Abstract: Data plane programmability is redesigning how we manage and operate network devices. Recent efforts from industry and academia have shown an increased shift from control plane decisions to data plane-based ones (e.g., routing decisions). Most of these algorithmic decisions executed by data planes are still deterministic and dependent on the control plane. However, we believe that we can break this control-loop dependency and allow data planes to learn by themselves network states and make appropriate choices automatically. In this project, we propose SkyNet, the first effort to implement, operate, and orchestrate Artificial Neural Networks (ANN) in programmable data planes. ANNs will allow, in the near future, the design and operation of domain-specific applications for data planes, allowing them to be trained for a multitude of purposes. Despite the existence of a few initiatives to implement data plane intelligent mechanisms, the design and operation of ANNs are still challenging for two reasons. First, programmable data plane architectures are still quite restricted regarding arithmetic operations (e.g., floating-point operations). Second, data plane virtualization is still in the early stages, making the execution of multiple applications in parallel difficult to realize. To fill in these gaps, SkyNet aims to ease the burden of describing, implementing and operating ANNs for network developers and network infrastructure operators. The project unfolds into four objectives: (i) designing a framework for ANN developing, allowing high-level description and code generation for programmable network infrastructures; (ii) designing a lightweight virtualization environment for programmable data planes; (iii) optimizing the management and operation of ANNs in programmable infrastructures; and (iv) providing a programmable testbed, among the participating institutions, for ANN experimentation. The results to be obtained in this project have potential to cause a lasting impact within its research area in the coming years.
Ano Inicio: 2021
Ano Fim: 2026
Coordenador Local: Luciano Paschoal Gaspary
Agência de Fomento: FAPESP