SiPanda: Rethinking the Network Datapath
Current network I/O vendors force developers and network infrastructure architects to choose between high performance and flexibility. They just re-spin old industry paradigms, providing incremental improvement but never solving the real problem for atscale distributed network architectures – the need for extreme performance and network datapath flexibility. This is why hyperscalers have built their own network infrastructure, but even they are looking for a better solution. The goal of SiPanda is to provide cloud service providers (CSPs) and hyperscalers with the capabilities that complex, distributed at-scale implementations require to be competitive. Today, SiPanda is introducing the “Programming and Network Datapath Acceleration” model (PANDA) network framework and parser – the first piece of our combined network architecture. PANDA is an open-source project that breaks the traditional network datapath paradigm and creates an intuitive, open programming capability complex, at-scale applications. This new model network architects to program the capabilities that they want to add to their network in C, parse it using PANDA, and then run it on their existing network hardware. PANDA parses the I/O into multiple parallel streams at scale, efficiently enabling the network flexibility needed by cloud service providers and hyperscalers. You can access the github for PANDA at https://github.com/panda-net/panda. The hardware compliment to PANDA is Panda-X, SiPanda’s domain-specific hardware that extends the OpenISA RISC-V architecture to accelerate the network datapath created by PANDA in a similar way to how GPUs accelerate graphics. Unlike current solutions that run out of speed at 100Gbps, Panda-X is designed to excel at speeds in excess of 1 Tbps, while still delivering the flexibility that at-scale networks need. Rethink the Network with SiPanda.
Tom Herbert, SiPanda CTO (Feb 18, 2021)