Status
This is a relatively stable repository for customizable experiment crowdsourcing across platforms provided by voluteers (such as mobile devices and data centers). Currently implemented scenarios include universal, customizable, multi-dimensional, multi-objective GCC/LLVM crowd-tuning. See public crowd-tuning results in the [http://cknowledge.org/repo CK live repository]!
Description
This repository is based on CK machine-learning based autotuning. It crowdsources experiments (using optimization knobs exposed via CK such as OpenCL, compiler flag, CUDA, etc) across many machines while building a realistic, large and representative training set.
This is a continuation of Grigori Fursin's original postdoctoral proposal for the MILEPOST project in 2005, i.e. crowdsource training of a machine-learning based compiler across any shared computational resource such as mobile phones (supported by the non-profit cTuning foundation since 2008).
Publications
- http://arxiv.org/abs/1506.06256
- http://bit.ly/ck-date16
- http://arxiv.org/abs/1407.4075
- http://hal.inria.fr/hal-01054763
- https://hal.inria.fr/inria-00436029
Authors
- Grigori Fursin, cTuning foundation (France) / dividiti (UK)
License
- BSD, 3-clause
Prerequisites
- Collective Knowledge Framework: http://github.com/ctuning/ck
Usage
See https://github.com/ctuning/ck/wiki/Crowdsource_Experiments
Notes
We and the community added various analysis of variation of empirical characteristics such as execution time and energy: min, max, mean, expected values from histogram, normality test, etc.
Users can decide how to calculate improvements based on available statistics and their requirements. For example, when trying to improve compilers or hardware, we compare minimal characteristics (execution time, energy, etc), i.e. the best what we can squeeze from this hardware when there are no cache effects, contentions, etc.
Later, we suggest to calculate improvements using expected values - we noticed that computer systems has "states" (similar to electron energy states in physics), hence such improvements will show how a given program will behave in non-ideal conditions.
Furthermore, when there is more than one expected behavior, i.e. several states, we suggest to analyze such cases by the community and find missing experiment features that could explain and separate such states such as CPU/GPU frequency.
See our papers for more details.