[R] Size of data vs. needed memory...rule of thumb?

Prof Brian Ripley ripley at stats.ox.ac.uk
Thu Jan 25 09:04:58 CET 2007

On Wed, 24 Jan 2007, WILLIE, JILL wrote:

> I have been searching all day & most of last night, but can't find any
> benchmarking or recommendations regarding R system requirements for very
> large (2-5GB) data sets to help guide our hardware configuration.  If
> anybody has experience with this they're willing to share or could
> anybody point me in a direction that might be productive to research, it
> would be much appreciated.  Specifically:  will R simply use as much
> memory as the OS makes available to it, unlimited?

Under most OSes.  Because Windows has no means to limit the amount made 
available, R under Windows does have it own limiting mechanism (which you 
hit in the examples below).  R under Linux will allow you to run a 4GB 
process on a machine with 2GB RAM, but you are likely to get around 0.7% 
usage.  (One of my colleagues did that on a server earlier this week, 
hence the very specific answer.)

> Is there a multi-threading version R, packages?

Not to run computations in R.  Some parts of R (e.g. GUIs) and some 
libraries (e.g. some BLAS) are multithreaded.  There are multiprocess
packages, e.g. Rmpi, rpvm, snow.

> Does the core R package support 64-bit

Yes, and has for many years.

> & should I expect to see any difference in how memory's handled under 
> that version?

yes, because the address space will not get seriously fragmented.  See the 
appropriate section in R-admin.html (referenced from INSTALL).

> Is 3 GB of memory to 1GB of data a reasonable ballpark?

I'd say it was a bit low, but it really depends on the analysis you are 
doing, how 1GB of data is made up (many rows?, many cols?, etc) and so on.
Had you asked me to suggest a ratio I would have said 5.

> Our testing thus far has been on a windows 32-bit box w/1GB of RAM & 1
> CPU; it appears to indicate something like 3GB of RAM for every 1GB of
> sql table (ex-indexes, byte-sized factors).  At this point, we're
> planning on setting up a dual core 64-bit Linux box w/16GB of RAM for
> starters, since we have summed-down sql tables of approx 2-5GB
> generally.
> Here's details, just for context, or in case I'm misinterpreting the
> results, or in case there's some more memory-efficient way to get data
> in R's binary format than going w/the data.frame.

Well, sqlQuery returns a data frame directly.  I think you need to tell 
RODBC to translate your 'byte-sized factors' to numeric, as it will be 
going through character if these are a type it does not know about.

> R session:
> 		> library(RODBC)
> 		> channel<-odbcConnect("psmrd")
> 		> FivePer <-data.frame(sqlQuery(channel, "select * from
> AUTCombinedWA_BILossCost_5per"))
> 		Error: cannot allocate vector of size 2000 Kb
> 		In addition: Warning messages:
> 		1: Reached total allocation of 1023Mb: see
> help(memory.size)
> 		2: Reached total allocation of 1023Mb: see
> help(memory.size)
> ODBC connection:
> 		Microsoft SQL Server ODBC Driver Version 03.86.1830
> 		Data Source Name: psmrd
> 		Data Source Description:
> 		Server: psmrdcdw01\modeling
> 		Database: OpenSeas_Work1
> 		Language: (Default)
> 		Translate Character Data: Yes
> 		Log Long Running Queries: No
> 		Log Driver Statistics: No
> 		Use Integrated Security: Yes
> 		Use Regional Settings: No
> 		Prepared Statements Option: Drop temporary procedures on
> disconnect
> 		Use Failover Server: No
> 		Use ANSI Quoted Identifiers: Yes
> 		Use ANSI Null, Paddings and Warnings: Yes
> 		Data Encryption: No
> Please be patient, I'm a new R user (or at least I'm trying to be...at
> this point I'm mostly a new R-help-reader); I'd appreciated being
> pointed in the right direction if this isn't the right help list to send
> this question to...or if this question is poorly worded (I did read the
> posting guide).
> Jill Willie
> Open Seas
> Safeco Insurance
> jilwil at safeco.com
> 	[[alternative HTML version deleted]]
> ______________________________________________
> R-help at stat.math.ethz.ch 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.

Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595

More information about the R-help mailing list