MILEPOST Project

1
MILEPOST Project
Machine Learning for Embedded Programs
Optimisation

• Grigori Fursin, PhD
• Research Scientist at INRIA, France

2
Agenda / content

• Motivation
• MILEPOST project introduction
• MILEPOST framework
• Preliminary results
• Dissemination through open source software
• Conclusions

3
Motivation
Current innovation in science and industry demand
ever-increasing computing resources while placing
strict requirements on system performance, power
consumption, size, response, reliability,
portability and design time. High-performance
computing systems tend to evolve toward complex
heterogeneous multi-core systems.
dramatically increased optimization time
4
Motivation
Current innovation in science and industry demand
ever-increasing computing resources while placing
strict requirements on system performance, power
consumption, size, response, reliability,
portability and design time. High-performance
computing systems tend to evolve toward complex
heterogeneous multi-core systems.
dramatically increased optimization time
Optimizing compilers play a key role in producing
executable codes quickly and automatically while
satisfying all the above requirements for a broad
range of programs and architectures.
5
Motivation

• Developing and tuning current compilers for
  rapidly evolving architectures is a tedious and
  time consuming process
• - simplistic hardware models for rapidly
  evolving hardware
• - inability to reuse optimization knowledge
  among different programs and architectures
• - lack of run-time information and inability to
  adapt to varying program and system behavior (or
  dataset) at run-time with low overhead
• Humans can't keep up

6
Motivation

• Developing and tuning current compilers for
  rapidly evolving architectures is a tedious and
  time consuming process
• - simplistic hardware models for rapidly
  evolving hardware
• - inability to reuse optimization knowledge
  among different programs and architectures
• - lack of run-time information and inability to
  adapt to varying program and system behavior (or
  dataset) at run-time with low overhead
• Humans can't keep up
• Need intelligent self-tuning compilers that can
  continuously and automatically learn how to
  optimize programs, and have an ability to make
  program adaptable at run-time for different
  behavior and constraints

7
MILEPOST project
Machine Learning for Embedded Programs
Optimization http//www.milepost.eu
Objective is to develop compiler technology that
can automatically learn how to best optimize
programs for re-configurable heterogeneous
embedded processors (with run-time adaptation)
and dramatically reduce the time to market.
Partners
Funded by Information Society Technologies (IST)
of the European Union under FP6
8
MILEPOST framework
Building the training set for learning
9
MILEPOST framework
Building the machine learning model
10
MILEPOST framework
Predicting good optimizations for a new program
11
Preliminary Results
Predicted speedups on ARC 725D processor and
MiBench benchmark for embedded systems
Details can be found in the following
paper Grigori Fursin, Cupertino Miranda, Olivier
Temam, Mircea Namolaru, Elad Yom-Tov, Ayal Zaks,
Bilha Mendelson, Phil Barnard, Elton Ashton, Eric
Courtois, Francois Bodin, Edwin Bonilla, John
Thomson, Hugh Leather, Chris Williams, Michael
O'Boyle. MILEPOST GCC machine learning based
research compiler. Proceedings of the GCC
Developers' Summit, Ottawa, Canada, June 2008
12
Dissemination through open source software

• All tools developed in MILEPOST project are
  publicly available or will be available in the
  near future as a free open source software
• MILEPOST website
• http//www.milepost.eu
• GCC Interactive Compilation Interface
• http//gcc-ici.sourceforge.net
• Hence it is easy to extend this software through
  collaboration with academic institutions and
  industry. Some technology can later be
  commercialized.

13
Conclusions
We believe that MILEPOST project opens many new
exciting research directions and will help to
create future self-tuning intelligent systems
that is necessary to continue innovation in
science and industry. We plan to use statistical
machine learning techniques for the following
topics - Continuous program and architecture
optimization, re-configuration and adaptation
(beyond the separation between compiler, runtime
and hardware) - Automatic parallelization and
new programming paradigms - Programmability,
optimizing and simulation tools, run-time
adaptability for re-configurable heterogeneous
multi-core computer architectures
14
Thank you

• Contact and details
• http//www.milepost.eu
• grigori.fursin_at_inria.fr
• http//fursin.net/research

SMART09 3rd Workshop on Statistical and Machine
learning approaches to ARchitectures and
compilaTion http//www.hipeac.net/smart-workshop.h
tml Sponsored by MILEPOST