Thread: Performance gain from reduction of GROUP BY memory allocations
In PostgreSQL Weekly News, David Fetter wrote: > Please test the new beta. Some of the new features are at > http://developer.postgresql.org/docs/postgres/release.html#RELEASE-8-1 I notice that Neil's patch regarding reducing the number of memory allocations during aggregation operations isn't mentioned. It was originally discussed in 8.0beta (2-3?) time. What happened there? - patch not committed in the end - committed but not mentioned, as a dropped item - committed but not mentioned, since part of a larger patch Seemed like a good performance gain to me... I *have* tried to find this, but Neil's work on memory allocation has been extensive... (please take the compliment...). Best Regards, Simon Riggs
Simon Riggs <simon@2ndquadrant.com> writes:
> I notice that Neil's patch regarding reducing the number of memory
> allocations during aggregation operations isn't mentioned. It was
> originally discussed in 8.0beta (2-3?) time.
> What happened there?
> - patch not committed in the end
> - committed but not mentioned, as a dropped item
> - committed but not mentioned, since part of a larger patch
Are you speaking of these patches?
2005-04-06 19:56 neilc
* src/backend/utils/adt/: float.c, numeric.c: Apply the "nodeAgg"optimization to more of the builtin transition
functions.Thispatch optimizes int2_sum(), int4_sum(), float4_accum() andfloat8_accum() to avoid needing to copy the
transitionfunction'sstate for each input tuple of the aggregate. In an extreme case(e.g. SELECT sum(int2_col) FROM
tablewhere table has a singlecolumn), it improves performance by about 20%. For more complexqueries or tables with
widerrows, the relative performanceimprovement will not be as significant.
2005-04-04 19:50 neilc
* src/backend/utils/adt/numeric.c: This patch changesint2_avg_accum() and int4_avg_accum() use the nodeAgg
performancehackTom introduced recently. This means we can avoid copying thetransition array for each input tuple if
thesefunctions areinvoked as aggregate transition functions.To test the performance improvement, I created a 1 million
rowtablewith a single int4 column. Without the patch, SELECT avg(col)FROM table took about 4.2 seconds (after the data
wascached); withthe patch, it took about 3.2 seconds. Naturally, the performanceimprovement for a less trivial query
(ora table with wider rows)would be relatively smaller.
2005-03-12 15:25 tgl
* contrib/intagg/int_aggregate.c,contrib/intagg/int_aggregate.sql.in, doc/src/sgml/xaggr.sgml,doc/src/sgml/xfunc.sgml,
src/backend/executor/nodeAgg.c,src/backend/utils/adt/int8.c:Adjust the API for aggregate functioncalls so that a
C-codedfunction can tell whether it is being usedas an aggregate or not. This allows such a function to
avoidre-pallocinga pass-by-reference transition value; normally itwould be unsafe for a function to scribble on an
input,but in theaggregate case it's safe to reuse the old transition value. Makeint8inc() do this. This gets a useful
improvementin the speed ofCOUNT(*), at least on narrow tables (it seems to be swamped by I/Owhen the table rows are
wide). Per a discussion in early Decemberwith Neil Conway. I also fixed int_aggregate.c to check this,thereby
turningit into something approaching a supportabletechnique instead of being a crude hack.
I don't recall how Neil's original patch differed from what got
applied...
regards, tom lane
On Mon, 2005-08-29 at 20:25 -0400, Tom Lane wrote: > Simon Riggs <simon@2ndquadrant.com> writes: > > I notice that Neil's patch regarding reducing the number of memory > > allocations during aggregation operations isn't mentioned. It was > > originally discussed in 8.0beta (2-3?) time. > > > What happened there? > > - patch not committed in the end > > - committed but not mentioned, as a dropped item > > - committed but not mentioned, since part of a larger patch > > Are you speaking of these patches? Yes, those look like the ones I mentioned. Those seem to have a useful performance improvement? At very least, the change in Aggregate function API should be mentioned, no? > 2005-04-06 19:56 neilc > > * src/backend/utils/adt/: float.c, numeric.c: Apply the "nodeAgg" > optimization to more of the builtin transition functions. This > patch optimizes int2_sum(), int4_sum(), float4_accum() and > float8_accum() to avoid needing to copy the transition function's > state for each input tuple of the aggregate. In an extreme case > (e.g. SELECT sum(int2_col) FROM table where table has a single > column), it improves performance by about 20%. For more complex > queries or tables with wider rows, the relative performance > improvement will not be as significant. > > 2005-04-04 19:50 neilc > > * src/backend/utils/adt/numeric.c: This patch changes > int2_avg_accum() and int4_avg_accum() use the nodeAgg performance > hack Tom introduced recently. This means we can avoid copying the > transition array for each input tuple if these functions are > invoked as aggregate transition functions. > > To test the performance improvement, I created a 1 million row > table with a single int4 column. Without the patch, SELECT avg(col) > FROM table took about 4.2 seconds (after the data was cached); with > the patch, it took about 3.2 seconds. Naturally, the performance > improvement for a less trivial query (or a table with wider rows) > would be relatively smaller. > > 2005-03-12 15:25 tgl > > * contrib/intagg/int_aggregate.c, > contrib/intagg/int_aggregate.sql.in, doc/src/sgml/xaggr.sgml, > doc/src/sgml/xfunc.sgml, src/backend/executor/nodeAgg.c, > src/backend/utils/adt/int8.c: Adjust the API for aggregate function > calls so that a C-coded function can tell whether it is being used > as an aggregate or not. This allows such a function to avoid > re-pallocing a pass-by-reference transition value; normally it > would be unsafe for a function to scribble on an input, but in the > aggregate case it's safe to reuse the old transition value. Make > int8inc() do this. This gets a useful improvement in the speed of > COUNT(*), at least on narrow tables (it seems to be swamped by I/O > when the table rows are wide). Per a discussion in early December > with Neil Conway. I also fixed int_aggregate.c to check this, > thereby turning it into something approaching a supportable > technique instead of being a crude hack. I'll search CVS directly next time. Thanks. Best Regards, Simon Riggs
Simon Riggs wrote: > On Mon, 2005-08-29 at 20:25 -0400, Tom Lane wrote: > > Simon Riggs <simon@2ndquadrant.com> writes: > > > I notice that Neil's patch regarding reducing the number of memory > > > allocations during aggregation operations isn't mentioned. It was > > > originally discussed in 8.0beta (2-3?) time. > > > > > What happened there? > > > - patch not committed in the end > > > - committed but not mentioned, as a dropped item > > > - committed but not mentioned, since part of a larger patch > > > > Are you speaking of these patches? > > Yes, those look like the ones I mentioned. > > Those seem to have a useful performance improvement? > > At very least, the change in Aggregate function API should be mentioned, > no? > > 2005-03-12 15:25 tgl > > > > * contrib/intagg/int_aggregate.c, > > contrib/intagg/int_aggregate.sql.in, doc/src/sgml/xaggr.sgml, > > doc/src/sgml/xfunc.sgml, src/backend/executor/nodeAgg.c, > > src/backend/utils/adt/int8.c: Adjust the API for aggregate function > > calls so that a C-coded function can tell whether it is being used > > as an aggregate or not. This allows such a function to avoid > > re-pallocing a pass-by-reference transition value; normally it > > would be unsafe for a function to scribble on an input, but in the > > aggregate case it's safe to reuse the old transition value. Make > > int8inc() do this. This gets a useful improvement in the speed of > > COUNT(*), at least on narrow tables (it seems to be swamped by I/O > > when the table rows are wide). Per a discussion in early December > > with Neil Conway. I also fixed int_aggregate.c to check this, > > thereby turning it into something approaching a supportable > > technique instead of being a crude hack. I don't usually document internal API changes in the release notes. Should I? -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001+ If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania19073
On Tue, Aug 30, 2005 at 10:23:49AM -0400, Bruce Momjian wrote: > > > 2005-03-12 15:25 tgl > > > > > > * contrib/intagg/int_aggregate.c, > > > contrib/intagg/int_aggregate.sql.in, doc/src/sgml/xaggr.sgml, > > > doc/src/sgml/xfunc.sgml, src/backend/executor/nodeAgg.c, > > > src/backend/utils/adt/int8.c: Adjust the API for aggregate function > > > calls so that a C-coded function can tell whether it is being used > > > as an aggregate or not. This allows such a function to avoid > > > re-pallocing a pass-by-reference transition value; normally it > > > would be unsafe for a function to scribble on an input, but in the > > > aggregate case it's safe to reuse the old transition value. Make > > > int8inc() do this. This gets a useful improvement in the speed of > > > COUNT(*), at least on narrow tables (it seems to be swamped by I/O > > > when the table rows are wide). Per a discussion in early December > > > with Neil Conway. I also fixed int_aggregate.c to check this, > > > thereby turning it into something approaching a supportable > > > technique instead of being a crude hack. > > I don't usually document internal API changes in the release notes. > Should I? Doesn't this potentially affect user-defined aggregates? -- Alvaro Herrera <alvherre[]alvh.no-ip.org> Architect, www.EnterpriseDB.com "Lo esencial es invisible para los ojos" (A. de Saint Exúpery)
Alvaro Herrera wrote: > On Tue, Aug 30, 2005 at 10:23:49AM -0400, Bruce Momjian wrote: > > > > > 2005-03-12 15:25 tgl > > > > > > > > * contrib/intagg/int_aggregate.c, > > > > contrib/intagg/int_aggregate.sql.in, doc/src/sgml/xaggr.sgml, > > > > doc/src/sgml/xfunc.sgml, src/backend/executor/nodeAgg.c, > > > > src/backend/utils/adt/int8.c: Adjust the API for aggregate function > > > > calls so that a C-coded function can tell whether it is being used > > > > as an aggregate or not. This allows such a function to avoid > > > > re-pallocing a pass-by-reference transition value; normally it > > > > would be unsafe for a function to scribble on an input, but in the > > > > aggregate case it's safe to reuse the old transition value. Make > > > > int8inc() do this. This gets a useful improvement in the speed of > > > > COUNT(*), at least on narrow tables (it seems to be swamped by I/O > > > > when the table rows are wide). Per a discussion in early December > > > > with Neil Conway. I also fixed int_aggregate.c to check this, > > > > thereby turning it into something approaching a supportable > > > > technique instead of being a crude hack. > > > > I don't usually document internal API changes in the release notes. > > Should I? > > Doesn't this potentially affect user-defined aggregates? I read it as something that _could_ be used by user-defined aggregates, but not something that would require a changes to a user-defined aggregate. -- Bruce Momjian | http://candle.pha.pa.us pgman@candle.pha.pa.us | (610) 359-1001+ If your life is a hard drive, | 13 Roberts Road + Christ can be your backup. | Newtown Square, Pennsylvania19073
Bruce Momjian <pgman@candle.pha.pa.us> writes:
> Alvaro Herrera wrote:
>> On Tue, Aug 30, 2005 at 10:23:49AM -0400, Bruce Momjian wrote:
>>> I don't usually document internal API changes in the release notes.
>>> Should I?
>>
>> Doesn't this potentially affect user-defined aggregates?
> I read it as something that _could_ be used by user-defined aggregates,
> but not something that would require a changes to a user-defined
> aggregate.
I tend to agree with Bruce here. If he documented changes of this size
the release notes would be twice as long and even fewer people would
read them.
regards, tom lane