[R] Memory Problems with CSV and Survey Objects

[Gabor Grothendieck]

Note that read.csv.sql in the sqldf package could be used to avoid
most of the setup:

library(sqldf)
DF <- read.csv.sql("myfile.csv", sql = "select ...")

It will setup the database, read the file into it, apply the select
statement, place the result into data frame DF and destroy the
database all in one line.  See Example 13 on the home page:
http://sqldf.googlecode.com

On Fri, Oct 23, 2009 at 1:24 PM,  <tlumley at u.washington.edu> wrote:
>
>
> Yes, a 350Mb data frame is a bit big for 32-bit R to handle conveniently.
>
> As you note, the survey package doesn't yet do database-backed
> replicate-weight designs. You can get the same effect yourself without too
> much work.
>
> First, put the data into a database, such as SQLite.  If you have the data
> frame read in then dbWriteTable will do it.
>
> Now, drop most of the variables, keeping the sampling weights, replicate
> weights, and a couple of other variables.
>
> Create a svrepdesign() with the reduced data set.
>
> When you want to do an analysis, use dbGetQuery() to load the variables you
> need for the analysis, and put them in the $variables component of the
> svrepdesign.
>
> That's exactly what the database-backed functions do for svydesign objects.
>
> [If you only ever want to use a small subset of the variables, it's even
> easier: drop all the extraneous variables and create a svrepdesign with the
> variables you want]
>
>       -thomas
>
> On Fri, 23 Oct 2009, Anthony Damico wrote:
>
>> I'm working with a 350MB CSV file on a server that has 3GB of RAM, yet I'm
>> hitting a memory error when I try to store the data frame into a survey
>> design object, the R object that stores data for complex sample survey
>> data.
>>
>> When I launch R, I execute the following line from Windows:
>> "C:\Program Files\R\R-2.9.1\bin\Rgui.exe" --max-mem-size=2047M
>> Anything higher, and I get an error message saying the maximum has been
>> set
>> to 2047M.
>>
>> Here are the commands:
>>>
>>> library(survey)
>>
>> #this step takes more than five minutes
>>>
>>> data08<-read.csv("data08.csv",header=TRUE,nrows=210437)
>>
>>> object.size(data08)
>>
>> #329877112 bytes
>>
>> #Looking at Windows Task Manager, Mem Usage for Rgui.exe is already
>> 659,632K
>>
>>> brr.dsgn <-svrepdesign( data = data08 , repweights = data08[, grep(
>>
>> "^repwgt" , colnames( data08)) ], type = "BRR" , combined.weights = TRUE ,
>> weights = data08$mainwgt )
>> #Error: cannot allocate vector of size 254.5 Mb
>>
>> #The survey design object does not get created.
>>
>> #This also causes Windows Task Manager, Mem Usage to spike to 1,748,136K
>>
>> #And here are some memory diagnostics
>>>
>>> memory.limit()
>>
>> [1] 2047
>>>
>>> memory.size()
>>
>> [1] 1449.06
>>>
>>> gc()
>>
>>          used  (Mb) gc trigger   (Mb)  max used   (Mb)
>> Ncells   131148   3.6     593642   15.9  15680924  418.8
>> Vcells 45479988 347.0  173526492 1324.0 220358611 1681.3
>>
>> A description of the survey package can be found here:
>> http://faculty.washington.edu/tlumley/survey/
>>
>> I tried creating a work-around by using the database-backed survey objects
>> (DB SO), included in the survey package to conserve memory on larger
>> datasets like this one.  Unfortunately, I don't think the survey package
>> supports database connections for replicate weight designs yet, since I've
>> only been able to get a database connection working after creating a
>> svydesign object and not a svrepdesign object - and also because neither
>> the
>> DB SO website nor the svrepdesign help page make any mention of those
>> parameters.
>>
>> The DB SOs are described in detail here:
>> http://faculty.washington.edu/tlumley/survey/svy-dbi.html
>>
>> Any advice would be truly appreciated.
>>
>> Thanks,
>> Anthony Damico
>>
>>        [[alternative HTML version deleted]]
>>
>> ______________________________________________
>> R-help at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-help
>> PLEASE do read the posting guide
>> http://www.R-project.org/posting-guide.html
>> and provide commented, minimal, self-contained, reproducible code.
>>
>
> Thomas Lumley                   Assoc. Professor, Biostatistics
> tlumley at u.washington.edu        University of Washington, Seattle
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>