Embedded systems have become an integral part of our life in the last few years in multifarious ways, be it in mobile phones, portable audio players, smart watches or even cars. Most embedded systems fall under the category of consumer electronics, such as televisions, mobile devices, and wearable electronics. With several players competing in this market, manufacturers of embedded systems continue to add more functionality to these devices to make them more user-friendly, and often equip them with a very high resolution display and graphics support, and better computing and Internet capabilities. Unfortunately, they are often constrained by tight power/energy budgets, since battery capacity does not improve at the same rate as computing power. While there is clearly much progress to be made in harnessing all the possibilities of embedded systems, limitations in battery capacities, thermal constraints and power/energy budgets surely hinder this progress. Although technology scaling has traditionally addressed both the power minimization and high-performance requirements, with Moore's law nearing its limits, the development of energy-efficient system designs has become critically important. Thus, to be able to continue to provide new and improved features in embedded systems, design-time and run-time power management and minimization holds the key. As a consequence, power optimization has become one of the most defining aspects of designing modern embedded systems. To design such high-performance and energy-efficient embedded systems, it is extremely important to address two basic issues: (1) accurate estimation of power consumption of all system components during early design stages and (2) deriving power optimization solutions that do not negatively impact system performance. In this thesis, we aim to address these two issues for one of the most important components in modern embedded systems: DRAM memories. Towards this, we propose a high-precision DRAM power model (DRAMPower) and a set of performance-neutral DRAM power-down strategies. DRAMPower is a high-level DRAM power model that performs high-precision modeling of the power consumption of different DRAM operations, state transitions and power-saving modes at the cycle-accurate level. To further improve the accuracy of DRAMPower's power/energy estimates, we derive better than worst-case and realistic measures for the JEDEC current metrics instead of vendor provided worst-case measures from device datasheets. Towards this, we modify a SPICE-based circuit-level DRAM architecture and power model and derive better than worst-case current measures under nominal operating conditions applicable to a majority of DRAM devices (>97%) with any given configuration (capacity, data width and frequency). Besides these better than worst-case current measures, we also propose a generic post-manufacturing power and performance characterization methodology for DRAMs that can help identify the realistic current estimates and optimized set of timing measures for a given DRAM device, thereby further improving the accuracy of the power and energy estimates for that particular DRAM device. To optimize DRAM power consumption, we propose a set of performance-neutral DRAM power-down strategies coupled with a power management policy that for any given use-case (access granularity, page policy and memory type) achieves significant power savings without impacting its worst-case performance (bandwidth and latency) guarantees. We verify the pessimism in DRAM currents and four critical DRAM timing parameters as provided in the datasheets, by experimentally evaluating 48 DDR3 devices of the same configuration. We further derive optimal set of timings using the performance characterization algorithm, at which the DRAM can operate successfully under worst-case run-time conditions, without increasing its energy consumption. We observed up to of 33.3% and 25.9% reduction in DRAM read and write latencies and 17.7% and 15.4% improvement in energy efficiency. We validate DRAMPower model against a circuit-level DRAM power model and verify it against real power measurements from hardware for different DRAM operations. We observed between 1-8% difference in power estimates, with an average of 97% accuracy. We also evaluated the power-management policy and power-down strategies and observed significant energy savings (close to theoretical optimal) at very marginal average-case performance penalty without impacting any of the original latency and bandwidth guarantees.
|Qualification||Doctor of Philosophy|
|Award date||3 Oct 2014|
|Place of Publication||Delft|
|Publication status||Published - 2014|