CAPÍTULO II MARCO TEORICO
2.1. Eventos culturales
2.1.2. Imagen de la ciudad:
This chapter documented a selection of the current research in the area of power management, power-awareness in computational environments and source code cost analyses. Techniques for power management and cost analy- ses usually fall under one of three categories (Traditional and General Pur- poses, Micro and Hardware Level, Macro and Software Level) and this chap- ter described a number of tools that are associated with each of these cate- gories. Of particular interest is the movement from low-level hardware power management such as reducing cache misses [20] to high-level source code transformation [56][18][67].
ness is no longer restricted to areas of mobile computing or energy-limited computational environments but is gradually moving towards the areas of resource-critical computational environments such as parallel and distributed computational environments, and the Grid [30][43] where energy consump- tion has become a major factor to running cost [68]. Therefore to ensure a low computational running cost, it is essential to develop new approaches to predict an application’s energy consumption at a source code level and to include this as a metric when building performance models.
Power Analysis and Prediction
Techniques
3.1
Introduction
Whilst current research has produced a bank of techniques on power analysis which are either software or hardware focused, they share some common shortcomings:
Current techniques’ insufficiencies - The review of current power analy- sis methodologies in chapter 2 suggests some important areas, which are concerned with the development of designing a well-formed power analysis strategy, that still need to be addressed. In particular, the majority of analy- sis techniques that are currently available either require the analysers to have
specific knowledge such as the low-level machine code or require the use of specialised equipment. This technical knowledge and specialised equipments might not be available during power analysis and such dependencies will only hinder the flexibility of the analysis methodology. Furthermore, current methodologies such asinstruction level analysis[71] over emphasise the mea- surement of absolute energy consumption. In the case of [71] analysers must acquire the absolute measurement of the current drawn for every machine instruction and this can undermine the usefulness of the analysis technique itself.
In modern performance and cost analysis, there are three types of evalu- ation techniques: analytical modelling, simulation and measurement. These techniques offer different levels of accuracy, in particular, analytical modelling requires so many simplifications and assumptions that high level accuracy is not essential [39]. Unfortunately since current power analysis techniques sep- arate themselves from the general performance evaluation domain, they lack the ability to abstract the technicality of both target machine architectures and target applications. To develop a well-formed power analysis strategy means that such strategy should possess the flexibility similar to the ones in the performance domain, so that level of accuracy can be varied and mea- surement values can be relative.
Performance incompatibility - The current power analysis methodologies are simply not compatible with the current advance in performance evalua- tion and optimisation. Techniques which have been reviewed either neglect
performance efficiency or isolate energy consumption from other performance metrics such as execution time or memory utilisation. It is believed that electrical energy is a major cost when running some large scale applications and the cost of dissipating tens or potentially hundreds of megawatts is pro- hibitive. This means during an overall cost analysis on performance measure, energy consumption should be taken into account and should eventually be integrated into performance characterisation and evaluation.
No standard characterisation model- To compensate for the shortcom- ings of current power analysis strategies, a standard model is needed for power analysis and it should allow applications to be systematically or hier- archically optimised for energy consumption. As applications in recent years are moving toward execution environments which are heterogeneous, distrib- uted and even ubiquitous [16], without a standard model that can categorise and characterise the energy usage of application’s workload generically, cur- rent power analysis techniques will prove to be too inefficient and impractical. Also by using an analytical model, it allows measurements to be based on a hierarchical framework of relativity.
Following on from the weaknesses mentioned above, the proposed method- ologies are aimed at developers without expertise in technical areas such as low-level machine code and without specialised equipment to carry out energy measurements. During the preliminary stages of this research we propose an application-level power analysis and prediction technique which adopts the performance evaluation framework and techniques developed by the High
Performance Systems Group [35] at the University of Warwick. Furthermore this chapter introduces a theoretical concept to construct a power classifica- tion model based on benchmark workloads. This model allows a more rel- ative energy consumption prediction of an application, although this model has not yet been fully implemented, some insights in choosing the relevant characterisation units have been established.