Add best practices page to Dask cuDF docs #16821
Conversation
rjzamora
added
2 - In Progress
|
@jacobtomlinson @quasiben @pentschev - Interested in your feedback on the specific "best practices" guidelines I added here. Happy to revise. |
| Deployment and Configuration | ||
| ---------------------------- | ||
|
|
||
| Use Dask DataFrame Directly |
There was a problem hiding this comment.
Not sure what "section" this belongs to.
There was a problem hiding this comment.
Maybe just move it to the end of the Deployment and Configuration section? The rationale is this at the top may look like a suggestion, which is actually the opposite of what this subsection says. No strong opinions though.
| The ideal partition size is typically between 2-10% of the memory capacity | ||
| of a single GPU. Increasing the partition size will typically reduce the |
There was a problem hiding this comment.
Should we provide here a rule-of-thumb as to whether users should initially target more to the 2% or the 10% range, and how/when to increase/decrease that? Or is this too difficult to provide a good rule-of-thumb and the 2-10% phrasing is the best we can do? I understand it can be quite difficult to give more details for general purpose docs, so it's fine if you think the current phrasing is sufficient/best.
There was a problem hiding this comment.
Okay, I attempted to turn this into the more-explicit "rule of thumb" I personally use: 1/16 or less if the workflow is memory-intensive (i.e. shuffle intensive), and 1/8 otherwise. The "best" partition size is definitely difficult to know a priori.
|
|
||
| ``blocksize``: Use this argument to specify the maximum partition size. | ||
| The default is `"256 MiB"`, but larger values are usually more performant | ||
| (e.g. `1 GiB` is usually safe). Dask will use the ``blocksize`` value to map |
There was a problem hiding this comment.
Maybe provide a guideline for when 1 GiB is safe. I imagine this is safe for large devices that we're usually used to work with, but given the recent NO-OOM effort I don't think a small laptop GPU will be capable of handling 1GiB safely.
There was a problem hiding this comment.
I decided to remove the 1 GiB comment since we already discuss the 1/16-1/8 "rule of thumb" above.
| many of the details discussed in the `Dask DataFrames Best Practices | ||
| <https://docs.dask.org/en/stable/dataframe-best-practices.html>`__ | ||
| documentation also apply to Dask cuDF. |
There was a problem hiding this comment.
Are there any notable ones that are known NOT to apply to Dask cuDF and we should let users know here?
There was a problem hiding this comment.
I updated the wording to say the guidelines that are not pandas-specific also apply to Dask cuDF.
| Deployment and Configuration | ||
| ---------------------------- | ||
|
|
||
| Use Dask DataFrame Directly |
There was a problem hiding this comment.
Maybe just move it to the end of the Deployment and Configuration section? The rationale is this at the top may look like a suggestion, which is actually the opposite of what this subsection says. No strong opinions though.
| Sorting, joining and grouping operations all have the potential to | ||
| require the global shuffling of data between distinct partitions. |
There was a problem hiding this comment.
This also applies to repartition, no? Maybe we should mention the same arguments here or point https://github.com/rapidsai/cudf/pull/16821/files#diff-1c3b287013ea5f3b56726f0b7e7538bd0242e8cadcbd89e966a82d3d36719317R93-R95 to here as well to make it more explicit where users are dealing with in those cases.
There was a problem hiding this comment.
Hmmm. Repartition does not really require data "shuffling". Data shuffling requires "all-to-all", while repartitioning is usually limited to data movement between neighboring partitions.
| * Use a distributed cluster with Dask-CUDA workers | ||
| * Use native cuDF spilling whenever possible |
There was a problem hiding this comment.
Should we link to the sections above in here that deal with these two suggestions, for the benefit of the reader who may come directly to this section? Not sure if easy/possible, if not skipping is also fine.
There was a problem hiding this comment.
Yeah, good idea. Will need to figure out how to do that :)
rjzamora
added
3 - Ready for Review
|
@VibhuJawa @ayushdg @randerzander - I would consider you all dask-cudf power users. Let me know if these "best practices" seem reasonable to you. |
| ``False``, but ``aggregate_files=True`` is usually more performant when | ||
| the dataset contains many files that are smaller than half of ``blocksize``. | ||
|
|
||
| .. note:: |
There was a problem hiding this comment.
I like this note. Once we have more cloud IO specific optimizations it might make sense to add it to best practices or create a new one for cloud IO to discuss tips/tricks for those environments.
There was a problem hiding this comment.
Agree that we need a lot more remote-IO information. However, it doesn't feel like there is much to say yet :/
|
I will plan to merge this in a few hours if there aren't any more comments. |
Co-authored-by: Lawrence Mitchell <[email protected]>
rjzamora
added
5 - Ready to Merge
|
/merge |
Description
Adds a much-needed "best practices" page to the Dask cuDF documentation.
Checklist