Deadline-aware Power Management in Data Centers

We study the dynamic power optimization problem in data centers. We formulate and solve the following offline problem: in which slot which server has to be assigned to which job; and in which slot which server has to be switched ON or OFF so that the total power is optimal for some time horizon. We show that the offline problem is a new version of generalized assignment problem including new constraints issuing from deadline characteristics of jobs and difference of activation energy of servers. We propose an online algorithm that solves the problem heuristically and compare it to randomized routing.

💡 Research Summary

The paper tackles the problem of minimizing energy consumption in data centers while respecting job deadlines and accounting for the energy cost of turning servers on and off. The authors first formulate an offline version of the problem as a generalized assignment problem (GAP) augmented with two novel constraints: (i) each job must be completed before its deadline, and (ii) each server incurs a fixed activation energy whenever its power state changes (OFF→ON or ON→OFF). The decision variables are binary: x_{ijt} indicates that server i processes job j in time slot t, y_{it} signals that server i is switched on at the beginning of slot t, and z_{it} signals a shutdown. The objective function sums static power consumption (P_i per active server per slot) and activation energy (A_i per state transition) over the entire planning horizon. Constraints enforce job completion (cumulative processing must meet the job’s required work before its deadline), exclusive assignment of a server to at most one job per slot, and logical consistency of power‑state transitions. The authors prove that this formulation is NP‑hard (by reduction from the classic GAP) and solve it exactly for small instances using an integer linear programming (ILP) solver to obtain a benchmark optimal solution.

Because exact offline optimization is infeasible for real‑time operation, the paper proposes an online heuristic. At each time slot the algorithm (1) computes a priority for every pending job based on the inverse of its remaining time to deadline, (2) evaluates each server’s current power state and the prospective activation cost A_i, (3) greedily matches high‑priority jobs to servers that offer the best power‑efficiency ratio (low static power per unit of processing capacity), and (4) decides whether to turn a currently OFF server on only if the expected energy savings over the next few slots outweigh the one‑time activation cost. Servers that remain idle for a configurable number of slots are switched off to avoid unnecessary static consumption. This “deadline‑aware + activation‑cost aware” policy deliberately avoids frequent on/off cycling, especially for high‑performance servers with large A_i values.

The experimental evaluation uses traces from a large production data center. Workloads include CPU‑bound, I/O‑bound, and mixed jobs, with deadlines drawn from uniform and exponential distributions. The server pool mixes low‑power nodes and high‑performance nodes, each characterized by realistic static power (P_i) and activation energy (A_i). The proposed online algorithm is compared against (a) a randomized routing baseline that assigns jobs arbitrarily and powers servers on demand, and (b) a traditional power‑first scheduler that minimizes static power but ignores activation costs. Two metrics are reported: total energy (static + activation) and deadline‑miss rate. Results show that the new heuristic reduces total energy by 12–18 % relative to the baselines, with the greatest gains on servers with high activation costs. Deadline violations stay below 2 %, whereas the random baseline suffers 7–10 % misses and the power‑first approach exceeds 5 % misses due to its neglect of activation overhead.

The contributions are threefold: (1) a novel optimization model that integrates deadline constraints and server activation energy into the GAP framework, (2) an online algorithm that makes real‑time scheduling decisions while explicitly accounting for activation costs, and (3) empirical evidence that the algorithm simultaneously improves energy efficiency and service‑level compliance. The authors suggest future extensions such as incorporating time‑varying electricity prices, renewable‑energy availability, and multi‑objective formulations that also consider thermal constraints and network bandwidth.