Conserving energy in mobile and embedded systems is becoming an active area of research as hardware components are becoming more power-hungry than ever, and as battery technology is not able keep up with the growing demands. By exploiting the ability of modern hardware components to operate at multiple power levels, recent research has demonstrated that a significant amount of energy can be conserved. Due to the high-peak power demands of the processor, a large body of work has focused on reducing processor energy consumption. Weiser et al. [38] first demonstrated the effectiveness of using Dynamic Voltage Scaling (DVS) to reduce power dissipation in processors. Later work [29,2,11,14,,25,30,32,31] further explored the effectiveness of DVS techniques in both real-time and general-purpose systems.
There is also a large body of work that focused on reducing power in other system components, including wireless communication [36,18,21,10], disk drives [24,6,7,22], flash [5,28], cache [1,19,37], and main memory [23,8,9,3,4], while others [12,40,26,35] explored system-level approaches to extend/target the battery lifetime of systems, as opposed to saving energy for individual components.
Among the works dealing with main memory energy, in [23,8], Lebeck et al. studied the effects of various static and dynamic memory-controller policies to reduce power dissipated by the memory using extensive simulations. However, they assumed having additional hardware support to do very fine-grained idle time detection for each device so the controller can correlate this idle time with a power state for each device. In a later work, they used a stochastic Petri Nets approach to explore more complex policies [9]. Our work differs significantly in not assuming any additional hardware support or a particular memory architecture. Moreover, by elevating the decision-making to the OS level, we can use information known to the OS to conserve more energy without degrading performance. Finally, we have fully implemented a power-aware VM system that handles the complexities of a real, working system, and demonstrated its effectiveness when running real-world applications.
Delaluz et al. [3] took a compiler-directed approach, where power-state transition instructions are automatically inserted into compiled code based on offline profiling. The major drawback of this approach is that the compiler only works with one program at a time and has no information about other processes that may be present at runtime. Therefore, it needs to be either less aggressive or else it can trigger large performance and energy overheads when used in a multitasking system. This approach, however, is appropriate for DSP-like platforms where single-application systems are common.
Delaluz et al. [4] later showed a simple scheduler-based power-management policy. The basic idea is similar to our work, but is of much more limited scope. In our work, much effort is put into making the underlying physical page allocator to allocate pages by collaborating with the VM through a NUMA management layer so the energy footprint is reduced for each process, whereas they rely on the default page allocation and VM behaviors. As we have seen in Section 4.2, a substantial amount of power-saving opportunities remain unexploited even with our rudimentary implementation of PAVM, let alone when randomly allocating pages using the default page allocator. In [23], it was also noted that the default page allocation behavior has a detrimental impact on the energy footprints of processes. Second, we have explored advanced techniques such as library aggregation and page migration which are necessary for reducing memory footprints when complex sharing between processes in real operating systems is involved. Finally, in their work, the active nodes are determined using page faults and repeated scans of process page tables. Although this ensures only the truly active nodes are detected, it is intrusive and involves high operational overheads. In contrast, we take every precaution to avoid performance overheads and hide any unavoidable latencies in our implementation, and the end result is a PAVM system that can save a significant amount of energy with only a very small performance overhead.