Debunking the myth that tight packing is energy conserving.

Energy takes about half the operational expenses of data centers making energy conservation a critical goal. Fine-grained control over frequency of processors aids in reducing power at the cost of performance degradation. Placing tasks on servers is formulated as a bin packing and tight packing is considered to be energy conserving as idle servers can be shut down. However, we reveal that due to neglecting task deadline and operating at higher frequency leads to higher energy consumption compared to distributing of tasks over larger number of machines at lower frequencies. In this paper, we explore how to provision tasks in an energy-optimal manner. We show that task-based provisioning is a variable-sized bin packing problem and analyze energy efficiency of 14 classical and proposed heuristics for different distributions of task deadline requirements. We establish that tight packing leads to higher energy usage and we need to consider the task deadline while provisioning. Our heuristics saves as much as 65% energy compared to existing greedy heuristics for an instance of provisioning.