Paul Rogers created DRILL-5594:
----------------------------------
Summary: Excessive buffer reallocations during merge phase of
external sort
Key: DRILL-5594
URL: https://issues.apache.org/jira/browse/DRILL-5594
Project: Apache Drill
Issue Type: Bug
Affects Versions: 1.11.0
Reporter: Paul Rogers
Assignee: Paul Rogers
Priority: Minor
Fix For: 1.11.0
Consider the log file attached to DRILL-5513. The log shows an excessive number
of buffer reallocations while assembling the merged output of the sort:
{code}
2017-05-15 12:58:46,319 [26e5f7b8-71e8-afca-e72e-fad7be2b2416:frag:5:13] DEBUG
o.a.drill.exec.vector.BigIntVector - Reallocating vector
[$data$(BIGINT:REQUIRED)]. # of bytes: [32768] -> [65536]
2017-05-15 12:58:46,321 [26e5f7b8-71e8-afca-e72e-fad7be2b2416:frag:5:13] DEBUG
o.a.drill.exec.vector.UInt4Vector - Reallocating vector
[$offsets$(UINT4:REQUIRED)]. # of bytes: [16384] -> [32768]
...:5:13] DEBUG o.a.drill.exec.vector.UInt4Vector - Reallocating vector
[$offsets$(UINT4:REQUIRED)]. # of bytes: [16384] -> [32768]
...:5:13] DEBUG o.a.drill.exec.vector.UInt1Vector - Reallocating vector
[$bits$(UINT1:REQUIRED)]. # of bytes: [4096] -> [8192]
...:5:13] DEBUG o.a.drill.exec.vector.UInt1Vector - Reallocating vector
[$bits$(UINT1:REQUIRED)]. # of bytes: [4096] -> [8192]
...5:13] DEBUG o.a.drill.exec.vector.UInt1Vector - Reallocating vector
[$bits$(UINT1:REQUIRED)]. # of bytes: [4096] -> [8192]
...5:13] DEBUG o.a.drill.exec.vector.BigIntVector - Reallocating vector
[c(BIGINT:OPTIONAL)]. # of bytes: [32768] -> [65536]
..5:13] DEBUG o.a.drill.exec.vector.UInt1Vector - Reallocating vector
[$bits$(UINT1:REQUIRED)]. # of bytes: [4096] -> [8192]
...frag:5:13] DEBUG o.a.drill.exec.vector.Float8Vector - Reallocating vector
[d(FLOAT8:OPTIONAL)]. #
...
{code}
Hundreds of these lines appear. This means that the initial buffer allocation
is too small.
Given that the merge phase knows the number of rows it will put into the batch,
and has access to the "sizer" information to estimate Varchar widths, the merge
phase should predict, and allocate, the required buffer sizes to avoid repeated
reallocation. Each reallocation requires:
* Allocate new buffer (which puts memory pressure on the sort's memory)
* Copy data from old to new buffer
* Zero-fill the new half (zero-fill is not done on the first allocation,
strangely.)
* Free the unneeded original buffer.
Since the sort is already slow, the above extra work just makes the problem
worse.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
