Re: [HACKERS] Optimizing aggregates

[Andres Freund]

On 2016-08-31 19:07:00 +0300, Heikki Linnakangas wrote:
> On 08/31/2016 06:51 PM, Andres Freund wrote:
> > I've first combined the projection for all the aggregates, ordered set,
> > or not, into one projetion. That got rid of a fair amount of overhead
> > when you have multiple aggregates.  I attached an, probably out of date,
> > WIP version of that patch.
> 
> A-ha, I also considered doing just that! I also considered a variant where
> we call ExecProject once for all non-ordered aggregates, and a separate
> ExecProject() for each ordered one. But just switching back to straight
> ExecEvalExprs seemed easier.

The issue is that might, I think, end up iteratively deforming the
underlying tuple. The projection machinery takes care of that, if we do
it in one go.

Greetings,

Andres Freund


-- 
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers

Re: [HACKERS] Optimizing aggregates

[Heikki Linnakangas]

On 08/31/2016 06:51 PM, Andres Freund wrote:

On 2016-08-31 17:47:18 +0300, Heikki Linnakangas wrote:

We actually used to call ExecEvalExpr() directly for each argument, but that
was changed by the patch that added support for ordered set aggregates. It
looks like that was a bad idea, from a performance point of view.


I complained about that as well
http://archives.postgresql.org/message-id/20160519175727.ymv2y5tye4qgcmqx%40alap3.anarazel.de


Ah, missed that!


I propose that we go back to calling ExecEvalExpr() directly, for
non-ordered aggregates, per the attached patch. That makes that example
query about 10% faster on my laptop, which is in line with the fact that
ExecProject() accounted for about 13% of the CPU time.


My approach is a bit different.

I've first combined the projection for all the aggregates, ordered set,
or not, into one projetion. That got rid of a fair amount of overhead
when you have multiple aggregates.  I attached an, probably out of date,
WIP version of that patch.


A-ha, I also considered doing just that! I also considered a variant 
where we call ExecProject once for all non-ordered aggregates, and a 
separate ExecProject() for each ordered one. But just switching back to 
straight ExecEvalExprs seemed easier.



Secondly, I'm working on overhauling expression evaluation to be
faster. Even without the ExecProject overhead, the computations very
quickly become the bottleneck. During that I pretty much merged
ExecProject and ExecEvalExpr into one - they're really not that
different, and the distinction serves no purpose, except to increase the
number of function calls. The reason I'm working on getting rid of
targetlist SRFs is precisely that. A proof of concept of that is
attached to
http://archives.postgresql.org/message-id/20160714011850.bd5zhu35szle3n3c%40alap3.anarazel.de


Cool, yes, all that should help.

- Heikki



--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers

Re: [HACKERS] Optimizing aggregates

[Andres Freund]

Hi,

On 2016-08-31 17:47:18 +0300, Heikki Linnakangas wrote:
> #     ..  .
> 
> #
> 25.70% 0.00%  postmaster  [unknown]  [k] 
> 14.23%13.75%  postmaster  postgres   [.] ExecProject

> ExecProject stands out. I find that pretty surprising.
> 
> We're using ExecProject to extract the arguments from the input tuples, to
> pass to the aggregate transition functions. It looks like that's a pretty
> expensive way of doing it, for a typical aggregate that takes only one
> argument.
> 
> We actually used to call ExecEvalExpr() directly for each argument, but that
> was changed by the patch that added support for ordered set aggregates. It
> looks like that was a bad idea, from a performance point of view.

I complained about that as well
http://archives.postgresql.org/message-id/20160519175727.ymv2y5tye4qgcmqx%40alap3.anarazel.de


> I propose that we go back to calling ExecEvalExpr() directly, for
> non-ordered aggregates, per the attached patch. That makes that example
> query about 10% faster on my laptop, which is in line with the fact that
> ExecProject() accounted for about 13% of the CPU time.

My approach is a bit different.

I've first combined the projection for all the aggregates, ordered set,
or not, into one projetion. That got rid of a fair amount of overhead
when you have multiple aggregates.  I attached an, probably out of date,
WIP version of that patch.

Secondly, I'm working on overhauling expression evaluation to be
faster. Even without the ExecProject overhead, the computations very
quickly become the bottleneck. During that I pretty much merged
ExecProject and ExecEvalExpr into one - they're really not that
different, and the distinction serves no purpose, except to increase the
number of function calls. The reason I'm working on getting rid of
targetlist SRFs is precisely that. A proof of concept of that is
attached to
http://archives.postgresql.org/message-id/20160714011850.bd5zhu35szle3n3c%40alap3.anarazel.de

Greetings,

Andres Freund
>From 384845bea72d28952d88e58e55f81aaa5addd930 Mon Sep 17 00:00:00 2001
From: Andres Freund 
Date: Tue, 12 Jul 2016 01:01:28 -0700
Subject: [PATCH] WIP: Only perform one projection in aggregation.

---
 src/backend/executor/nodeAgg.c | 112 -
 1 file changed, 88 insertions(+), 24 deletions(-)

diff --git a/src/backend/executor/nodeAgg.c b/src/backend/executor/nodeAgg.c
index f655aec..4499d5f 100644
--- a/src/backend/executor/nodeAgg.c
+++ b/src/backend/executor/nodeAgg.c
@@ -210,6 +210,9 @@ typedef struct AggStatePerTransData
 	 */
 	int			numInputs;
 
+	/* offset of input columns in Aggstate->evalslot */
+	int			inputoff;
+
 	/*
 	 * Number of aggregated input columns to pass to the transfn.  This
 	 * includes the ORDER BY columns for ordered-set aggs, but not for plain
@@ -836,14 +839,20 @@ advance_aggregates(AggState *aggstate, AggStatePerGroup pergroup)
 	int			setno = 0;
 	int			numGroupingSets = Max(aggstate->phase->numsets, 1);
 	int			numTrans = aggstate->numtrans;
+	TupleTableSlot *slot = aggstate->evalslot;
+	AggStatePerTrans pertrans;
 
-	for (transno = 0; transno < numTrans; transno++)
+	/* compute input for all aggregates */
+	if (aggstate->evalproj)
+		ExecProjectIntoSlot(aggstate->evalproj, aggstate->evalslot);
+
+	for (transno = 0, pertrans = aggstate->pertrans; transno < numTrans;
+		 transno++, pertrans++)
 	{
-		AggStatePerTrans pertrans = >pertrans[transno];
 		ExprState  *filter = pertrans->aggfilter;
 		int			numTransInputs = pertrans->numTransInputs;
 		int			i;
-		TupleTableSlot *slot;
+		int			inputoff = pertrans->inputoff;
 
 		/* Skip anything FILTERed out */
 		if (filter)
@@ -857,13 +866,10 @@ advance_aggregates(AggState *aggstate, AggStatePerGroup pergroup)
 continue;
 		}
 
-		/* Evaluate the current input expressions for this aggregate */
-		slot = ExecProject(pertrans->evalproj, NULL);
-
 		if (pertrans->numSortCols > 0)
 		{
 			/* DISTINCT and/or ORDER BY case */
-			Assert(slot->tts_nvalid == pertrans->numInputs);
+			Assert(slot->tts_nvalid >= pertrans->numInputs);
 
 			/*
 			 * If the transfn is strict, we want to check for nullity before
@@ -876,7 +882,7 @@ advance_aggregates(AggState *aggstate, AggStatePerGroup pergroup)
 			{
 for (i = 0; i < numTransInputs; i++)
 {
-	if (slot->tts_isnull[i])
+	if (slot->tts_isnull[i + inputoff])
 		break;
 }
 if (i < numTransInputs)
@@ -888,10 +894,22 @@ advance_aggregates(AggState *aggstate, AggStatePerGroup pergroup)
 /* OK, put the tuple into the tuplesort object */
 if (pertrans->numInputs == 1)
 	tuplesort_putdatum(pertrans->sortstates[setno],
-	   slot->tts_values[0],
-	   slot->tts_isnull[0]);
+	   slot->tts_values[inputoff],
+	   slot->tts_isnull[inputoff]);
 else
-