Characterizing Variability in Heterogeneous Edge Systems: A Methodology &amp; Case Study

Abdelhafez, Hazem A.; Halawa, Hassan H.; Almoallim, Amr; Ahmadi, Amirhossein; Pattabiraman, Karthik; Ripeanu, Matei

doi:10.1109/sec54971.2022.00016

Cited by 3 publications

(2 citation statements)

References 45 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In [21], authors propose a methodology to characterize variability of performance and power, on a single edge node and across multiple homogeneous edge nodes, for both NVIDIA Jetson AGX and Nano AI edge platforms. The study is conducted for a set of CNN and Rodinia benchmarks.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Quantifying the Variability of Power and Energy Consumption for IoT Edge Nodes

Tofaily,

Raïs,

Anshus

2023

2023 19th International Conference on Distributed Computing in Smart Systems and the Internet of Things (DCOSS-IoT)

View full text Add to dashboard Cite

For IoT and edge systems, measuring, predicting and optimizing energy consumption is an open field. It is important to accurately and precisely characterize power and energy consumption of edge nodes, as energy can be a scarce and key resource. However, there are no fine-grain studies that aim at understanding the potential variability of power and energy consumption of edge nodes. Existing research works give minor or no significance to this potential variability.This paper addresses this problem by quantifying the variability of power and energy consumption on a single edge node, and among multiple homogeneous edge nodes, for three scenarios: Idle, CPU intensive "matrix product" and RAM intensive "flip". These scenarios are found in edge applications. Identical controlled experiments are repeated thoroughly, for each scenario.Results show that power and energy variability exist for all studied scenarios. On a single node, power and energy variability measurements are relatively low. On multiple homogeneous nodes, the variability can be significant. For example, for CPU intensive "matrix product", the variability in energy is equivalent to an idle up-time, in a month, of 7 hours and 52 hours, when considering single and multiple homogeneous nodes, respectively.

show abstract

Section: Related Workmentioning

confidence: 99%

“…Few papers measure power and energy variability for multiple homogeneous edge nodes, but with only one single run per node, or with few nodes. Paper [21] is the closest to our contribution, as it thoroughly studies power variability. However, it focuses on high level benchmarks.…”

Section: Related Workmentioning

confidence: 99%