Efficient collective communication is crucial to parallel-computing applications such as distributed training of large-scale recommendation systems and natural language processing models. Existing collective communication libraries focus on optimizing operations for dense inputs, resulting in transmissions of many zeros when inputs are sparse. This counters current trends that see increasing data sparsity in large models.

Traffic monitoring is essential to manage large networks and validate Service Level Agreements. Passive monitoring is particularly valuable to promptly identify transient fault episodes and react in a timely manner. This article proposes a novel, non-invasive and flexible method to passively monitor large backbone networks. By using only packet counters, commonly available on existing hardware, we can accurately measure packet losses, in different segments of the network, affecting only specific flows. We can monitor not only end-to-end flows, but any generic flow with packets following several different paths in the network (multipoint flows). We also sketch a possible extension of the method to measure average one-way delay for multipoint flows, provided that the measurement points are synchronized. Through various experiments we show that the method is effective and enables easy zooming in on the cause of packet losses. Moreover, the method can scale to very large networks with a very low overhead on the data plane and the management plane.

Many data center applications nowadays rely on distributed computation models like MapReduce and Bulk Synchronous Parallel (BSP) for data-intensive computation at scale [4]. These models scale by leveraging the partition/aggregate pattern where data and computations are distributed across many worker servers, each performing part of the computation. A communication phase is needed each time workers need to synchronize the computation and, at last, to produce the final output. In these applications, the network communication costs can be one of the dominant scalability bottlenecks especially in case of multi-stage or iterative computations [1].

Programmable data plane hardware creates new opportunities for infusing intelligence into the network. This raises a fundamental question: what kinds of computation should be delegated to the network?

Virtual Network Functions (VNFs) are often realized using virtual machines (VMs) because they provide an isolated environment compatible with classical cloud computing technologies. However, VMs are demanding in terms of required resources (CPU and memory) and therefore not suitable for low-cost devices like residential gateways. Such equipment often runs a Linux-based operating system that includes by default a (large) number of common network functions, which can provide some of the services otherwise offered by simple VNFs, but with reduced overhead. In this paper those native software components are made available through a Network Function Virtualization (NFV) platform, thus making their use transparent from the VNF developer point of view.