SLIDE 1
GI using Deep Parameter Tuning
Mark Harman Yue Jia Jens Krinke Wes Weimer Fan Wu
Functional Requirements Non-Functional Requirements
Why GI for non-functional properties
humans have to define these a machine can
- ptimise these
GI using Deep Parameter Tuning Mark Fan Wu Wes Weimer Yue Jia - - PowerPoint PPT Presentation
GI using Deep Parameter Tuning Mark Fan Wu Wes Weimer Yue Jia - - PowerPoint PPT Presentation
GI using Deep Parameter Tuning Mark Fan Wu Wes Weimer Yue Jia Jens Krinke Harman Why GI for non-functional properties Non-Functional Functional Requirements Requirements humans have to a machine can define these optimise these
SLIDE 2
SLIDE 3
Traditional GI Process
Program
SLIDE 4
Traditional GI Process
Program
SLIDE 5
Traditional GI Process
Program
GP
SLIDE 6
Traditional GI Process
Program
GP Test cases Fitness
SLIDE 7
Traditional GI Process
Program
GP Test cases Fitness
Program’
SLIDE 8
Traditional GI Process
Program
GP Test cases Fitness
Program’
SLIDE 9
Parameter Tuning
Program
Tuneable Parameters
e.g. GCC -O3 foo.c
SLIDE 10
Parameter Tuning
Program
Tuneable Parameters
e.g. GCC -O3 foo.c
Good alternative for GI?
SLIDE 11
Parameter Tuning
Program
Tuneable Parameters
Not Many Tuneable Parameters Limited power Good alternative for GI?
SLIDE 12
Parameter Tuning
Program
Tuneable Parameters
Many software systems contain undocumented internal variables or expressions that also affect the performance of the systems.
SLIDE 13
Parameter Tuning
Program
Tuneable Parameters
Many software systems contain undocumented internal variables or expressions that also affect the performance of the systems.
Program
Tuneable Implicit Parameters Tuneable Parameters
Our approach
SLIDE 14
Deep Parameter Tuning
Program
Tuneable Implicit Parameters Tuneable Parameters (Shallow Parameters)
(Deep Parameters) NSGAII Opentuner Dynamic Knobs
SLIDE 15
Is this GI?
Genetic Improvement (GI) aims to find improved versions of existing programs that retain some partial semantics of the
- riginal (and possibly some
- f its syntax too)
SLIDE 16
Comparing to GI
EASY Lightweight Easy to use Less fragile
SLIDE 17
Our approach
Program
GP Test cases Fitness
Program’
SLIDE 18
Our approach
Program
SLIDE 19
Our approach
Program
Mutation Testing Test cases
SLIDE 20
Mutation Operators
C Mutation Testing tool : Milu
SLIDE 21
Our approach
Program
Mutation Testing Pseudo Equivalent Mutants: Mutants are not killed by any tests
SLIDE 22
Our approach
Non-dominated Rank Pseudo Equivalent Mutants
SLIDE 23
Our approach
Exposing Deep Parameters
SLIDE 24
Our approach
Exposing Deep Parameters Given a location L, and E is the expression at L L arithmetic expression logical expression
SLIDE 25
Our approach
Exposing Deep Parameters Given a location L, and E is the expression at L L arithmetic expression logical expression
SLIDE 26
Why to apply GI?
Program GI
SLIDE 27
Why to apply GI?
Program Program library GI GI library library GI GI
SLIDE 28
Why to apply GI?
Program Program library GI GI library library GI GI
General purpose memory allocator dlmalloc by Doug Lea
SLIDE 29
Experiment Subjects
SLIDE 30
Experiment Settings
9 Deep Parameters Repeat 10 times (cpu time and memory) Repeated 20 times
SLIDE 31
Experiments
Espresso gawk flex sed dlmalloc dlmalloc dlmalloc dlmalloc
Shallow Parameters Only Deep + Shallow Parameters Only
SLIDE 32
Experiments
Espresso gawk flex sed dlmalloc dlmalloc dlmalloc dlmalloc
Shallow Parameters Only Deep + Shallow Parameters Only Random NSGAII
SLIDE 33
Results : Optimisation Quality
SLIDE 34
Results : Optimisation Quality
SLIDE 35
Results: Reduction
SLIDE 36
Results : Execution cost
SLIDE 37
Deep Parameter Example
Managed by allocator
Heap
SLIDE 38
Deep Parameter Example
Managed by allocator In use
Heap
SLIDE 39
Deep Parameter Example
Managed by allocator In use
Heap
SLIDE 40
Deep Parameter Example
Managed by allocator In use
Heap
SLIDE 41
Deep Parameter Example
Managed by allocator In use
Heap
SLIDE 42
Related work
Software Tuning panel for Autonomic Control (STAC) Autotuner / PowerDial
SLIDE 43
Summary
Mutation analysis based approach to expose deep parameters 12% on time, 21% on memory
Input Domain ?
Program Program library GI GI