Multitenant In-Network Acceleration with SwitchVM

Explore a groundbreaking conference talk on multitenant in-network acceleration using SwitchVM. Delve into a practical approach for implementing multitenancy on programmable network switches, making in-network acceleration accessible to cloud users. Learn about the Switch Virtual Machine (SwitchVM) deployed on switches, offering an expressive instruction set and program state abstractions. Discover how tenant programs, called Data-Plane filters (DPFs), execute on SwitchVM in a sandbox with memory, network, and state isolation policies controlled by network operators. Understand the Turing-complete nature of DPFs, their ability to maintain state in packets and switch virtual memory, form dynamic chains, and steer packets to desired destinations while enforcing operator policies. Examine the practical implementation of SwitchVM in P4 on Intel Tofino switches, and explore various use cases it supports. Gain insights into three complex applications implemented using SwitchVM: key-value store cache, load-aware load balancer, and Paxos accelerator. Learn about SwitchVM's strong performance isolation, zero-overhead runtime programmability, and its capacity to hold significantly more in-switch programs than existing techniques while supporting thousands of concurrent tenants with private states.

NSDI '24 - Multitenant In-Network Acceleration with SwitchVM