ctuning / ck-crowdtuning Goto Github PK

when new crowdtuning engine is ready, gradually implement

• program benchmarking
• automatic program modelling (hierarchical execution time/energy/IPC modeling (collection of models), feature selection, run-time adaptation)
• optimal algorithm selection
• validation of algorithm scalability
• analysis of numerical stability
• crowdsource detection of program/compiler bugs

See

• https://hal.inria.fr/hal-01054763
• http://arxiv.org/abs/1506.06256

We mainly support coarse-grain crowd-benchmarking/crowd-tuning, i.e. global stats per whole program or one main function. Next, we need to improve repository to keep stats per multiple functions/kernels within analyzed program (while keeping track of any optimizatoins on this level, i.e. compiler flags, passes, OpenCL/CUDA/OpenMP/MPI parameters, etc). This will be useful to unify Caffe/TensorFlow/BLAS crowd-benchmarking and crowd-tuning under one engine ...

I try to pull a repo by "ck pull repo:ck-crowtuning" and the following message shows up:

cd /home/tao/CK/ck-crowtuning
git clone https://github.com/ctuning/ck-crowtuning /home/tao/CK/ck-crowtuning

Cloning into '/home/tao/CK/ck-crowtuning'...
Username for 'https://github.com':

And if I input my own github username and password, the following message shows up:
remote: Repository not found.
fatal: repository 'https://github.com/ctuning/ck-crowtuning/' not found

WARNING: repository update likely failed OR IN A DIFFERENT BRANCH/CHECKOUT (git exit code: 32768)

Would you like to continue (Y/n)?: Y
CK error: [repo] problem writing text file=/home/tao/CK/ck-crowtuning/.ckr.json ([Errno 2] No such file or directory: '/home/tao/CK/ck-crowtuning/.ckr.json')!

We have more and more useful statistics about optimizations in cknowledge.org/repo as well as DNN working via CK, so it would be interesting to try to apply DNN to figure out useful features in programs, compiler IR, binaries, data sets, OS, hardware which could predict such optimizations. It can be a nice internship or a GSOC project.

See related ticket for MILEPOST GCC: ctuning/reproduce-milepost-project#2

XYZ issue created by Giree

Current shared workloads are from cTuning/MILEPOST projects and are quite outdated. We need to provide more workloads and include ones from GCC/LLVM test suites. We should also add BLAS, DNN, etc ...

See related issue from ck-env repo: ctuning/ck-env#7

I'm having a problem with this.

After installing ck crowdtuning with:
$ git clone http://github.com/ctuning/ck
$ PATH=$PWD/ckbin:$PATH
$ ck pull repo:ck-crowdtuning

When running:
$ ck crowdtune program --gcc

The exepriments goes to completion, but apparently some tools are missing and in particular frequency settings end up to be skipped apparently.

ck does not find apparently these programs:

• ck-print-gpu-freq
• ck-set-gpu-performance
• ck-set-cpu-performance

I get these messages in the output:
...
GPU name: NVIDIA Corporation G98 [Quadro NVS 295](rev a1)
GPU vendor: NVIDIA
Trying to read GPU frequency:
ck-print-gpu-freq >/tmp/tmp-ck-Qlb8TQ
sh: ck-print-gpu-freq: command not found

Non-zero return code :32512 - likely failed
...
Setting CPU frequency to max (if supported) ...
CMD to set CPU frequency:
ck-set-cpu-performance
sh: ck-set-cpu-performance: command not found

Warning: setting frequency possibly failed - return code 32512
...
Setting GPU frequency to max (if supported) ...
CMD to set GPU frequency:
ck-set-gpu-performance
sh: ck-set-gpu-performance: command not found

Warning: setting frequency possibly failed - return code 32512

When running collaborative program optimization on mobile devices, our Android app should request only small data sets (i.e. < 1Mb), otherwise there is a very heavy traffic which can be problematic, if mobile device doesn't use WiFi ...
I think we should just provide an extra tag that the dataset is small ...

I noticed in our public repository that we may record the same optimizations as different solutions (unroll loops, etc). We should fix that and merge those results...