Approximate Non-volatile Memory in Intermittent Computing
We plan on exploring how to improve the energy efficiency of battery-less Internet of Things(IoT) devices at the cost of introducing errors during the computation, causing a decrease in the Quality of Results (QoR). Battery-less IoT devices are extremely resource-constrained energy-harvesting devices. Due to erratic energy patterns from the ambient, their executions become intermittent; periods of active computation interleave periods of recharging small energy buffers. During the recharging period the device is turned off and the data stored in volatile memory is lost; thus, to cross periods of energy unavailability, the device must persist application and system state onto Non-Volatile Memory (NVM) in anticipation of energy failures. This state persistence operation represents an energy overhead since it consumes a portion of the harvested energy without making any progress in the program execution. Moreover, since NVMs are energy-hungry hardware, this overhead typically consumes a significant portion of the total energy stored in the energy buffer.
We control the write current of Spin-Transfer Torque Magnetic Random-Access Memory (STT-MRAM) used as NVM for an energy harvesting device, resulting in a reduction of the state persistence overhead yet introducing errors in the computation that impact the program's QoR. In our research, we want to illustrate whether or not this is a gamble worth taking, when, and where.
Furthermore, we observe that not all errors have the same impact on the QoR, some data are more resilient to errors, and even by strongly approximating them, the QoR decrease is negligible. We aim to develop an optimization algorithm that assigns each data an optimal approximation level, maximizing the energy overhead reduction while maintaining an acceptable QoR.
To have a comprehensive assessment of the efficacy of our technique, we will evaluate the efficacy of our technique on a broad and heterogeneous set of benchmarks and various low-power microcontrollers commonly used in intermittent systems.
During the first phase we plan on experimenting on a simulated environment, but to obtain more precise results we can also consider moving on to the actual hardware.
In the long run, our objective is to expand the concept of approximate states within intermittent computing to also consider scenarios where the intermittent system utilizes a Non-Volatile Memory (NVM) other than STT MRAM.
Back to Current Students