Supporting Run-time Adaptation in Packet Processing Syste ms

Implementors of packet-processing applications on multi-core processors must balance two requirements: (1) adapt processor allocations dynamically to reduce the overall resource provisioning requirement for the system, achieve robustness to traffic fluctuations, and reduce energ y consumption; and (2) utilize, for each application stage, r esources (e.g., memory levels, inter-processor communication mechanisms, etc.) closer or local to the processors on which the stages are mapped to achieve the highest possible throughput. In this paper, we describe the design and implementation of a run-time adaptation system that can meet these two requirements simultaneously. Our design allows each application stage to utilize local resources whenever possible in the steady state. Upon adapting the allocation of processors to stages, the run-time system (1) binds each resource usage within a stage to a new resource instance; and (2) checkpoints and migrates the state from the previous resource instance to the newly-bound resource instance . We describe the design and implementation of our adaptation system in the context of a packet processing system designed using the Intel R © IXP2400 network processor. We show that our design has little impact (14%) on the steadystate throughput of the system. We further show that our design is able to perform resource adaptation for a real application in less than 100ms, allowing processor allocatio ns to be adapted at a very fine time-scale.