removed CLI table. Added first results of runtime verification.

mpicbg-scicomp / gearshifft_publication

Publication Manuscript for results obtained with gearshifft

Other

0 stars 1 forks source link

removed CLI table. Added first results of runtime verification. #47

Closed tdd11235813 closed 7 years ago

tdd11235813 commented 7 years ago

~ WIP ~

This PR relates to #46

tdd11235813 commented 7 years ago

added some plots to illustrate gearshifft's almost zero overhead (for K80, cufft, inplace r2c, float). text is not finished yet, so wait with merging.

psteinb commented 7 years ago

great to see that everything is coming together. I am looking forward to the description. Please document the steps taken to produce the data. maybe in a new directory scripts/validation. we also need to document in the publication repo how we obtained the gearshifft csv files. feel free to put slurm scripts into the respective results directory. I;d like to try to make this publication as reproducible as possible.

tdd11235813 commented 7 years ago

merged your PR #48 #49 #50 #52 #53. That was fun, git diff --word-diff did a good job to guide me through this battlefield, at least I hope to get all changes right. I now would proceed with this PR and your recommendations.

psteinb commented 7 years ago

sounds good to me. good luck!

tdd11235813 commented 7 years ago

I suggest to put the cufft standalone code and the validation + run scripts into gearshifft repo, since it is relevant there, too. And it would be more convenient to refer to it in the paper, what do you think?

tdd11235813 commented 7 years ago

... hmm a big disadvantage is, that each user of this repository will download these files as well, which I would dislike. well, I gonna put these files into publication repo as you suggested; we should refer to it in the gearshifft repo, so we can keep the reference to gearshifft github url in the paper.

tdd11235813 commented 7 years ago

improved text on gearshifft validation
code for validation and job scripts added
- for the sake of correctness I rerun the validation and I am waiting for the job scheduler again (Taurus is busy)
- (csv also mistakenly contained Inplace Complex instead of Inplace Real, which however did not affect the diagrams)
fftw wisdom explained more in detail (wisdom -> FFTW_PATIENT)
- in the results section I removed the warning which discourages consequently using wisdom
- ... I mentioned the phenomenon that PATIENT mode should provide faster transforms than MEASURE and although I used the same flags, we have to find out, if plan-search is deterministic -- even in a multi-thread environment (read doc or ask fftw devs)
reduced words in abstract to fit 150 words limit

psteinb commented 7 years ago

Fine with me.

psteinb commented 7 years ago

Ok, that's a lot of work you put into this PR. Thanks. I highly appreciate that. Feel free to merge at will.

psteinb commented 7 years ago

btw, you could think about moving the overhead/validation analysis into the implementation section. I think the context there would be more appropriate

tdd11235813 commented 7 years ago

this merge happens more due to its growing size. Still had no time to proofread all my text, I hope to finish it this afternoon ..

I also think about your suggestion to move the validation part to the implementation section (atm I would tend to leave it in the results section as it is more a result as a consequence of an implementation than an implementation detail).

psteinb commented 7 years ago

ok, that's all fine with me