[ecoop-info] PhD thesis position "Compilation optimizations for peak power control"

Olivier Zendra Olivier.Zendra at inria.fr
Wed Sep 30 09:47:17 CEST 2009

PhD Thesis position "Compilation optimizations for peak power control"
(Open-PEOPLE N°3 by INRIA/Nancy Grand Est and UBS/Lab-STICC)

Starting date:
Fall 2009.

Context and motivation:
This work is part of the Open-PEOPLE (Open Power and Energy Optimization
PLatform and Estimator) project, financed since the end of 2008 by ANR (Agence
Nationale de la Recherche), the French National Research Agency.
Open-PEOPLE initially gathers 5 partners from academia and 2 from industry.
This project aims at providing a federative and open platform for the
estimation and optimization of power and energy consumption in computer
systems. The platform users will be able to evaluate application consumption
on a hardware architecture chosen among a set of provided typical, parametric
architectures. In the considered system, the components will be picked from a
library of hardware and software components, be they parametric or not. It
will be possible to perform the estimation at various stages of the
specification refinement, thanks to a methodology based on multi-level,
interoperable and exchangeable consumption models allowing an easy exploration
of the design space. Thus, estimations results may be used to check the energy
behaviour of a system developed with simulation platforms. Feedback about the
application functional properties will allow further refining of the
estimation results in Open-PEOPLE. A standardisation of consumption models
will be proposed in order to allow interoperability and have easier exchanges
with other platforms. The Open-PEOPLE library of consumption models will be
extendible: new component models will be added as the user applicative
requirements evolve and as implementation techniques progress. To do so, the
software estimation platform that will be accessible via an Internet portal
shall be linked to a hardware platform made of an automated measurement
testbench, which will be controllable from the software platform. A standalone
version will also be provided to meet the confidentiality requirements of
industry. A library of applications benchmarks will be proposed to
characterize new components and new architectures.
In addition to the research work required to build methods for multi-level
estimation in heterogeneous complex systems, research work shall be carried on
in order to offer new methods and techniques making it possible to optimize
consumption thanks to the results provided by Open-PEOPLE
Open-PEOPLE is hence geared towards academia to support research work
pertaining to consumption estimation and optimization methods, as well as
towards industry to estimate or optimize the consumption of future products.

PhD topic:
This Thesis work, as part of the Open-PEOPLE research proposal, consists in
tackling the issue of peak power from a software point of view, taking into
account both the executed programs and the underlying hardware architectures.
It aims at proposing new compiler-level optimizations to control peak power
[Benini2003a,Graybill2002a], hence thermal peaks, since temperature is
directly related to peak power [Skadron2004a]. Indeed, controlling temperature
is critical in mobile devices, for the comfort of the user, the reliability of
electronic components and the battery autonomy.
Moreover, although there is currently a lack of efficient high-level battery
models, it is known that the best usage pattern consists in minimizing current
peaks down to the average supply current. So it may be necessary to slow down
computation [Wu2005a,Wu2006a], thus spreading the required processing more
evenly throughout program execution, in order to control peak power. But when
execution time increases, energy also rises up and battery autonomy is
reduced. Thus, for a given context, the trade-off between time and power must
be found inside the technological constraints.
Such a problem, considering both thermal (through power) and energy limits,
has not be undertaken yet. We thus propose to integrate to the compiler
optimizations new possibilities to adapt the computation intensity (e.g
parallelism) to the constraints imposed by the designer. To this end,
analysing program behaviour phases and modelling their impact on power and
temperature shall be necessary. Hot spots will also be detected in order to
focus code optimizations.
To carry out this work, the Open-PEOPLE hardware platform will be used to
characterize thermal behaviour relatively to peak power and measure power and
energy performances. The resulting thermal models will be integrated into the
software platform and estimation tools, thus making it possible to validate
the effectiveness of the proposed optimization techniques and contributing to
the building and extension of the platform.

[Benini2003a] Luca. Benini, Mahmut. Kandemir, J. Ramanujam, "Compilers and
Operating Systems for Low Power", Kluwer Academic Publishers, 2003,
[Graybill2002a] Robert Graybill and Rami Melhem, "Power aware computing",
Kluwer Academic Publishers, Norwell, MA, USA, 2002, ISBN 0-306-46786-0.
[Skadron2004a] Kevin Skadron, "Hybrid Architectural Dynamic Thermal
Management". Design, Automation and Test in Europe (DATE), 2004, 10-15
[Wu2005a] Qiang Wu, Margaret Martonosi, Douglas W. Clark, V.J. Reddi, Dan
Connors, Youfeng Wu, Jin Lee, David Brooks, "A Dynamic Compilation Framework
for Controlling Microprocessor Energy and Performance," micro, pp. 271-282,
38th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO'05),
[Wu2006a] Qiang Wu, Margaret Martonosi, Douglas W. Clark, Vijay Janapa Reddi,
Dan Connors, Youfeng Wu, Jin Lee, David Brooks, "Dynamic-Compiler-Driven
Control for Microprocessor Energy and Performance," IEEE Micro, vol. 26, no.
1,  p. 119-129,  Jan/Feb,  2006.

Co-supervision by UBS/LabSTIC (Lorient, France) and INRIA Nancy Grand Est
(Nancy, France).

The PhD thesis will take place in Nancy, France.
Possible stays at Lorient, France.

Approximately 1584 Euros / month during two years, 1666 Euros / month during
one year, with benefits (social security, health insurance, unemployment
insurance, retirement plan) already paid, but income taxes (about 10%) not paid.

Olivier Zendra, INRIA Nancy Grand Est (Olivier.Zendra at inria.fr)
Email subject: "Application for Open-PEOPLE PhD Thesis #3"

Olivier ZENDRA, Ph.D., INRIA Researcher  | Email: Olivier.Zendra at inria.fr
INRIA Nancy - Grand Est / LORIA,         | Phone: (+33) 354 958 407
Office C-124,                            | Fax:   (+33) 383 552 573
615 Rue du Jardin Botanique, CS 20101    | Web:   www.loria.fr/~zendra

More information about the ecoop-info mailing list