General Observations

1
General Observations

• Importance of Data
• common glue underlying all tool components
• precise specification, storage, handling,
  management important
• database technology?
• Differences in Content
• prof ? gprof ? trace of profiles ? event traces
• Incompleteness of Data
• tool must be able to handle incomplete data
• needs to know how to get missing / additional
  information
• Conflicting / Contradictory Data
• profiles ? traces
• 1st run ? 2nd run

2
Experiments
Experiment
prepare / setup
DB
execute / run
control
analyze

• Types of experiments
• multi experiments (series of experiments)
• macro experiments (off-line tool full program
  execution)
• micro experiments (on-line tool one analysis
  loop)

3
Experiment Classification Space

• Technology (how)
• measurement
• profiling
• sampling, measured, ...
• traces of profiles
• event tracing
• modeling
• technique
• analytical, simulation, ...
• Goal
• extrapolation, detailed analysis, ...
• Metrics (what)
• time
• HW counters ...

• Coverage (where / when)
• location
• machine
• process
• thread
• program structure (regions)
• program
• module / file
• function
• basic block
• statement
• time
• complete execution
• partial

4
source repository
repository control system
source code
compilation system
executable
input data
output data
5
target config model
aspect model
extrapolation
model analysis
modeling results
6
(No Transcript)
7
Project / Subtask Description Form

• Name
• Description
• What part of the analysis process does it
  automate/support?
• Related / already existing tools / technology
• uniqueness w.r.t. these technologies
• reuse
• What market does it serve?
• Extent
• student research project
• PhD
• research project
• esprit rdt
• plus Duration (person years)

8
Project / Subtask Description Form (cont.)

• Input / Output specifications /
  requirements(external interface)
• APIs
• standards which can be used
• possible standards which should be defined
• Supported platforms
• programming languages
• programming models
• machines
• Evaluation of success
• validation suite
• benchmark suite
• metrics

9
Theory (Methodology) ToDo List

• T1) Properties Relationship structure (Tree?
  Graph?) structure walking algorithms to
  identify experiments to satisfy queries
• T2) Refine "Confidence" concept (data model
  update / refinement) explore uses / association
  of quantity of data
• T3) Experiment description language for
  expressing queries answers from experiments
  abstracted away from details how they are
  answered
• T4) Add distributed computing model (client/serve
  r, multi-tier, ...)
• T5) Add hierarchic model (MPIOpenMP)
• T6) Add event trace model to data model

10
Implementation Projects

• I1) Tool for checking performance properties
  with data access model (implementing properties
  as Java classes). Validate by hand on "trivial
  program.
• I2) "Demonstration Prototype"
• Initial Implementation of very simple
  experiment cycle
• I2a) static analyzer / constraints /
  prof/gprof like measurements
• I2b) based on message passing (MPI?) tracing
• I3) "Prototype 2nd Generation"
• - add experiment management, optimization
  planning
• - validate on "trivial" programs
• I4) query gt answer by experiment using existing
  tools main topic tool interoperability,
  wrappers, "standard" APIs, ...

11
Implementation Projects (cont.)

• I5) "Prototype 3nd Generation"
• I5a) use machine learning techniques
• I5b) prioritization of constraints
  (predicted data may be replaced by
  monitor data later)
• I6) User Interface Design
• I7) user guidance ("wizard") for existing
  performance tools
• I8) automatic instrumentation wrap existing
  instrumenters (compiler tools)
  providing "standard" instrumentation interface
• I9) automatic experiment support repeated
  execution of applications with varying
  CPU numbers / input data to determine speedup
  /scalability / stability of results

12
Implementation Projects (cont.)

• I10) automatic location of bottlenecks
• I10a) based on prof/gprof like measurements
• I10b) based on message passing (MPI?)
  tracing
• I11) automatic comparison of different
  experiments of the same program -
  detection of the main differences - summarizing
  results