Accuracy-Aware Power Management for Many-Core Systems Running Error-Resilient Applications

Abstract

Power capping techniques based on dynamic voltage and frequency scaling (DVFS) and power gating (PG) are oriented towards power actuation, compromising on performance and energy. Inherent error resilience of emerging application domains such as Internet-of-Things (IoT) and machine learning provide opportunities for energy and performance gains. Leveraging accuracy-performance trade-offs in such applications, we propose approximation (APPX) as another knob for close-looped power management, to complement power knobs with performance and energy gains. We design a power management framework, APPEND+, that can switch between accurate and approximate modes of execution subject to system throughput requirements. APPEND+ considers the sensitivity of the application to error to make disciplined alteration between levels of approximation such that performance is maximized while error is minimized. We implement a power management scheme that uses APPX, DVFS and PG knobs hierarchically. We evaluated our proposed approach over machine learning and signal processing applications along with two case studies on IoT -early warning score system and fall detection. APPEND+ yields 1.9× higher throughput, improved latency up to 5×, better performance per energy and dark silicon mitigation compared to state-of-the-art power management techniques over a set of applications ranging from high to no error resilience.

Hierarchical adaptive Multi-objective resource management for many-core systems

“…On the other hand, the assignment of more than one task per PE minimizes the hop number and the number of active PEs and leads to energy savings [23] . Some proposals [6,16,20,21] deploy distinct algorithms to map one task per processor in square shapes (third column). Another mapping approach to meet power constraints is the pattern mapping [24] .…”

“…The overhead of a task migration is low for homogeneous many-cores without shared memory [25] . As some works [6,16,20] do not support multi-tasking mapping, task remapping brings no significant advantages concerning power and resources, and then it is not employed (fourth column). Pathania et al [21] benefit from an exponential property of contiguous square shapes to optimally remap running tasks and avoid resources fragmentation due to dynamic workload.…”

“…In the work by Rahmani et al [16] , the application enters the system if there are available processors, but can also be killed suddenly if the power overcomes the capping. An alternative approach for AA assumes that application tasks can have approximated versions to trade-off power and performance [20] . MORM can remap running tasks to share PEs and map incoming applications in a reduced number of PEs to open power and resources room.…”

Implementación de Algoritmos de Procesamiento Digital de Señales en Hardware Paralelo: Artículo de revisión

“…La implementación de algoritmos de PDS en esta arquitectura es reportada en la literatura logrando tiempos de ejecución diez veces más rápidas comparadas con la implementación del algoritmo tradicional (Ramalakshmi and Kompala, 2017). El factor que ha tomado relevancia actualmente es la potencia disipada de los procesadores, que al tener más núcleos se incrementa; existen diversas técnicas para bajar la potencia disipada y mejorar la eficiencia energética; dos técnicas son el escalamiento dinámico de voltaje y frecuencia (DVFS por sus siglas en inglés) y potencia canalizada (PG por sus siglas en inglés) (Chen et al, 2017) (Kanduri, … and. Con el fin de descargar de trabajo al procesador y aumentar la potencia de ejecución de algoritmos se introdujo el uso de las unidades de procesamiento gráfico, con altas velocidades de procesamiento y gran cantidad de núcleos (480 en GeForce GTX 295).…”

Power profiling and monitoring in embedded systems: A comparative study and a novel methodology based on NARX neural networks

“…This decade-old question is still generating studies aiming to enhance accuracy while accounting for newly released technologies [3]. Furthermore, with the emergence of new factors, such as the Internet of Things (IoT), and the ubiquity of embedded systems in general, the interest in power consumption is no longer limited to just mobile devices, and now includes most embedded electronics [4,5].…”