Add cudf (RAPIDS)

[jangorecki]

https://github.com/rapidsai/cudf

[jangorecki]

As of now installation of rapids required anaconda which will heavily complicate reproducibility of this benchmark. Currently it is waiting for rapidsai/cudf#331 to be solved

[mattdowle]

Do the suggestions in rapidsai/cudf#331 work?

[jangorecki]

@mattdowle It didn't, but I asked @abal5 for help in setting up environment. He finished his part, I just need to get back to that.

[jangorecki]

As of now we were unable to setup cudf on our gpu machine. Soon it will hopefully be be easier once rapidsai/cudf#285 will be resolved, PR is already under review rapidsai/cudf#567

[mattdowle]

That cudf issue 285 contains this comment :

That does not sound too bad. Did you try that and it failed with error messages?
The 2nd issue (567) was merged yesterday.

[jangorecki]

Installation from pip is still not that easy going as normally: rapidsai/cudf#285 (comment)

[jangorecki]

I successfully installed cuda toolkit 9.2 and trying on it, as cuda toolkit 10.0 was raising errors trying to access 9.2 shared libraries. Currently stuck on rapidsai/cudf#736
fyi @abal5

[jangorecki]

slowly moving forward, stuck on NVIDIA/nvstrings#8

[jangorecki]

cudf can be now much more easily installed from pip, don't need to deal with conda envs anymore. Although it is not ready to be added to benchmark due to NVIDIA/nvstrings#8, rapidsai/cudf#847 and rapidsai/cudf#848, rapidsai/cudf#894

[jangorecki]

validation script needs to be adjusted for appropriate tolerance: rapidsai/cudf#2494

[jangorecki]

Current status is that q1-q5 are implemented, but unfortunately 1e8 and 1e9 are failing due to out of memory. 1e7 k=2 segfaults. Also 1e7 k=2 affected by rapidsai/cudf#2495 (possible that this caused a segfault)

[jangorecki]

cuDF has been added to groupby benchmark fully, answers validated on 1e7, k=1e2 (head and tail).
I will close this issue once fresh run of benchmark timings will get published.

---

Pending issues related to cuDF:

• extend GPU memory to run cuDF for medium and big data extend GPU memory to run cuDF for medium and big data #97
• investigate cuDF memory usage investigate cuDF memory usage #94
• validation of chk should have tolerance within single solution validation of chk should have tolerance within single solution #95
• q6-q9 not yet implemented in cuDF, add when ready
  • q6: median v3 sd v3 by id2 id4 [FEA] Series.median() rapidsai/cudf#1085
  • q7: max v1 - min v2 by id2 id4 Support callables in DataFrame.assign rapidsai/cudf#2591
  • q8: largest two v3 by id2 id4 [FEA] top N rows by group rapidsai/cudf#2592
  • q9: regression v1 v2 by id2 id4 [FEA] libcudf Series correlation (Pearson) rapidsai/cudf#1267