improve bitmap vector offset to report contiguous groups#11039
improve bitmap vector offset to report contiguous groups#11039gianm merged 3 commits intoapache:masterfrom
Conversation
| currentVectorSize = to; | ||
| } | ||
|
|
||
| if (currentVectorSize > 1) { |
There was a problem hiding this comment.
Why not do this branch if currentVectorSize == 1?
There was a problem hiding this comment.
we could, and I considered it, but ultimately I decided that 1 didn't really "feel" like it met the threshold for me to consider it really contiguous. I also figured it didn't really make a lot of difference between which read method it would end up in because there is only 1 offset either way. But I don't have strong opinions on it though since either seems valid, do you think it logically should be considered contiguous?
There was a problem hiding this comment.
I guess it doesn't really matter. I thought I might have been missing some reason why it wasn't ok to do it for the currentVectorSize == 1 case.
| @Override | ||
| public int getStartOffset() | ||
| { | ||
| if (isContiguous) { |
There was a problem hiding this comment.
According to the interface, getOffsets is supposed to throw an exception if isContiguous is true, so we should check it there too.
2 participants
Description
This PR is a follow-up to #11004, which improves
BitmapVectorOffsetto report the current set of offsets as "contiguous" when they are, in order to take advantage of any optimizations available when processing contiguous vectorized gets.In #11004, I noticed the improvements to vectorized decoding of long values in my column scan benchmarks only seemed to impact a full scan. Looking closer, it was due to how the benchmark uses a
BitmapVectorOffsetto simulate filtering rows, which prior to this pr would never report the current set of offsets as contiguous, even if it was. Since #11004 primarily improved performance of contiguous vectorized gets, when using aBitmapVectorOffsetwe were missing out on these improvements entirely.To help measure the improvements of allowing contiguous blocks to be processed with the contiguous get methods, I improved the benchmarks added in #11004 to now have several different filter "distributions" to simulate different filtering patterns, instead of previously only using a random distribution. These include:
BitmapVectorOffset, bitmap is set randomly for desired number of rowsNoFilterVectorOffsetfrom starting offset 0 to desired number of rowsBitmapVectorOffset, bitmap set from 0 to desired number of rowscontiguous-bitmap-startBitmapVectorOffset, position chosen randomly and up to 1000 contiguous values set until chosen number of rows is setThe results of this change is a pretty decent improvement to vectorized column read speed when there are contiguous vectors to read:
before:
after:
This PR has: