[mapserver-dev] Trying to improve perfs of MapServer with PostGIS
Patrick Valsecchi
patrick.valsecchi at camptocamp.com
Mon Dec 14 22:26:22 PST 2015
Paul,
The resultFormat parameter was already there back in 2003 with Postgres 7.4
when PQexecWithParams was introduced [1]. I hope the current minimum
supported version of Postgres does not predate that.
Another question I have before I do a PR: what is the usefulness of all
those #if TRANSFER_ENCODING? It cannot even be modified through cmake. I'm
very tempted to remove all that.
Thanks for your feedback.
[1]
http://www.postgresql.org/docs/7.4/static/libpq-exec.html#LIBPQ-EXEC-MAIN
On Mon, Dec 14, 2015 at 8:15 PM, Paul Ramsey <pramsey at cleverelephant.ca>
wrote:
> Please bear in mind that the original implementation using binary
> cursors was deliberately *removed* (by me) because it added so much
> complexity to the code base. (In terms of having to maintain a correct
> transaction state throughout the query lifecycle.)
>
> However, your patch is very small, I see. Perhaps pgsql support of
> binary has improved a lot since the days of binary cursors. (Which
> could only be used in transaction, at the time.)
>
> Adding TWKB to mapnik has substantially improved network performance [1]
> for us.
>
> http://blog.cartodb.com/smaller-faster/
>
> Removing precision in map rendering works like a charm, I'd highly
> recommend it. However, Mapserver's dual nature as both a feature and
> map server would make a change to support that more tricky than the
> Mapnik change.
>
> The only issue I'd have to this being a PR is some careful research as
> to when this RESULTSET_TYPE flag was added to PQexecParams I suppose
> we could have missed it all those years ago, and thus used binary
> cursors, but it could also be an add-on as of a certain date. Knowing
> when it was added would be good. (Probably long enough ago that we
> don't even need to #ifdef a version test on it, but I would like
> confirmation.)
>
> ATB,
>
> P
>
> On Mon, Dec 14, 2015 at 7:10 AM, Patrick Valsecchi
> <patrick.valsecchi at camptocamp.com> wrote:
> > Daniel: I do agree with your analysis.
> >
> > Even: That is a bit out of scope for this patch and I'm no fan of
> reducing
> > the precision, even for WMS.
> >
> > On Fri, Dec 11, 2015 at 5:06 PM, Even Rouault <
> even.rouault at spatialys.com>
> > wrote:
> >>
> >> Le vendredi 11 décembre 2015 13:38:56, Daniel Morissette a écrit :
> >> > Hi Patrick,
> >> >
> >> > We have run into similar behaviors with a quite similar setup on
> Amazon
> >> > a little while ago, in our case we had a dedicated DB server with SSD
> >> > (like yours) and multiple rendering nodes connecting to it, and adding
> >> > more nodes or more powerful ones still resulted in a constant
> throughput
> >> > and the additional rendering nodes CPU running mostly idle. After much
> >> > hair pulling we finally found out that we were hitting the 1 Gbps
> limit
> >> > of the network connection between the DB and rendering node.
> >> >
> >> > I suspect you are running into the same thing here: if you could
> verify
> >> > that the size of the original vs binary response is about the double,
> >> > then that would confirm this possibility. We used PGBouncer to look at
> >> > the volume of data transferred by the DB server, it produces output
> like
> >> > this:
> >> >
> >> > 2014-06-06 18:49:44.597 27891 LOG Stats: 852 req/s, in 435114 b/s,
> out
> >> > 83424883 b/s,query 7133 us
> >> > 2014-06-06 18:50:44.597 27891 LOG Stats: 818 req/s, in 418046 b/s, out
> >> > 79177139 b/s,query 5761 us
> >> > 2014-06-06 18:52:44.598 27891 LOG Stats: 898 req/s, in 458041 b/s, out
> >> > 88441708 b/s,query 7949 us
> >> >
> >> >
> >> > Then to really confirm it, look at the number of hits per second in
> each
> >> > test case (instead of the processing time per request). You will
> likely
> >> > notice that it is constant for a given zoom level once you hit the
> >> > network limit no matter how many parallel requests you have. This is
> >> > because the number of hits per seconds that can be served are limited
> by
> >> > the volume of data per second that the DB server can deliver to
> >> > MapServer (around ~800Mbits/sec for a 1 Gbps connection).
> >> >
> >> > In our case, if I remember correctly, we used to hit a limit of about
> 40
> >> > maps/second, where each map was averaging 2MB of data... so we were
> >> > capping at 4 hits x 2MB ~= 80MBytes per second from the DB (or
> >> > ~800Mbits/sec) ... which is as I wrote above in line with a 1Gbps
> limit.
> >>
> >> If bandwidth is the bottleneck, an even more efficient solution than the
> >> binary
> >> WKB representation would be to use http://postgis.net/docs/manual-
> >> dev/ST_AsTWKB.html . This would require a TWKB decoder to be developped
> in
> >> MapServer and be likely only appropriate for WMS like scenarios where
> you
> >> can
> >> afford loosing some precision to coordinates values.
> >>
> >> Even
> >>
> >> >
> >> > Good luck!
> >> >
> >> > Daniel
> >> >
> >> > On 2015-12-11 4:13 AM, Patrick Valsecchi wrote:
> >> > > Hi,
> >> > >
> >> > > I was a bit disapointed by the perfs of reading the features from
> >> > > PostGIS compared to reading them from a local Shapefile. So I've
> >> > > pocked
> >> > > my nose around the code and I saw that geometries were fetched from
> >> > > the
> >> > > DB in EWKB encoded in hex.
> >> > >
> >> > > So I went ahead and changed the code to use libpq's binary mode when
> >> > > we
> >> > > are fetching features. To avoid conversion nightmares when fetching
> >> > > attributes, I'm tricking Postgres into sending them in their textual
> >> > > form. That way, the behavior of MapServer should not change
> regarding
> >> > > the other columns.
> >> > >
> >> > > The patch can be found in github [1].
> >> > >
> >> > > General testing algo:
> >> > > for each zoom level
> >> > > start at the same time N threads that does
> >> > > for each iteration
> >> > > compute random BBOX
> >> > > get the 1000x1000 image (throw the content out, but read it)
> >> > > Wait for all the threads to stop
> >> > > Collect the stats from the threads (throwing out the slowest result
> >> > > and
> >> > > the fastest result).
> >> > >
> >> > > The DB is on a db.m4.large from Amazon with SSD disk. Mapserver is
> on
> >> > > apache with fcgid (up to 20 processes) and runs on a m4.xlarge. The
> >> > > measures are taken from another Amazon machine to avoid hitting the
> >> > > bandwidth limit too fast.
> >> > >
> >> > > Times are for one query from one thread and are given in
> milliseconds
> >> > > along with their standard deviation. Each time, I did a full run
> >> > > before
> >> > > taking measures.
> >> > >
> >> > > Data used is the swiss villages (total 2479 features) with polygon
> >> > > contours in a 1.7MB shapefile file [2], imported as is in PostGIS.
> >> > >
> >> > > nbThreads=1 nbIterations=20
> >> > > zoom level 1.00 4.00 16.00 32.00
> >> > > original 964± 14 222± 126 66± 18 68± 19
> >> > > binary 807± 13 194± 111 87± 28 79± 24
> >> > > shapefile 554± 94 187± 107 72± 26 56± 3
> >> > >
> >> > >
> >> > > nbThreads=5 nbIterations=20
> >> > > zoom level 1.00 4.00 16.00 32.00
> >> > > original 3686± 946 403± 264 84± 37 70± 22
> >> > > binary 1710± 486 340± 242 105± 59 89± 32
> >> > > shapefile 519± 225 278± 166 91± 58 80± 34
> >> > >
> >> > > nbThreads=10 nbIterations=20
> >> > > zoom level 1.00 4.00 16.00 32.00
> >> > > original 7287±1936 800± 575 119± 79 110± 81
> >> > > binary 3737± 647 471± 294 123± 70 110± 54
> >> > > shapefile 884± 241 412± 269 111± 119 98± 57
> >> > >
> >> > > nbThreads=20 nbIterations=20
> >> > > zoom level 1.00 4.00 16.00 32.00
> >> > > original 14969±2507 1643±1221 239± 231 166± 103
> >> > > binary 7649± 730 857± 576 210± 121 181± 77
> >> > > shapefile 1455± 438 483± 326 143± 97 126± 75
> >> > >
> >> > > What is what:
> >> > > * original: Mapserver 7.0 (git 157fa47)
> >> > > * binary: Same as above with a small patch to configure libpg to
> use
> >> > >
> >> > > binary transfer.
> >> > >
> >> > > * shapefile: Same as original, but using a shapefile on the local
> >> > > disk
> >> > >
> >> > > (just here for comparison).
> >> > >
> >> > > We can see that when the machine gets parallel queries, we quickly
> get
> >> > > a
> >> > > factor 2 in perfs when there are a lot of features (low zoom level).
> >> > > There is no measurable negative impact at higher zoom levels and
> lower
> >> > > loads.
> >> > >
> >> > > Now, what do you guys think? Do you see in risk? Or should do a pull
> >> > > request?
> >> > >
> >> > > Thanks.
> >> > >
> >> > >
> >> > > [1]
> >> > >
> >> > >
> https://github.com/pvalsecc/mapserver/commit/45bd3d5795c9108618c37cc8c747
> >> > > 2809cff54d16 [2]
> >> > >
> >> > >
> http://www.swisstopo.admin.ch/internet/swisstopo/en/home/products/landsca
> >> > > pe/swissBOUNDARIES3D.html
> >> > >
> >> > >
> >> > >
> >> > > _______________________________________________
> >> > > mapserver-dev mailing list
> >> > > mapserver-dev at lists.osgeo.org
> >> > > http://lists.osgeo.org/mailman/listinfo/mapserver-dev
> >>
> >> --
> >> Spatialys - Geospatial professional services
> >> http://www.spatialys.com
> >> _______________________________________________
> >> mapserver-dev mailing list
> >> mapserver-dev at lists.osgeo.org
> >> http://lists.osgeo.org/mailman/listinfo/mapserver-dev
> >
> >
> >
> > _______________________________________________
> > mapserver-dev mailing list
> > mapserver-dev at lists.osgeo.org
> > http://lists.osgeo.org/mailman/listinfo/mapserver-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/mapserver-dev/attachments/20151215/2dc8ae56/attachment.html>
More information about the mapserver-dev
mailing list