UC San Diego

Reconfigurable match tables (RMT) have been widely adopted in practice over high-speed packet processing pipelines. Coupled with P4, a number of useful application-specific tasks such as in-network telemetry, key-value caching, aggregation and load balancing, have found their way into the network. However, achieving multi-tenancy on such devices has not been a trivial task. RMT switches can run only one program per processing pipeline and multi-tenancy is currently achieved using static program composition with the inability to perform runtime updates. Moreover, memory is local to processing stages making it difficult to achieve efficient resource utilization. I first present ActiveRMT, a capsule-based approach to leveraging computation within the network using a general purpose memory-efficient packet processing model that pre-configures match tables to execute user-defined programs at runtime. Using a fast coordinated approach to dynamic memory allocation along with a constraint-guided approach to synthesizing stateful active programs, I present a unique method of hitlessly provisioning computationally cheap tasks with low memory footprint, that operate on a per-packet basis, onto a programmable switch. However, a capsule-based approach limits the scope of network functionality, particularly in terms of behavioral inspection. Hence, I present vRMT, a system that expands the set of tasks that can be deployed over such a packet processing runtime to include both commonly used network functions and application offloads. I show how network functions that perform behavioral inspection on arbitrary packets -- using programs defined by an authorized third-party (such as a network operator) -- can co-execute with application-specific tasks using automated filter composition and function chain synthesis. Generalizing recirculation-to-completion as a technique to accommodate such function chains, I present a unique method of deploying such combinations of network functions over a best-effort programmable networking substrate. We address a key challenge to supporting complex function chains by showing how to effectively manage switch backplane bandwidth when recirculating packets through RMT pipelines.