Meeting 01-08-2014

MAGEEC Meeting 01/08/2014 - Bristol

WP2

• GCC:
  • New patch available to extract more features (41 features)

• LLVM:
  • No new pass manager available. Problem with plugins.
  • Shall we write a plugin interface? No. No time. → changes or interface with releases.
  • Feature Extractor has had no progress. Needs some.
• [SC] contact David Malcom at Redhat re abstracting GCC plugin interface to make consistant across releases.

WP5

• All done.
• Review of initial implementation was Greg's blog post.

WP6

• How to get per-function energy basis?
  • Sampling with triggers have too much latency/error.
  • Template-based approach (a la crypto)?
  • Use cycle-accurate simulation to predict when functions occur and sample only at these points.
  • [SC to give priority to solving the measurement problem]

• James measurement errors:

• AVRs with non-grounded pins have big energy variation → need to ground.
• AVR position in ZIF socket has a 5% variation left → right
• [JP] Blog post on how not-to measure

• Board-board variation of energy consumption of ATMEGA328 ~10%. Batch is “1404”

WP7

[JB] to chase Joern to see what progress is.

WP8

• Currently using -O0 with some prediction, selecting from -O2 passes. Has made some decisions, but query how much different from base -O2 we have really got.

• George: 7h to do 2.5 runs of BEEBS V2 on one board. JP says it should take 4000s to do this. Problem is time-outs.
• AB has rolled out compilation patches for BEEBS.
  • DejaGNU is working.
  • V2 BEEBS still on target for end-of-August release
• Data variance:
  • The order of data elements as well as number can influence.
  • Best/worst/average case analysis.
    • Large number of runs would be needed
    • Ideally, we'd auto-generate best/worst cases or hand-program cases
• [Future RQ] A separate evaluation: for those categorisable programs and data sets, we can hand craft tests to see how MAGEEC responds on best/worst/avg cases
• Case studies
  • [SG] to run MAGEEC over the weather station.
  • [SG] get the satellite MSP430 code working
  • RTOS (battery manager on RTOS) – perhaps too ambitious for this project.

Next project review is 8^th September in Lymington @ 10:30am

For this, two major targets:

• Working case studies to demonstrate
• BEEBS V2.0 release

WP9:

Paper:

[SH+KIE] by next meeting structure of monster MAGEEC journal paper with all authors, describing process of developing a system with machine learning, plus the evaluation. Adds in the software engineering + evaluation. Not a push on the novelty of the machine learning.

• What's the story
• What results do we need?

• BEEBS paper. Explains V2.0 benchmarks suite. Evaluation after MAGEEC.

• Linux Plumbers event in Dortmund? (ENTRA+MAGEEC)
• Innovate UK

• [JB] Update events page
• [SH] Add CASES + Craig's ILP (using BEEBS) to page
• FOSDEM 2015
  • Proposal for 1 day workshop on compilers (morning: compilers; afternoon: compilers and energy efficiency)
• [SC+ABOpen] BEEBS.eu needs setting up.
  

• Blog post schedule (roughly chronological)
  • [ABOpen] “If you want a wand, put orders in now”. End of September target for fabrication.
  • [GC] Update PCA blog draft and publish
  • [OR] Blog post on what we are not looking at in project and ILP paper.
  • [SG] Case study of weather station with MAGEEC running
  • [GF] Initial analysis of pass effectiveness.
  • [JB] GNU cauldron
  • [Aburgess] BEEBS 2.0 (when released)

• “WAND” approved. Backronym to be considered ;)
• [ABOpen] Update wiki accordingly.

Summer work

George:

• Need to improve the running of the tests

With use of Placket-Burnham analysis to select tests, we need to see which flags had what weighting for

Experimental design (stats)

TODO: Ask if there are any EngMaths people who can help with the stats of the work.

Mann-Whitney comparison algorithms,

Bootstrap, Jacknife sampling algorithms.

[2 Week goal]: Minimal set of analysis done.

Greg

Some BEEBS work, keeping AB happy.

• PCA anlysis on BEEBS V2 per function feature vectors for x86 (680 total).
  • Looking at variation of features
  • TODO: Extend to include energy.

• [2 week goal: GC ]: Apply C5 decision tree on the PCA results to see if you can produce a (supervised learning) prediction for energy based on PCAs. Builds on Moon's work from last year.
  • OR has a lot of knowledge about how to go about this.
    

• [JB] To talk to Atmel about what output from cycle-accurate simulation can be made available.

Stoil

• GIMPL Tree analysis of instruction distribution of benchmarks (BEEBS V1)
• Also AVR for static and dynamic
• Reason is to check that BEEBS has a broad spread.
• How to group the instructions?
  • In a way that includes e.g. compare and skip (count it twice?)
    • Or between register and non-register architectures
  • Data movement should be split between register-register and memory operations.
  •  ? Do we separate read and write?

• Suggested set:

1. 1. ALU operations (inc. compare)
   2. Memory ops
      1. Each access to a memory logged separately
      2. Reads/writes
   3. Move register/accumulator
   4. Control flow
   5. Floating point

Priorities:

1. [2 week goal] Get the data for BEEBS V2 in AVR and Cortex M3/4. Post on github. Blog post
2. Then weather station running MAGEEC'd code.