The demand for electricity related to Information and Communications Technologies is constantly growing and significantly contributes to the increase in global greenhouse gas emissions. To reduce this harmful growth, it is necessary to address this problem from different perspectives. Among these is changing the computing scale, such as migrating, if possible, algorithms and processes to the most energy efficient resources. In this context, this paper explores the possibility of running scientific and engineering programs on personal computers and compares the obtained power efficiency on these systems with that of mainframe computers and even supercomputers. Anecdotally, this paper also shows how the power efficiency obtained for the same workloads on personal computers is similar to that obtained on supercomputers included in the Green500 ranking.
For information and communication technology power consumption to be sustainable, the energy efficiency of computing systems must grow at least as fast as the demand for computing services. It is therefore crucial to understand how energy efficiency is evolving and how it will trend in the future, in order to take appropriate measures where possible. This article analyses the evolution of this parameter by analysing high-performance computers from 2008 to 2023, contrasting the results with those from Koomey's Law. It is concluded, after comparing the two that in the studied period and in the near future, energy efficiency continues to grow exponentially but at a slower rate than that established by Koomey's Law (maximum energy efficiency doubles every 2.29 years instead of every 1.57 years). Another interesting result is that energy efficiency grows at a slower rate (doubling every 2.29 years) than performance (doubling every 1.85 years).
Due to energy consumption’s increasing importance in recent years, energy-time efficiency is a highly relevant objective to address in High-Performance Computing (HPC) systems, where cost significantly impacts the tasks executed. Among these tasks, classification problems are considered due to their great computational complexity, which is sometimes aggravated when processing high-dimensional datasets. In addition, implementing efficient applications for high-performance systems is not an easy task since hardware must be considered to maximize performance, especially on heterogeneous platforms with multi-core CPUs. Thus, this article proposes an efficient distributed K-Nearest Neighbors (KNN) for Electroencephalogram (EEG) classification that uses minimum Redundancy Maximum Relevance (mRMR) as a feature selection technique to reduce the dimensionality of the dataset. The approach implements an energy policy that can stop or resume the execution of the program based on the cost per Megawatt. Since the procedure is based on the master-worker scheme, the performance of three different workload distributions is also analyzed to identify which one is more suitable according to the experimental conditions. The proposed approach outperforms the classification results obtained by previous works that use the same dataset. It achieves a speedup of 74.53 when running on a multi-node heterogeneous cluster, consuming only 13.38% of the energy consumed by the sequential version. Moreover, the results show that financial costs can be reduced when energy policy is activated and the importance of developing efficient methods, proving that energy-aware computing is necessary for sustainable computing.
24 October 2023 in Journal of Parallel and Distributed Computing
Software has become more computationally demanding nowadays, turning out high-performance software in many cases, implying higher energy and economic expenditure. Indeed, many studies have arisen within the IT community to mitigate the environmental impact of software. Collecting and measuring software’s power consumption has become an essential task. This paper proposes the design of a standard interface for any currently available smart meter, which is programmatically accessible from any software application and can collect consumption data transparently while a program is executed. This interface is structured into two layers. The former is a driver that provides an OS-level standard interface to the meter, while the latter is a proxy offering higher-level API for a concrete programming language. This design provides many benefits. It makes it possible to substitute the meter for a different device without affecting the proxy layer. It also allows the presence of multiple proxy implementations to offer a programmatic interface to the meter for several languages. A prototype of the proposed interface design has been implemented for a concrete smart meter and OS to demonstrate its feasibility. It has been tested with two experiments. Firstly, its correct functioning has been validated. Later, the prototype has been applied to monitor the execution of a high-performance program, a machine learning application to select the most relevant features of electroencephalogram data.