Thread: Parallel Aggregate costs don't consider combine/serial/deserial funcs
Hi, I realised a few days ago that the parallel aggregate code does not cost for the combine, serialisation and deserialisation functions at all. I've attached a patch which fixes this. One small point which I was a little unsure of in the attached is, should the "if (aggref->aggdirectargs)" part of count_agg_clauses_walker() be within the "if (!context->combineStates)". I simply couldn't decide. We currently have no aggregates which this affects anyway, per; select * from pg_aggregate where aggcombinefn <> 0 and aggkind <> 'n'; so for now I've left it outwith. Another thing I thought of is that it's not too nice that I have to pass 3 bools to count_agg_clauses() in order to tell it what to do. I was tempted to invent some bitmask flags for this, then modify create_agg_path() to use the same flags, but I thought I'd better not cause too much churn with this patch. -- David Rowley http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Attachment
On Mon, Apr 11, 2016 at 12:47:29AM +1200, David Rowley wrote: > Hi, > > I realised a few days ago that the parallel aggregate code does not > cost for the combine, serialisation and deserialisation functions at > all. > > I've attached a patch which fixes this. > > One small point which I was a little unsure of in the attached is, > should the "if (aggref->aggdirectargs)" part of > count_agg_clauses_walker() be within the "if > (!context->combineStates)". I simply couldn't decide. We currently > have no aggregates which this affects anyway, per; select * from > pg_aggregate where aggcombinefn <> 0 and aggkind <> 'n'; so for now > I've left it outwith. > > Another thing I thought of is that it's not too nice that I have to > pass 3 bools to count_agg_clauses() in order to tell it what to do. I > was tempted to invent some bitmask flags for this, then modify > create_agg_path() to use the same flags, but I thought I'd better not > cause too much churn with this patch. [This is a generic notification.] The above-described topic is currently a PostgreSQL 9.6 open item. Robert, since you committed the patch believed to have created it, you own this open item. If that responsibility lies elsewhere, please let us know whose responsibility it is to fix this. Since new open items may be discovered at any time and I want to plan to have them all fixed well in advance of the ship date, I will appreciate your efforts toward speedy resolution. Please present, within 72 hours, a plan to fix the defect within seven days of this message. Thanks.
On Sun, Apr 10, 2016 at 8:47 AM, David Rowley <david.rowley@2ndquadrant.com> wrote: > I realised a few days ago that the parallel aggregate code does not > cost for the combine, serialisation and deserialisation functions at > all. Oops. > I've attached a patch which fixes this. I've committed this patch. I wonder if it's going to produce compiler warnings for some people, complaining about possible use of an uninitialized variable. That would kind of suck. I don't much mind having to insert a dummy assignment to shut the compiler up; a smarter compiler will just throw it out anyway. I'm less enthused about a dummy MemSet. The compiler is less likely to be able to get rid of that, and it's more expensive if it doesn't. But let's see what happens. > One small point which I was a little unsure of in the attached is, > should the "if (aggref->aggdirectargs)" part of > count_agg_clauses_walker() be within the "if > (!context->combineStates)". I simply couldn't decide. We currently > have no aggregates which this affects anyway, per; select * from > pg_aggregate where aggcombinefn <> 0 and aggkind <> 'n'; so for now > I've left it outwith. The direct arguments would be evaluated in the worker, but not in the leader, right? Or am I confused? > Another thing I thought of is that it's not too nice that I have to > pass 3 bools to count_agg_clauses() in order to tell it what to do. I > was tempted to invent some bitmask flags for this, then modify > create_agg_path() to use the same flags, but I thought I'd better not > cause too much churn with this patch. I'm kinda tempted to say this should be using an enum. I note that serialStates has a subtly different meaning here than in some other places where you have used the same term. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On 13 April 2016 at 08:52, Robert Haas <robertmhaas@gmail.com> wrote: > On Sun, Apr 10, 2016 at 8:47 AM, David Rowley > <david.rowley@2ndquadrant.com> wrote: >> I realised a few days ago that the parallel aggregate code does not >> cost for the combine, serialisation and deserialisation functions at >> all. > > Oops. > >> I've attached a patch which fixes this. > > I've committed this patch. I wonder if it's going to produce compiler > warnings for some people, complaining about possible use of an > uninitialized variable. That would kind of suck. I don't much mind > having to insert a dummy assignment to shut the compiler up; a smarter > compiler will just throw it out anyway. I'm less enthused about a > dummy MemSet. The compiler is less likely to be able to get rid of > that, and it's more expensive if it doesn't. But let's see what > happens. Thanks for committing. I wondered that too, so checked a couple of compilers and got no warnings, but the buildfarm should let us know. The other option would be to palloc() them, and have them set to NULL initially... that's not very nice either... Another option would be to protect the final parallel path generation with if (grouped_rel->partial_pathlist && grouped_rel->consider_parallel). I'd imagine any compiler smart enough to work out that uninitialised is not possible would also be able to remove the check for grouped_rel->consider_parallel, but *shrug*, I don't often look at the assembly that compilers generate, so I might be giving them too much credit. >> One small point which I was a little unsure of in the attached is, >> should the "if (aggref->aggdirectargs)" part of >> count_agg_clauses_walker() be within the "if >> (!context->combineStates)". I simply couldn't decide. We currently >> have no aggregates which this affects anyway, per; select * from >> pg_aggregate where aggcombinefn <> 0 and aggkind <> 'n'; so for now >> I've left it outwith. > > The direct arguments would be evaluated in the worker, but not in the > leader, right? Or am I confused? That seems right, but I just can't think of how its possible to parallelise these aggregates anyway. >> Another thing I thought of is that it's not too nice that I have to >> pass 3 bools to count_agg_clauses() in order to tell it what to do. I >> was tempted to invent some bitmask flags for this, then modify >> create_agg_path() to use the same flags, but I thought I'd better not >> cause too much churn with this patch. > > I'm kinda tempted to say this should be using an enum. I note that > serialStates has a subtly different meaning here than in some other > places where you have used the same term. hmm, I'm not sure how it's subtly different. Do you mean the preference towards costing the finalfn when finalizeAggs is true, and ignoring the serialfn in this case? nodeAgg.c should do the same, although it'll deserialize in such a case. We can never finalize and serialize in the same node. -- David Rowley http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On Tue, Apr 12, 2016 at 5:38 PM, David Rowley <david.rowley@2ndquadrant.com> wrote: >>> One small point which I was a little unsure of in the attached is, >>> should the "if (aggref->aggdirectargs)" part of >>> count_agg_clauses_walker() be within the "if >>> (!context->combineStates)". I simply couldn't decide. We currently >>> have no aggregates which this affects anyway, per; select * from >>> pg_aggregate where aggcombinefn <> 0 and aggkind <> 'n'; so for now >>> I've left it outwith. >> >> The direct arguments would be evaluated in the worker, but not in the >> leader, right? Or am I confused? > > That seems right, but I just can't think of how its possible to > parallelise these aggregates anyway. Well, if you could ensure that each worker would see a whole group, you could do it, I think. But it's probably fine to just leave this for now. It's not like it can't be changed if somebody figures out some cool thing to do in this area. >>> Another thing I thought of is that it's not too nice that I have to >>> pass 3 bools to count_agg_clauses() in order to tell it what to do. I >>> was tempted to invent some bitmask flags for this, then modify >>> create_agg_path() to use the same flags, but I thought I'd better not >>> cause too much churn with this patch. >> >> I'm kinda tempted to say this should be using an enum. I note that >> serialStates has a subtly different meaning here than in some other >> places where you have used the same term. > > hmm, I'm not sure how it's subtly different. Do you mean the > preference towards costing the finalfn when finalizeAggs is true, and > ignoring the serialfn in this case? nodeAgg.c should do the same, > although it'll deserialize in such a case. We can never finalize and > serialize in the same node. I mean that, IIUC, in some other places where you use serialStates, true means that (de)serialization is known to be needed. Here, however, it only means it might be needed, contingent on whether the serial/deserial functions are actually present. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company