Using programmable network devices to aid in-network machine learning has been the focus of significant research. However, most of the research was of a limited scope, providing a proof of concept or describing a closed-source algorithm. To date, no general solution has been provided for mapping machine learning algorithms to programmable network devices. In this paper, we present Planter, an opensource, modular framework for mapping trained machine learning models to programmable devices. Planter supports a wide range of machine learning models, multiple targets and can be easily extended. The evaluation of Planter compares different mapping approaches, and demonstrates the feasibility, performance, and resource efficiency for applications such as anomaly detection, financial transactions, and quality of experience. The results show that Planterbased in-network machine learning algorithms can run at line rate, have a negligible effect on latency, coexist with standard switching functionality, and have no or minor accuracy trade-offs.

No abstract

Very-large-scale network-on-chip (VLS-NoC) has become a promising fabric for supercomputers, but this fabric may encounter the many-fault problem. This article proposes a deterministic routing algorithm to tolerate the effects of many faults in VLS-NoCs. This approach generates routing tables offline using a breadth-first traversal algorithm and stores a routing table locally in each switch for online packet transmission. The approach applies the Tarjan algorithm to degrade the faulty NoC and maximizes the number of available nodes in the reconfigured NoC. In 2D NoCs, the approach updates routing tables of some nodes using the deprecated channel/node rules and avoids deadlocks in the NoC. In 3D NoCs, the approach uses a forbidden-turn selection algorithm and detour rules to prevent faceted rings and ensures the NoC is deadlock-free. Experimental results demonstrate that the proposed approach provides fault-free communications of 2D and 3D NoCs after injecting 40 faulty links. Meanwhile, it maximizes the number of available nodes in the reconfigured NoC. The approach also outperforms existing algorithms in terms of average latency, throughput, and energy consumption.

No abstract