Distributed deep learning (DDL) requires a great investment in cloud infrastructure, including accelerated compute nodes and networking hardware capable of supporting high-performance networking, e.g., Remote Direct Memory Access (RDMA). When a host running a DDL application becomes unreachable, the cost can be high as application-level failure recovery is slow and disruptive. When the host is unreachable due to host failure, this is unavoidable; however, when the network components involved in attaching the host to the core data center network fail, we argue that this cost is avoidable. This paper introduces THORN-ML, a hardware-offloaded resilient network architecture that is completely transparent to DDL applications and works with commodity hardware. We evaluate THORN-ML on a cluster of 5 nodes with Nvidia A100 GPUs and Mellanox ConnectX-5 NICs, with several applications leveraging model parallelism and/or data parallelism, and find that THORN-ML reduces disruption from minutes (impacting the whole cluster) to milliseconds (impacting packets that can be re-transmitted).

Recent advances in network switch designs have enabled machine learning inference directly within the switch at line speed. However, hardware constraints limit switches capabilities of tracking stateful features essential for accurate inference, as the demand for these features grows rapidly with line rates. To address this, we propose DIDA, a distributed in-network machine learning approach. In DIDA, feature extraction occurs at the host, features are transmitted via in-band telemetry, and inference is performed on the switches. In this paper, we evaluate the effectiveness and efficiency of this architecture. We examine its impact on network bandwidth, CPU and memory usage at the host, and its robustness across different feature sets and deep neural network classifications.

With the rise of the Network Softwarization era, eBPF has become a hot technology for efficient packet processing on commodity hardware. However the development of custom eBPF solutions is a challenging process that requires highly qualified human resources. Indeed, in eBPF, it is difficult to devise truly modular applications since the development model does not favour the use of pre-compiled functions and libraries. In addition, for safety purposes, each eBFF program must pass a binary code verifier of the Linux kernel, which may increase the difficulty of the development process. To overcome such difficulties and enable a new development model, in this paper we propose the eCLAT framework with the goal to lower the learning curve of engineers by re-using eBPF code in a programmable way. eCLAT offers a high level programming abstraction to eBPF based network programmability, allowing a developer to create custom application logic with no need of understanding the complex details of regular eBPF programming. A developer can write eCLAT scripts in a python-like language to compose eBPF programs. To support such abstraction at the eBPF level, we created an eBPF framework called HIKe which brings code reuse and modularity in eBPF. The eCLAT/HIKe solution does not require any kernel modification. The new development model is tested through two concrete examples and compared with other proposed frameworks in the eBPF world.

Software-Defined Wide Area Networks (SD-WANs) are used to provide services to enterprises with geographically dispersed locations in a flexible and efficient way. We focus on SD-WAN services based on the Segment Routing over IPv6 (SRv6) technology. Performance Monitoring solutions are needed in SD-WANs to detect performance degradation and outages, and optimize network operations. In this paper, we describe a high performance solution for end-to-end delay monitoring for SRv6 based SD-WAN services. The proposed solution leverages the Simple Two-way Active Measurement Protocol (STAMP) to monitor the delay of an SRv6 path between two nodes called STAMP Session-Sender and Session-Reflector. We describe three implementations of the STAMP Session-Sender and Session-Reflector for a Linux software router and compare their performance. In particular, two implementations are based on user space processing and one is based on eBPF. The results show that the eBPF-based implementation outperforms the user space implementations and has a negligible impact on the forwarding capacity of the Linux software router.

Segment Routing over IPv6 (SRv6 in short) is a networking architecture suitable for both IP backbones and datacenters. The research, standardization and implementation of this architecture are actively progressing and SRv6 is already adopted in a number of large scale deployments. Effective solutions for Performance Monitoring (PM) of SRv6 networks are strongly needed and there is a lot of activity in this area. A full blown Performance Monitoring solution needs to include: i) Data Plane (as needed to measure metrics such as packet loss and delay); ii) Control Plane (to send commands to the monitoring entities in the nodes); iii) Management Plane (e.g., to collect the measured metrics). Moreover, Big-Data tools and solutions can be applied inside or above the traditional Management Plane boundaries to store and analyze the collected data. In this article we describe SRv6-PM, a solution for Performance Monitoring of SRv6 networks that deals with all the aspects discussed above. SRv6-PM features a cloud-native architecture that supports: i) the ingestion, processing, storage and visualization of PM data using Big-Data tools; ii) the SDN-based control of network routers to drive the performance monitoring operations. In particular, we focus on Loss Monitoring and consider a solution capable of tracking single packet loss events operating in near-real time (e.g., with a time granularity in the order of 10-20 seconds). SRv6-PM is released as open source. We offer a re-usable and extensible platform that can be automatically deployed in different environments, from a single host to multiple servers on private/public clouds and includes a set of Big-Data tools and the SDN control plane. We also provide a reproducible Data Plane environment for PM experiments in SRv6 networks based on the Mininet emulator.