Packet processing systems (e.g., routers, virus scanners,
intrusion detectors, SSL accelerators, etc.) provision
sufficient number of processors to handle the expected
maximum workload. The observed load at any instant,
however, is often substantially lower; further, the
load fluctuates significantly over time. These properties
offer an opportunity to conserve energy (e.g., by deactivating
idle processors or running them in low-power
mode). In this paper, we present an on-line algorithm
for adapting the number of activated processors such that
(1) the total energy consumption of the system across a
packet trace is minimized and (2) the additional delay
suffered by packets as a result of adaptation is deterministically
bounded. The resulting Power Management Algorithm
(PMA) is simple, but it accounts for system reconfiguration
overheads, copes with the unpredictability
of packet arrival patterns, and consumes nearly the same
energy as an optimal off-line strategy. A conservative version
of the algorithm (CPMA), turns off processors less
aggressively than is optimal but still provides good energy
savings while reducing the additional packet latency
introduced by power management. We demonstrate that
for a set of trace workloads both algorithms can reduce
the core power consumption by 60-70% while increasing
the average packet processing delay by less than 560µs
(PMA) and less than 170µs (CPMA).