Joint power management of memory and hard disks

Power reduction becomes increasingly important in computer system design. Lowering energy consumption extends the usage time of battery-powered devices; limiting power consumption of server systems reduces heat-related malfunction and utility bills. When devices are under light workload, power management can reduce power consumption by slowing down the devices or switching the devices to low-power modes. Among all devices, memory and hard disks are significant sources of power consumption. The interaction between memory and hard disks, such as prefetching and caching, provides power-saving opportunities. Prefetching data or caching previously accessed data in memory reduces disk accesses. Using larger amounts of memory can reduce more disk accesses so that the power manager of disks can save more power. However, larger amounts of memory consumes more power to retain data. Choosing optimal memory sizes is important for power optimization of memory and hard disks. Power reduction should consider the power management of memory and hard disks together. In this thesis, I first discuss joint power management of memory and a hard disk when a memory buffer prefetches data from the disk for sequential data access. I develop an analytical model to compute the optimal size of buffer memory and the optimal mode-transition period of managing the disk to minimize the power consumed by buffer memory and the disk. This model is generalized for data buffering of any two interactive devices: one is the data producer while the other acts as the data consumer. The case studies in sensor nodes and a PDA-like computer show up to 20% power savings. Second, I discuss joint power management of memory and a hard disk when the memory caches previously accessed data for data reuse. My method adjusts the size of cache memory and the value of the timeout to switch the disk to low-power modes for reducing power. Lastly, I extend the joint power management to multiple disks. My method allocates the cache memory to the disks and adjust their rotation speeds to reduce the power consumption under performance constraints. Simulation results based on server workload show that the joint methods outperform other power management methods considering the disks or the memory individually.