Please use this identifier to cite or link to this item:
|Title:||Basic Block energy prediction using machine learning methods - Εφαρμογή Τεχνικών Μηχανικής Μάθησης στη Μοντελοποίηση Κατανάλωσης Ενέργειας Στοιχειωδών Δομών Λογισμικού|
basic block energy consumption
|Abstract:||Over the past few decades, there has been a significant effort to find energy-efficient solutions due to climate change. One approach to reducing energy consumption is to minimize the energy consumption of code, particularly for large-scale scenarios such as data centers. Basic block energy consumption prediction is an essential step in creating energy-optimized compilers and schedulers that can significantly decrease code energy consumption. However, this task poses several challenges due to its fine granularity and hardware-specific sensors. To address these challenges, this research investigates the efficiency of machine learning approaches in basic block energy consumption prediction. A basic block en- ergy measurement dataset created in previous research was used for training several machine learning architectures, including both traditional regression methods and deep neural network architectures. Sophisticated code representation techniques were imple- mented to enable this task, given the nature of the basic block as a sequence of Assembly instructions. In particular, sequential neural networks, such as LSTMs, were used to achieve high accuracy in predicting basic block energy consumption, while traditional methods like linear regression and support vector machines proved to perform great as well. The fi- nal best-performing models were able to predict basic block energy consumption with a Mean Absolute Error fluctuating around 0.25 (∗61μJ) for energy measurements with standard deviation around 0.7 (∗61μJ). Despite these promising results, the research has discovered that basic block energy consumption is highly correlated with its pre- ceding basic blocks, therefore the findings were deteriorated due to the lack of data for basic blocks sequences. The final contributions of this research can be summed up as: an extensive empirical study of machine learning for energy consumption prediction, the realization that basic block energy prediction necessitates contextual information for preceding basic blocks, the recognition that whole-program error is an inadequate measure for evaluating a basic block dataset, and lastly, an informed outline of future work.|
|Appears in Collections:||Διπλωματικές Εργασίες - Theses|
Items in Artemis are protected by copyright, with all rights reserved, unless otherwise indicated.