Mihaly Szjatinya has posted comments on this change. ( 
http://gerrit.cloudera.org:8080/24579 )

Change subject: IMPALA-15139: Support DELTA_LENGTH_BYTE_ARRAY Parquet encoding
......................................................................


Patch Set 21:

(1 comment)

Thanks for review!

http://gerrit.cloudera.org:8080/#/c/24579/16/be/src/benchmarks/parquet-delta-length-byte-array-benchmark.cc
File be/src/benchmarks/parquet-delta-length-byte-array-benchmark.cc:

http://gerrit.cloudera.org:8080/#/c/24579/16/be/src/benchmarks/parquet-delta-length-byte-array-benchmark.cc@31
PS16, Line 31: //    all-long   146 iters/ms   1X      (zero-copy: only pointer 
stored, no Smallify)
             : //    mixed M-L  137 iters/ms   0.94X   (short-first phase: 
S,L,S,L)
> >core warms up to its full turbo frequency
Re-ran as you suggested (no_turbo=1, scaling_governor=performance), pinned to a 
single P-core. The phase gap persists essentially unchanged: the M-L vs L-M gap 
is ~13% without turbo (stable across runs, +/-0.5%) vs ~15% with turbo. So the 
ordering effect is real and not a turbo/frequency-scaling artifact. This 
corrects my earlier note above; the effect that I observed must have been some 
other transient ramp.

The absolute throughput dropped ~7x (all-long ~147 -> ~20 iters/ms) as expected 
at fixed base frequency, but the ratios held. So I think we're fine with what's 
already commited.



--
To view, visit http://gerrit.cloudera.org:8080/24579
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings

Gerrit-Project: Impala-ASF
Gerrit-Branch: master
Gerrit-MessageType: comment
Gerrit-Change-Id: I9e4395fa30cd7fb2c89ef5a2b16b043d4de60b52
Gerrit-Change-Number: 24579
Gerrit-PatchSet: 21
Gerrit-Owner: Mihaly Szjatinya <[email protected]>
Gerrit-Reviewer: Balazs Hevele <[email protected]>
Gerrit-Reviewer: Csaba Ringhofer <[email protected]>
Gerrit-Reviewer: Impala Public Jenkins <[email protected]>
Gerrit-Reviewer: Mihaly Szjatinya <[email protected]>
Gerrit-Reviewer: Zoltan Borok-Nagy <[email protected]>
Gerrit-Comment-Date: Tue, 15 Sep 2026 13:49:23 +0000
Gerrit-HasComments: Yes

Reply via email to