[SPARK-53266][TESTS] Regenerate benchmark results #52005
Conversation
|
Thank you, @HyukjinKwon . |
| Compression 10000 times at level 1 with buffer pool 580 581 0 0.0 58038.0 1.1X | ||
| Compression 10000 times at level 2 with buffer pool 612 615 3 0.0 61246.1 1.1X | ||
| Compression 10000 times at level 3 with buffer pool 721 734 11 0.0 72106.4 0.9X | ||
| Compression 10000 times at level 1 without buffer pool 265 267 1 0.0 26513.9 1.0X |
There was a problem hiding this comment.
This becomes much faster.
| Compression 4 times at level 1 with buffer pool 2568 2571 4 0.0 642087500.8 1.0X | ||
| Compression 4 times at level 2 with buffer pool 4211 4212 1 0.0 1052833529.0 0.6X | ||
| Compression 4 times at level 3 with buffer pool 6290 6291 2 0.0 1572505716.0 0.4X | ||
| Compression 4 times at level 1 without buffer pool 2764 2764 1 0.0 690899114.5 1.0X |
There was a problem hiding this comment.
Although it's a known issue, this is still slower than Java 17.
| arrayOfAnyAsObject 6 6 0 1611.8 0.6 1.0X | ||
| arrayOfAnyAsSeq 174 175 1 57.5 17.4 0.0X | ||
| arrayOfInt 393 395 1 25.4 39.3 0.0X | ||
| arrayOfIntAsObject 419 419 1 23.9 41.9 0.0X |
There was a problem hiding this comment.
The relationship between arrayOfInt and arrayOfIntAsObject is switched.
| Common Codecs 4821 4894 64 0.2 4820.6 1.0X | ||
| Java 2565 2572 10 0.4 2564.8 1.9X | ||
| Spark 3811 3812 1 0.3 3810.7 1.3X | ||
| Spark Binary 2758 2759 1 0.4 2757.9 1.7X |
There was a problem hiding this comment.
The relationship between Java and Spark Binary is switched. Now, Java is faster than Spark.
| Without bloom filter, blocksize: 16777216 790 794 5 126.6 7.9 1.0X | ||
| With bloom filter, blocksize: 16777216 792 798 9 126.3 7.9 1.0X | ||
| Without bloom filter, blocksize: 16777216 827 835 10 120.9 8.3 1.0X | ||
| With bloom filter, blocksize: 16777216 536 542 5 186.5 5.4 1.5X |
There was a problem hiding this comment.
This become faster and consistent with Java 21 result.
| Without bloom filter, blocksize: 33554432 427 430 3 234.2 4.3 1.0X | ||
| With bloom filter, blocksize: 33554432 508 520 12 196.9 5.1 0.8X | ||
| Without bloom filter, blocksize: 33554432 507 520 10 197.1 5.1 1.0X | ||
| With bloom filter, blocksize: 33554432 444 465 32 225.3 4.4 1.1X |
There was a problem hiding this comment.
This becomes faster and expected one.
| UTF-16 52085 52137 74 0.2 5208.5 0.6X | ||
| UTF-8 30150 30156 9 0.3 3015.0 1.1X | ||
| UTF-32 56295 56403 153 0.2 5629.5 1.0X | ||
| UTF-16 50644 50653 13 0.2 5064.4 1.1X |
There was a problem hiding this comment.
UTF-32 becomes suddenly the slowest among all. Previously, UTF-16 is slowest in both Java 17 and 21.
| columnar deserialization + columnar-to-row 179 220 36 5.6 178.7 1.0X | ||
| row-based deserialization 171 219 70 5.9 170.5 1.0X | ||
| columnar deserialization + columnar-to-row 222 257 41 4.5 222.3 1.0X | ||
| row-based deserialization 140 178 63 7.2 139.8 1.6X |
There was a problem hiding this comment.
Row-base deserialization becomes much faster relatively.
|
I'm going to merge this as a new baseline for further work about the above observations. |
2 participants
What changes were proposed in this pull request?
This PR aims to regenerate benchmark results.
Why are the changes needed?
We have 3 goals.
RecursiveCTEBenchmark-jdk21-results.txtspark/sql/core/benchmarks/RecursiveCTEBenchmark-results.txt
Line 6 in ebf4dd1
Does this PR introduce any user-facing change?
No.
How was this patch tested?
Manual review.
Was this patch authored or co-authored using generative AI tooling?
No.