[GRASSLIST:9360] Re: [GRASS5] Re: Re: v.in.ascii problems

[Roger Bivand]

On Wed, 7 Dec 2005, Roger Bivand wrote:

> On Thu, 8 Dec 2005, Hamish wrote:
> 
> > > There is an on-going discussion about this on the GRASS development
> > > list. >>From a simple test I ran last night, v.in.ascii -b (the -b
> > > flag is new in GRASS 6.1 CVS) does not build topology, and this
> > > removes one of the two humps in memory consumption. The other hump (>
> > > 200MB for a 1M point file with a single attribute) was associated with
> > > writing the dbf file (the file is 60MB), and is where things stick
> > > now. In addition, the -b flag leaves the vector data set at level 1
> > > topology (absent), and almost all vector commands need level 2. 
> > > 
> > > I do now know whether the use of a different database driver than the
> > > default would help. The dbf writing stage preceeds the topology
> > > building, so the two memory-intensive humps are separate, with
> > > topology being a little larger. Reading 1M points on a 1.5GHz P4 with
> > > topology took about 7 minutes, without about half that time.
> > 
> > 
> > Use the -z and -t flags to avoid making the table. (and the z= option)
> > If the input is just x,y,z data there is no need for a table.
> 
> For me in an x-y location my data with 1M points and -zbt now read in just
> under a minute; lidaratm2.txt in effectively the same time (64 rather than
> 57 seconds, z here is double not int) and v.in.ascii stays at a
> respectable 3.3MB size. d.vect works, but as you say prints a warning.

Emboldened by the 1M points case, I've tried 13.6M points, with:

GRASS > date; v.in.ascii -zbt input=gptsL.txt output=gptsLz x=1 y=2 z=3 ; date
Wed Dec  7 15:44:23 CET 2005
Maximum input row length: 37
Maximum number of columns: 3
Minimum number of columns: 3
Wed Dec  7 15:57:20 CET 2005

so scaling quite decently at about 1M points a minute. Once again the 
v.in.ascii process stayed at just over 3MB the whole time. So putting the 
single attribute of interest in the z column, not doing topology and not 
creating a database table seems to work - the coords file is now about 
480MB, so on 32-bit machines, its size may be a further limiting factor.

> 
> > 
> > At minimum, we need v.info, v.surf.rst, v.univar, v.out.ascii (points) 
> > and some sort of subsampling module (ie s.cellstats port) working with 
> > this data. d.vect works already (with a warning). maybe v.surf.idw too.
> > 
> > Probably not many more modules though? -- I think if Radim doesn't want 
> > this to be common-place use of the vector model then it probably shouldn't 
> > be. He knows it better than anyone.. So for now massive point datasets 
> > need to be treated as a special case to the vector model & only a 
> > work-around solution.
> > 
> > 
> > e.g. with the sample LIDAR data (GRASS downloads page)
> > 
> > G61> v.in.ascii -zbt in=lidaratm2.txt out=lidaratm2 x=1 y=2 z=3 fs=,
> > 
> > The first 250k points take about 20 seconds to load.
> > 
> > 
> > If I use the full million it gets stuck on the scanning step:
> > 
> > D3/3: row 374430 : 28 chars
> > 
> > Interesting, that line is the second value with elevation > 100.
> > 
> > changing the first z value to 500.054 it segfaults pretty quick:
> > 
> > D5/5: Vect_hist_write()
> > D4/5: G_getl2: ->-75.622346,35.949693,500.054<-
> > D3/5: row 1 : 28 chars
> > D4/5: token: -75.622346
> > D4/5: is_latlong north: -75.622346
> > D4/5: row 1 col 0: '-75.622346' is_int = 0 is_double = 1
> > D4/5: token: 35.949693
> > D4/5: is_latlong north: 35.949693
> > D4/5: row 1 col 1: '35.949693' is_int = 0 is_double = 1
> > D4/5: row 1 col 2: '500.054' is_int = 0 is_double = 1
> > D4/5: G_getl2: ->-75.629469,35.949693,11.962<-
> > D3/5: row 2 : 27 chars
> > D4/5: token: -75.629469
> > D4/5: is_latlong north: -75.629469
> > D4/5: row 2 col 0: 'H629469' is_int = 0 is_double = 0
> > Segmentation fault
> > 
> > where is 'H629469' coming from?
> 
> I was also seeing seg-faults with my data in a long-lat location, so 
> switched to x-y (current CVS 6.1).
> 
> Roger
> 
> > 
> > v.in.ascii/points.c
> >  tmp_token is getting corrupted, cascades from there
> > 
> > int points_analyse (){
> > ...
> >     char **tokens;
> > ...
> >     tmp_token=(char *) G_malloc(256);
> > ...
> >     while (1) {
> > ...
> >         tokens = G_tokenize (buf, fs);
> > ...
> >         for ( i = 0; i < ntokens; i++ ) {
> > ...
> > [*]                 sprintf(tmp_token, "%f", northing);
> > ...
> > 		    /* replace current DMS token by decimal degree */
> >                     tokens[i]=tmp_token;
> > 
> > BOOM. pointer abuse. (bug is new lat/lon scanning code, only in 6.1CVS)
> > 
> > [*] and if northing column is longer than 256 without hitting the fs, 
> >    buffer overflow??  add ,int maxlength, parameter to G_tokenize()?
> >    or can %f never be more than 256 bytes long?
> >     same %f effectively cutting down precision of lat/lon coords to 6 
> >    spots after the decimal place? (be that pretty small on the ground)
> > 
> > 
> > improvements come one bug at a time...
> > 
> > Hamish
> > 
> 
> 

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Helleveien 30, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: Roger.Bivand at nhh.no