Document DASK_JIT_UNSPILL

[madsbk]

Fixes #597

Also moving Spilling from device to the ADDITIONAL FEATURES section

[codecov-commenter]

Codecov Report

❗ No coverage uploaded for pull request base (branch-21.06@1d71749). Click here to learn what that means.
The diff coverage is n/a.

❗ Current head 151f436 differs from pull request most recent head 2794ce8. Consider uploading reports for the commit 2794ce8 to get more accurate results

@@               Coverage Diff               @@
##             branch-21.06     #604   +/-   ##
===============================================
  Coverage                ?   90.31%           
===============================================
  Files                   ?       15           
  Lines                   ?     1569           
  Branches                ?        0           
===============================================
  Hits                    ?     1417           
  Misses                  ?      152           
  Partials                ?        0           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1d71749...2794ce8. Read the comment docs.

[quasiben]

This is great @madsbk, thanks for working on it! It's probably worth mentioning the benefits (and potential problems) of using it. Something like, "when experimenting with JIT-UNSPILL we generally see better performance and lower memory usage". @beckernick do you have some general idea on percentages here ? Is it 20% better?

[beckernick]

@beckernick do you have some general idea on percentages here ? Is it 20% better?

This will be workload dependent. However, for workloads that require significant spilling (such as large joins on infrastructure with less available memory than data) we have often seen greater than 50% improvement (i.e., something taking 300 seconds might take only 110 seconds). For workloads that do not, we would not expect to see much delta.

[madsbk]

@quasiben updated the docs to include some more information

[quasiben]

Thanks @madsbk !

[quasiben]

@gpucibot merge