Data center networks undergo the coexistence of latency-sensitive mice flows and bandwidth-intensive elephant flows. Jointly optimizing the performance of both traffic classes poses complex challenges. Existing flow schedulers either rely on detailed flow size information or require numerous physical priority queues (PQs) within network switches, thus facing practical challenges.In this work, we propose a novel flow scheduling algorithm, namely Multi-Path Multi-Level Feedback Queueing (MPMLFQ), to overcome these limitations. MP-MLFQ leverages the spatial diversity and regularity of DCNs to realize a scheduler with numerous logical priority levels while occupying as few as 2 physical PQs at each switch port. We designed MP-MLFQ to run atop modern programmable networks, and highlighted how to implement it without modifications at the end-hosts’ stacks. Our simulation results show that MP-MLFQ outperforms existing flow size-agnostic solutions in minimizing the flow completion time, when only two PQs are available.

Excessive network packet loss is one of the strongest symptoms of the presence of some hardware or software-related infrastructure anomaly. As such, multiple techniques to rapidly and efficiently measure packet loss in physical network equipment have been developed throughout the years. Yet, most of the proposed techniques fall short in the case of cloud scenarios which combine the presence of physical and virtual network devices. In this paper, we tackle the problem of providing lightweight and fine-grained packet loss monitoring at virtual switches. We achieve our goal by combining packet coloring, with efficient packet loss signal extraction and aggregation entirely within the virtual switch level. To understand its feasibility in a production environment, the proposed system has been implemented and evaluated in a synthetic scenario and for real-world use cases. Our implementation on top of OvS-DPDK shows that the proposed system achieves 95% accurate packet loss measurement while introducing negligible switching throughput degradation.

Several applications for connected cars leverage the mobility information periodically broadcasted by cars through standard vehicle-to-vehicle messages. We propose an architecture in which each car generates and sends reports including the messages received from its neighbors to the access network infrastructure. The infrastructure collects and stores the received reports through multiple blockchains, each of them referring to a different geographical area. A smart contract is then executed to verify the spatial coherence among the received data. We implement a proof-of-concept of such a solution using Hyperledger Fabric, and we investigate the scalability of our solution in terms of resource consumption in a MEC system.

In SDN stateful data planes, switches can execute algorithms to process traffic based on local states. This approach permits to offload decisions from the controller to the switches, thus reducing the latency when reacting to network events. We consider distributed network applications that process traffic at each switch based on local replicas of network-wide states. Replicating a state across multiple switches poses many challenges, because the number of state replicas and their placement affects both the data traffic distribution and the amount of synchronization traffic among the replicas. In this paper, we formulate the optimal placement problem for replicated states, taking into account the data traffic routing, to ensure that traffic flows are properly managed by network applications, and the synchronization traffic between replicas, to ensure state coherence. Due to the high complexity required to find the optimal solution, we also propose an approximated algorithm to scale to large network instances. We numerically show that this algorithm, despite its simplicity, well approximates the optimal solution. We also show the beneficial effects of state replication with respect to the single-replica scenario, so far considered in the literature. Finally, we provide an asymptotic analysis to find the optimal number of replicas.

Recently proposed approaches to minimize the Flow Completion Time (FCT) in data centers do not require any a-priory information about the flow size, thus appear to be both practical and efficient. These solutions are based on a system of multiple priority queues (PQs) at both the servers and the switches and they may require to solve a complex algorithm to optimally split the traffic across the different PQs. However, the actual availability of priority queues at the switches is typically limited, thus restricting the applicability of these approaches. In our work we propose a novel approach, NOS2, which requires only 2 PQs at the switches while maintaining multiple PQs at the servers, and leverages a central controller that optimally coordinates the traffic split among the different priority levels. We show by simulation that NOS2 is able to achieve performance close to state-of-art solutions with significantly smaller implementation complexity. Thus, NOS2 is expected to provide a better trade- off between performance and implementation complexity.

Recently proposed architectures for high- performance data centers advocate the adoption of a centralized control that coordinates the packet transfers within the data center network. In such architectures, centralized algorithms perform decisions regarding packet scheduling (i.e., when a packet is transferred from the server to the switches) and packet routing (i.e., the sequence of traversed switches to reach the destination server) with the aim of optimizing the overall performance. Notably, centralized control permits to reduce packet contention and to minimize the delay introduced by the data center network, but may rely on expensive mechanisms such as synchronous transmission of the packets from the servers. In our work we compare a generic synchronous architecture for the centralized control of a data center with a generic asynchronous architecture, that relaxes the strict packet-by-packet control required by the synchronous architecture and enables a simpler rate-based implementation at the servers. We show that the two architectures achieve near identical performance in terms of throughput, fairness and delays. We finally conclude that asynchronous architectures offer a better trade-off in terms of complexity and performance, with better scaling properties to very large sizes.