[GRASS-dev] Backport r.li memory allocation fixes?

Tue Oct 14 18:35:17 PDT 2014

Hi Markus and others,

I would do a backport or r61812 but I'm not sure if it works correctly. It
seems to work and I got the same results as with the old version but I
remember that you (Markus) had some doubts at one point. Now I actually see
that you said you will backport it, so if you have your plan disregard this
message.

I also wanted to use the things bellow to write a test but I don't know how.

Vaclav

http://trac.osgeo.org/grass/changeset/61812

---------- Forwarded message ----------
From: Sandra MacFadyen <sandramf at live.co.za>
Date: Mon, Sep 29, 2014 at 9:24 PM
Subject: RE: [GRASS-user] Solved: execGRASS("r.diversity") Done. No rasters
created (Large rasters)
To: Vaclav Petras <wenzeslaus at gmail.com>
Cc: Markus Neteler <neteler at osgeo.org>, GRASS user list <
grass-user at lists.osgeo.org>, Duccio Rocchini <ducciorocchini at gmail.com>,
Luca Delucchi <lucadeluge at gmail.com>

Dear Vaclav,

I did a quick manual test to see if the results were correct for a few 5x5
windows and yes, everything appears 100%:

# Example data from one 5x5 window

test1 = c(572,617,661,638,616,

          662,617,593,638,616,

          662,639,572,661,616,

          662,662,639,662,616,

          594,572,594,639,593)

# Calculate info needed for diversity indices

freq1 = as.data.frame(table(test1)) # Calculate frequency table

S = length(unique(test1)) # Total number of species (S)

N = length(test1) # Total number of individuals (N)

ln.S = log(S) #Natural log of species (ln S)

ln.N = log(N) # Natural log of individuals (ln N)

Pi = freq1$Freq/N # Proportion of individuals that belong to each species

Pi2 = (Pi ^ 2)

ln.pi = log(Pi)

pi.ln.pi = Pi * ln.pi

pi.ln.pi2 = (ln.pi^2)*Pi

# Calculate different diversity indices

(M = (S-1)/ln.N) # Margalef's index (M)

(i.Di = 1/sum(Pi2)) # Simpson's index [Inverse] (i.Di)

(i.Dc = 1 - (sum(Pi2))) # Simpson's index [Complement] (i.Dc) * This one
calculated with r.diversity

(H = sum(pi.ln.pi)*-1) # Shannon-Wiener index (H')

(J = H/ln.S) # Pielou's index (J)

Thanks again

Cheers

Sandra

From: Vaclav Petras [mailto:wenzeslaus at gmail.com]
Sent: 18 September 2014 02:46 AM
To: Sandra MacFadyen
Cc: Markus Neteler; GRASS user list; Duccio Rocchini; Luca Delucchi
Subject: Re: [GRASS-user] Solved: execGRASS("r.diversity") Done. No rasters
created (Large rasters)

On Wed, Sep 17, 2014 at 7:40 PM, Sandra MacFadyen <sandramf at live.co.za>
wrote:

Dear Markus,

Excellent! Thanks to everyone for their help.
I tested r.diversity in the new revision (61840) on a raster with 12933032
cells and it took under 3 minutes to complete r.li.simpson, r.li.shannon,
r.li.pielou and r.li.renyi outputs.
Amazing :)

> execGRASS("r.diversity",flags="overwrite", parameters=list(input="flow",
prefix="flow_div", alpha=0.5, size=3))
> r.li.simpson complete. Raster map <flow_div_simpson_size_3.0> created.
> r.li.shannon complete. Raster map <flow_div_shannon_size_3.0> created.
> r.li.pielou complete. Raster map <flow_div_pielou_size_3.0> created.
> r.li.renyi complete. Raster map <flow_div_renyi_size_3.0_alpha_0.5>
> created.
> Done.

Thanks again and keep well.

Good to hear that. Can you say if the results are correct? I computed
difference of one map before and after and it was OK. But it would be great
to have more tests. Let's start with: does the result make sense?

Vaclav

Cheers
Sandra

-----Original Message-----
From: neteler.osgeo at gmail.com [mailto:neteler.osgeo at gmail.com] On Behalf Of
Markus Neteler
Sent: 06 September 2014 05:15 AM
To: Sandra MacFadyen
Cc: GRASS user list; Duccio Rocchini; Luca Delucchi
Subject: Re: [GRASS-user] execGRASS("r.diversity") Done. No rasters created
(Large rasters)

Dear Sandra,

On Fri, Jun 20, 2014 at 7:49 AM, Markus Neteler <neteler at osgeo.org> wrote:
> On Sat, Jun 14, 2014 at 9:21 AM, Sandra MacFadyen <sandramf at live.co.za>
wrote:
>> I am using r.diversity (GRASS GIS 7.0.0svn build 60785 win32) through
>> R (R version 3.0.2 win32) on Windows 7 64bit.
...
>> However, when running the same code on a larger image (cells=6746328)
>> from my own location, although it reports Done, no rasters are
>> created. If I subset the image
>> (cells=1632830) and run it again its works (see # sub2Kruger # code and
results below).
>> So I'm guessing it is a memory issue?
...
> ... it consumes a lot of memory... will check on a bigger machine,
> perhaps a memory leak.

The assumption turned out to be right and I think we got it today!

Vaclav Petras checked it and discovered an "unfortunate" memory allocation
which he fixed in r.li.* in revision
http://trac.osgeo.org/grass/changeset/61812 ("r.li: fix memory handling
(memory leak in avl_to_array function))".

Now r.li has become very fast, on my laptop:

GRASS 7.1.svn (nc_spm_08_grass7):~ >  g.region -p rast=lsat5_1987_10 res=10
-a ...
rows:       1355
cols:       1503
cells:      2036565

GRASS 7.1.svn (nc_spm_08_grass7):~ > time -p r.li.simpson --o
input=lsat5_1987_10 at landsat conf=conf_diversity_5.0
output=lsat5_1987_div__simpson_size_5.0
r.li.simpson complete. Raster map <lsat5_1987_div__simpson_size_5.0>
created.
--> 29.32 seconds

or with a simulated higher resolution:

GRASS 7.1.svn (nc_spm_08_grass7):~ >  g.region -p rast=lsat5_1987_10
res=5 -aprojection: 99 (Lambert Conformal Conic) ...
rows:       2708
cols:       3005
cells:      8137540

GRASS 7.1.svn (nc_spm_08_grass7):~ > time -p r.li.simpson --o
input=lsat5_1987_10 at landsat conf=conf_diversity_5.0
output=lsat5_1987_div__simpson_size_5.0
r.li.simpson complete. Raster map <lsat5_1987_div__simpson_size_5.0>
created.
--> 227.37 seconds (used to be > 2 hours)

So, to grab this improvement for Windows, grab the version from here:
http://wingrass.fsv.cvut.cz/grass71/

(or via OSGeo4W installer). Be sure that the revision is at least
r61812 which is indicated in the file name.

Please let us know if all works to avoid that the change has any negative
impact.
Tests here did not show any changes in the output except for the speed
improvement and solved memory leak.

Subsequently also r.diversity should behave now.

I'll backport it to GRASS 7.0 release branch after some testing.

Markus

_______________________________________________
grass-user mailing list
grass-user at lists.osgeo.org
http://lists.osgeo.org/mailman/listinfo/grass-user
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/grass-dev/attachments/20141014/2b61212a/attachment.html>