[BUG] validation failure of power run results on gpu with nvcomp and power run results on cpu about spark-rapids-benchmarks HOT 7 CLOSED

Please try using the latests rapids-4-spark_2.12-23.12.0 release from artifactory.

Also, were there any errors in the power-run output? You should capture the driver/executor output and see if there were any errors.

from spark-rapids-benchmarks.

To attempt to repro this, I ran the full nds power run at scale 100 on my desktop for orc/snappy. Output from the cpu and gpu matched for all of the queries except for this one known difference:

=== Comparing Query: query79 ===
Collecting rows from DataFrame
Collected 100 rows in 0.11637043952941895 seconds
Collecting rows from DataFrame
Collected 100 rows in 0.10641264915466309 seconds
Row 99: 
[None, None, 'Bethel', 12213922, None, None]
[None, None, 'Bethel', 22767968, None, None]

The difference is in the ss_ticket_number field, which is in the group by key, but not in the sort order for the output, so this is likely just ambiguous ordering in the output. I have verified that I get the same result on GPU with different compressors. This is a difference you may need to account for when doing comparisons.
Note that I built with the cuda11 jars.

from spark-rapids-benchmarks.

@yinqingh can you upload the query file and maybe point us to the input orc data path? if you still facing such diff after using the latest snapshot jar.

from spark-rapids-benchmarks.

The issue is found on the released nvcomp version is 3.0.3, we will check if the latest snapshot version of nvcomp still has the issue. If yes, well update it here.

from spark-rapids-benchmarks.

from spark-rapids-benchmarks.

@jbrennan333 @wjxiz1992 Thanks for the information! We will keep testing with the latest nvcomp and update here if the same issue can be repro in 3.0.4.

from spark-rapids-benchmarks.

Since we did not observed this issue with the daily rapids-plugin jar(containing the 3.0.4 nvcomp libs), I suppose we can close this issue first.

If it is repro again, we can re-open later this later.

from spark-rapids-benchmarks.