accelerate infrastructure functions, enhancing throughput, reducing latency, and alleviating the burdens on host CPUs. Their ability to handle complex networking, storage, and security tasks at line speed revolutionizes cloud deployments, optimizing resource utilization and enabling seamless scalability while mitigating the performance bottlenecks encountered with conventional NICs. In this paper, we explore offloading NVMe over Fabrics (NVMe-oF) to SmartNIC. We present an experimental design on the FABRIC testbed to perform NVMe-oF by offloading it to a SmartNIC. Our overall goal is to compare CPU performance with and without offloading, benchmark SmartNIC over a regular NIC, and estimate CPU cycle savings.

As the software development landscape transitions from monolithic applications to agile, cloud-native microservices, new challenges in communication have emerged, necessitating the evolution of network service meshes to effectively manage these complexities. While addressing these communication challenges, implementing service mesh solutions introduces higher resource utilization, primarily due to deploying an additional per-microservice infrastructure known as the “sidecar” container. A sidecarless approach, as exemplified by Cilium, leverages extended Berkeley packet filter (eBPF) technology to minimize operational overhead and optimize resource utilization. The advent of data processing units (DPU) represents a crucial evolution in cloud accelerators, designed to offload and accelerate networking, security, and storage tasks traditionally handled by server CPUs. This demonstration integrates DPU technology with Cilium’s sidecarless architecture and proposes to offload Cilium’s control plane to DPUs. Our proposed approach improves system efficiency by deploying Cilium’s control plane components on DPUs in a separate host mode. Further, our approach conserves CPU cycles on the primary hosts and shifts network processing tasks to the DPU. This strategic offload aims to enhance throughput, reduce latency, and unburden CPUs from network processing tasks to improve the performance of service mesh architectures for cloud-native ecosystems.

Increasingly, academic campus networks support large-scale data transfer workflows for data-intensive science. These data transfers rely on high-performance, scalable, and reliable protocols for moving large amounts of data over a high-bandwidth, high-latency network. GridFTP is a widely used protocol for wide area network (WAN) data movement. However, as the GridFTP protocol does not share connection information with the network-layer, network operators have reduced flexibility, particularly in identifying/managing flows across the network. We address this problem by deploying a production “application-aware” software defined network (SDN) for managing GridFTP transfers for data-intensive science workflows. We first propose a novel application-aware architecture called SNAG (SDN-managed Network Architecture for GridFTP transfers). SNAG combinesapplication-layer and network-layer collaboration(termed “application-awareness”) with SDN-enabled network management to classify, monitor and to manage network resources actively. Until now, our SNAG deployment has successfully classified over1.5 BillionGridFTP connections at the Holland Computing Center (HCC), University of Nebraska-Lincoln (UNL). Next, we develop an application-aware SDN system to provide differentiated network services for distributed computing workflows. At HCC, we also demonstrate how our system ensures the quality of service (QoS) for high-throughput workflows such as Compact Muon Solenoid (CMS) and Laser Interferometer Gravitational-Wave Observatory (LIGO). Further, we also demonstrate how application-aware SDN can be exploited to createpolicy-drivenapproaches to achieve accurate resource accounting for each workflow. We present strategies for implementing differentiated network services and discuss their capacity improvement benefits. Lastly, we provide some guidelines and recommendations for developing application-aware SDN architectures for general-purpose applications.

In this paper, we propose ERGO (edge architecture for Ag-IoT), an edge-computing architecture for infrastructureless smart agriculture environments. We also develop Ag-IoT application APIs and the associated microservice infrastructure. Our implementation and evaluations show that ERGO can operate independently of cloud-backed assistance, is highly scalable, modular, and affords composability benefits to Ag-IoT systems. We also demonstrate that ERGO outperforms traditional infrastructure in response latencies and transactional throughput, on average, by over 54% and 77%, respectively.

In this paper, we propose an application-aware intelligent load balancing system for high-throughput, distributed computing, and data-intensive science workflows. We leverage emerging deep learning techniques for time-series modeling to develop an application-aware predictive analytics system for accurately forecasting GridFTP connection loads. Our solution integrates with a major U.S. CMS Tier-2 site; we use a real dataset representing 670 million GridFTP transfer connections measured over 18 months to drive our predictive analytics solution. First, we perform extensive analysis on this dataset and use the connection loads as an example to study the temporal dependencies between various user-roles and workflow memberships. We use the analysis to motivate the design of a gated recurrent unit (GRU) based deep recurrent neural network (RNN) for modeling long-term temporal dependencies and predicting connection loads. We develop a novel application-aware, predictive and intelligent load balancer, APRIL, that effectively integrates application metadata and load forecast information to maximize server utilization. We conduct extensive experiments to evaluate the performance of our deep RNN predictive analytics system and compare it with other approaches such as ARIMA and multi-layer perceptron (MLP) predictors. The results show that our forecasting model, depending on the user-role, performs between 5.88%-92.6% better than the alternatives. We also demonstrate the effectiveness of APRIL by comparing it with the load balancing capabilities of an existing production Linux Virtual Server (LVS) cluster. Our approach improves server utilization, on an average, between 0.5 to 11 times, when compared with its LVS counterpart.