Thread: Add database to PGXACT / per database vacuuming
Hi, For the logical decoding patch I added support for pegging RecentGlobalXmin (and GetOldestXmin) to a lower value. To avoid causing undue bloat & cpu overhead (hot pruning is friggin expensive) I split RecentGlobalXmin into RecentGlobalXmin and RecentGlobalDataXmin where the latter is the the xmin horizon used for non-shared, non-catalog tables. That removed almost all overhead I could measure. During that I was tinkering with the idea of reusing that split to vacuum/prune user tables in a per db fashion. In a very quick and hacky test that sped up the aggregate performance of concurrent pgbenches in different databases by about 30%. So, somewhat worthwile ;). The problem with that is that GetSnapshotData, which computes RecentGlobalXmin, only looks at the PGXACT structures and not PGPROC which contains the database oid. This is a recently added optimization which made GetSnapshotData() quite a bit faster & scalable which is important given the frequency it's called at. What about moving/copying the database oid from PGPROC to PGXACT? Currently a single PGXACT is 12 bytes which means we a) have several entries in a single cacheline b) have ugly sharing because we will have PGXACTs split over more than one cacheline. Comments? Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 30.08.2013 19:01, Andres Freund wrote: > For the logical decoding patch I added support for pegging > RecentGlobalXmin (and GetOldestXmin) to a lower value. To avoid causing > undue bloat& cpu overhead (hot pruning is friggin expensive) I split > RecentGlobalXmin into RecentGlobalXmin and RecentGlobalDataXmin where > the latter is the the xmin horizon used for non-shared, non-catalog > tables. That removed almost all overhead I could measure. > > During that I was tinkering with the idea of reusing that split to > vacuum/prune user tables in a per db fashion. In a very quick and hacky > test that sped up the aggregate performance of concurrent pgbenches in > different databases by about 30%. So, somewhat worthwile ;). > > The problem with that is that GetSnapshotData, which computes > RecentGlobalXmin, only looks at the PGXACT structures and not PGPROC > which contains the database oid. This is a recently added optimization > which made GetSnapshotData() quite a bit faster& scalable which is > important given the frequency it's called at. Hmm, so you're creating a version of GetSnapshotData() that only takes into account backends in the same backend? > What about moving/copying the database oid from PGPROC to PGXACT? Might be worthwhile. > Currently a single PGXACT is 12 bytes which means we a) have several > entries in a single cacheline b) have ugly sharing because we will have > PGXACTs split over more than one cacheline. I can't get excited about either of these arguments, though. The reason for having separate PGXACT structs is that they are as small as possible, so that you can fit as many of them as possible in as few cache lines as possible. Whether one PGXACT crosses a cache line or not is not important, because when taking a snapshot, you scan through all of them. I don't know how big an impact adding the database oid would have, on the case that the PGPROC/PGXACT split was done in the first place. In the worst case it will make taking a snapshot 1/3 slower under contention. That needs to be tested. One idea is to have a separate PGXACT array for each database? Well, that might be difficult, but something similar, like group all PGXACTs for one database together, and keep a separate lookup array for where the entries for each database begins. - Heikki
On 30.08.2013 21:07, Heikki Linnakangas wrote: > On 30.08.2013 19:01, Andres Freund wrote: >> For the logical decoding patch I added support for pegging >> RecentGlobalXmin (and GetOldestXmin) to a lower value. To avoid causing >> undue bloat& cpu overhead (hot pruning is friggin expensive) I split >> RecentGlobalXmin into RecentGlobalXmin and RecentGlobalDataXmin where >> the latter is the the xmin horizon used for non-shared, non-catalog >> tables. That removed almost all overhead I could measure. >> >> During that I was tinkering with the idea of reusing that split to >> vacuum/prune user tables in a per db fashion. In a very quick and hacky >> test that sped up the aggregate performance of concurrent pgbenches in >> different databases by about 30%. So, somewhat worthwile ;). >> >> The problem with that is that GetSnapshotData, which computes >> RecentGlobalXmin, only looks at the PGXACT structures and not PGPROC >> which contains the database oid. This is a recently added optimization >> which made GetSnapshotData() quite a bit faster& scalable which is >> important given the frequency it's called at. > > Hmm, so you're creating a version of GetSnapshotData() that only takes > into account backends in the same backend? I mean, only takes account backends in the same database? - Heikki
On 2013-08-30 21:07:04 +0300, Heikki Linnakangas wrote: > On 30.08.2013 19:01, Andres Freund wrote: > >For the logical decoding patch I added support for pegging > >RecentGlobalXmin (and GetOldestXmin) to a lower value. To avoid causing > >undue bloat& cpu overhead (hot pruning is friggin expensive) I split > >RecentGlobalXmin into RecentGlobalXmin and RecentGlobalDataXmin where > >the latter is the the xmin horizon used for non-shared, non-catalog > >tables. That removed almost all overhead I could measure. > > > >During that I was tinkering with the idea of reusing that split to > >vacuum/prune user tables in a per db fashion. In a very quick and hacky > >test that sped up the aggregate performance of concurrent pgbenches in > >different databases by about 30%. So, somewhat worthwile ;). > > > >The problem with that is that GetSnapshotData, which computes > >RecentGlobalXmin, only looks at the PGXACT structures and not PGPROC > >which contains the database oid. This is a recently added optimization > >which made GetSnapshotData() quite a bit faster& scalable which is > >important given the frequency it's called at. > > Hmm, so you're creating a version of GetSnapshotData() that only takes into > account backends in the same backend? You can see what I did for logical decoding in http://git.postgresql.org/gitweb/?p=users/andresfreund/postgres.git;a=blobdiff;f=src/backend/storage/ipc/procarray.c;h=11aa1f5a71196a61e31b711e0a044b2a5927a6cc;hp=9bf0989c9206b5e07053587f517d5e9a2322a628;hb=edcf0939072ebe68969560a7d54a26c123b279b4;hpb=ff4fa81665798642719c11c779d0518ef6611373 So, basically I compute the normal RecentGlobalXmin, and then just subtract the "logical xmin" which is computed elsewhere to get the catalog xmin. What I'd done with the prototype of $topic (lost it, but I am going to hack it together again) was just to compute RecentGlobalXmin (for non-catalog, non-shared tables) at the same time with RecentGlobalDataXmin (for everything else) by just not lowering RecentGlobalDataXmin if pgxact->dboid != MyDatabaseId. So, the snapshot itself was the same, but because RecentGlobalDataXmin is independent from the other databases vacuum & pruning can cleanup way more leading to a smaller database and higher database. > >Currently a single PGXACT is 12 bytes which means we a) have several > >entries in a single cacheline b) have ugly sharing because we will have > >PGXACTs split over more than one cacheline. > > I can't get excited about either of these arguments, though. The reason for > having separate PGXACT structs is that they are as small as possible, so > that you can fit as many of them as possible in as few cache lines as > possible. Whether one PGXACT crosses a cache line or not is not important, > because when taking a snapshot, you scan through all of them. The problem with that is that we actually write to PGXACT pretty frequently (at least ->xid, ->xmin, ->nxids, ->delayChkpt). As soon as you factor that in, sharing cachelines between backends can hurt. Even a plain GetSnapshotData() will write to MyPgXact->xmin... > I don't know how big an impact adding the database oid would have, on the > case that the PGPROC/PGXACT split was done in the first place. In the worst > case it will make taking a snapshot 1/3 slower under contention. That needs > to be tested. Yes, definitely. I am basically wondering whether somebody has/sees fundamental probles with it making it pointless to investigate. > One idea is to have a separate PGXACT array for each database? Well, that > might be difficult, but something similar, like group all PGXACTs for one > database together, and keep a separate lookup array for where the entries > for each database begins. Given that we will have to search all PGXACT entries anyway because of shared relations for the forseeable future, I can't see that being really beneficial. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On Fri, Aug 30, 2013 at 2:29 PM, Andres Freund <andres@2ndquadrant.com> wrote: >> I don't know how big an impact adding the database oid would have, on the >> case that the PGPROC/PGXACT split was done in the first place. In the worst >> case it will make taking a snapshot 1/3 slower under contention. That needs >> to be tested. > > Yes, definitely. I am basically wondering whether somebody has/sees > fundamental probles with it making it pointless to investigate. I expect there will be a measurable performance degradation, though I'm willing to be proven wrong. I think the question is whether we get enough bang for the buck out of it to eat that. It seems quite likely that users with many databases will come out ahead, as such systems seem likely to be shared hosting environments where the machine is lightly loaded most of the time anyway, but where cross-database interactions cause headaches. But many users have One Big Database, and AFAICS this is just overhead for them. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company