ff: ff classes for representing (large) atomic data

Description

The ff package provides atomic data structures that are stored on disk but behave (almost) as if they were in RAM by mapping only a section (pagesize) into main memory (the effective main memory consumption per ff object). Several access optimization techniques such as Hyrid Index Preprocessing (as.hi, update.ff) and Virtualization (virtual, vt, vw) are implemented to achieve good performance even with large datasets. In addition to the basic access functions, the ff package also provides compatibility functions that facilitate writing code for ff and ram objects (clone, as.ff, as.ram) and very basic support for operating on ff objects (ffapply). While the (possibly packed) raw data is stored on a flat file, meta informations about the atomic data structure such as its dimension, virtual storage mode (vmode), factor level encoding, internal length etc.. are stored as an ordinary R object (external pointer plus attributes) and can be saved in the workspace. The raw flat file data encoding is always in native machine format for optimal performance and provides several packing schemes for different data types such as logical, raw, integer and double (in an extended version support for more tighly packed virtual data types is supported). flatfile data files can be shared among ff objects in the same R process or even from different R processes due to Memory-Mapping, although the caching effects have not been tested extensively.
Please do read and understand the limitations and warnings in LimWarn before you do anything serious with package ff.

Usage

ff( initdata  = NULL
, length      = NULL
, levels      = NULL
, ordered     = NULL
, dim         = NULL
, dimorder    = NULL
, bydim       = NULL
, symmetric   = FALSE
, fixdiag     = NULL
, names       = NULL
, dimnames    = NULL
, ramclass    = NULL
, ramattribs  = NULL
, vmode       = NULL
, update      = NULL
, pattern     = NULL
, filename    = NULL
, overwrite   = FALSE
, readonly    = FALSE
, pagesize    = NULL  # getOption("ffpagesize")
, caching     = NULL  # getOption("ffcaching")
, finalizer   = NULL
, finonexit   = NULL  # getOption("fffinonexit")
, FF_RETURN   = TRUE
, BATCHSIZE   = .Machine$integer.max
, BATCHBYTES  = getOption("ffbatchbytes")
, VERBOSE     = FALSE
)

Value

If (!FF_RETURN) then a ram object like those generated by vector, matrix, array but with attributes 'vmode', 'physical' and 'virtual' accessible via vmode, physical and virtual

If (FF_RETURN) an object of class 'ff' which is a a list with two components:

physical: an external pointer of class 'ff_pointer' which carries attributes with copy by reference semantics: changing a physical attribute of a copy changes the original
virtual: an empty list which carries attributes with copy by value semantics: changing a virtual attribute of a copy does not change the original

Arguments

initdata: scalar or vector of the .vimplemented vmodes, recycled if needed, default 0, see also as.vmode and vector.vmode
length: optional vector length of the object (default: derive from 'initdata' or 'dim'), see length.ff
levels: optional character vector of levels if (in this case initdata must be composed of these) (default: derive from initdata)
ordered: indicate whether the levels are ordered (TRUE) or non-ordered factor (FALSE, default)
dim: optional array dim, see dim.ff and array
dimorder: physical layout (default seq_along(dim)), see dimorder and aperm
bydim: dimorder by which to interpret the 'initdata', generalization of the 'byrow' paramter in matrix
symmetric: extended feature: TRUE creates symmetric matrix (default FALSE)
fixdiag: extended feature: non-NULL scalar requires fixed diagonal for symmetric matrix (default NULL is free diagonal)
names: NOT taken from initdata, see names
dimnames: NOT taken from initdata, see dimnames
ramclass: class attribute attached when moving all or parts of this ff into ram, see ramclass
ramattribs: additional attributes attached when moving all or parts of this ff into ram, see ramattribs
vmode: virtual storage mode (default: derive from 'initdata'), see vmode and as.vmode
update: set to FALSE to avoid updating with 'initdata' (default TRUE) (used by ffdf)
pattern: root pattern with or without path for automatic ff filename creation (default NULL translates to "ff"), see also argument 'filename'
filename: ff filename with or without path (default tmpfile with 'pattern' prefix); without path the file is created in getOption("fftempdir"), with path '.' the file is created in getwd. Note that files created in getOption("fftempdir") have default finalizer "delete" while other files have default finalizer "close". See also arguments 'pattern' and 'finalizer' and physical
overwrite: set to TRUE to allow overwriting existing files (default FALSE)
readonly: set to TRUE to forbid writing to existing files
pagesize: pagesize in bytes for the memory mapping (default from getOptions("ffpagesize") initialized by getdefaultpagesize), see also physical
caching: caching scheme for the backend, currently 'mmnoflush' or 'mmeachflush' (flush mmpages at each swap, default from getOptions("ffcaching") initialized with 'mmeachflush'), see also physical
finalizer: name of finalizer function called when ff object is removed (default: ff files created in getOptions("fftempdir") are considered temporary and have default finalizer delete, files created in other locations have default finalizer close); available finalizer generics are "close", "delete" and "deleteIfOpen", available methods are close.ff, delete.ff and deleteIfOpen.ff, see also argument 'finonexit' and finalizer
finonexit: logical scalar determining whether and finalize is also called when R is closed via q, (default TRUE from getOptions("fffinonexit"))
FF_RETURN: logical scalar or ff object to be used. The default TRUE creates a new ff file. FALSE returns a ram object. Handing over an ff object here uses this or stops if not ffsuitable
BATCHSIZE: integer scalar limiting the number of elements to be processed in update.ff when length(initdata)>1, default from .Machine$integer.max
BATCHBYTES: integer scalar limiting the number of bytes to be processed in update.ff when length(initdata)>1, default from getOption("ffbatchbytes"), see also .rambytes
VERBOSE: set to TRUE for verbosing in update.ff when length(initdata)>1, default FALSE

Physical object component

The 'ff_pointer' carries the following 'physical' or readonly attributes, which are accessible via physical:

`vmode`	see `vmode`
`maxlength`	see `maxlength`
`pattern`	see parameter 'pattern'
`filename`	see `filename`
`pagesize`	see parameter 'pagesize'
`caching`	see parameter 'caching'
`finalizer`	see parameter 'finalizer'
`finonexit`	see parameter 'finonexit'
`readonly`	see `is.readonly`
`class`	The external pointer needs class 'ff_pointer' to allow method dispatch of finalizers

Virtual object component

The 'virtual' component carries the following attributes (some of which might be NULL):

`Length`	see `length.ff`
`Levels`	see `levels.ff`
`Names`	see `names.ff`
`VW`	see `vw.ff`
`Dim`	see `dim.ff`
`Dimorder`	see `dimorder`
`Symmetric`	see `symmetric.ff`
`Fixdiag`	see `fixdiag.ff`
`ramclass`	see `ramclass`
`ramattribs`	see `ramattribs`

Class

You should not rely on the internal structure of ff objects or their ram versions. Instead use the accessor functions like vmode, physical and virtual. Still it would be wise to avoid attributes AND classes 'vmode', 'physical' and 'virtual' in any other packages. Note that the 'ff' object's class attribute also has copy-by-value semantics ('virtual'). For the 'ff' object the following class attritibutes are known:

vector	`c("ff_vector","ff")`
matrix	`c("ff_matrix","ff_array","ff")`
array	`c("ff_array","ff")`
symmetric matrix	`c("ff_symm","ff")`
distance matrix	`c("ff_dist","ff_symm","ff")`
reserved for future use	`c("ff_mixed","ff")`

Methods

The following methods and functions are available for ff objects:

Type	Name	Assign	Comment
			Basic functions
function	`ff`		constructor for ff and ram objects
generic	`update`		updates one ff object with the content of another
generic	`clone`		clones an ff object optionally changing some of its features
method	`print`		print ff
method	`str`		ff object structure
			Class test and coercion
function	`is.ff`		check if inherits from ff
generic	`as.ff`		coerce to ff, if not yet
generic	`as.ram`		coerce to ram retaining some of the ff information
generic	`as.bit`		coerce to `bit`
			Virtual storage mode
generic	`vmode`	`<-`	get and set virtual mode (setting only for ram, not for ff objects)
generic	`as.vmode`		coerce to vmode (only for ram, not for ff objects)
			Physical attributes
function	`physical`	`<-`	set and get physical attributes
generic	`filename`	<-	get and set filename
generic	`pattern`	<-	get pattern and set filename path and prefix via pattern
generic	`maxlength`		get maxlength
generic	`is.sorted`	`<-`	set and get if is marked as sorted
generic	`na.count`	`<-`	set and get NA count, if set to non-NA only swap methods can change and na.count is maintained automatically
generic	`is.readonly`		get if is readonly
			Virtual attributes
function	`virtual`	`<-`	set and get virtual attributes
method	`length`	`<-`	set and get length
method	`dim`	`<-`	set and get dim
generic	`dimorder`	`<-`	set and get the order of dimension interpretation
generic	`vt`		virtually transpose ff_array
method	`t`		create transposed clone of ff_array
generic	`vw`	`<-`	set and get virtual windows
method	`names`	`<-`	set and get names
method	`dimnames`	`<-`	set and get dimnames
generic	`symmetric`		get if is symmetric
generic	`fixdiag`	`<-`	set and get fixed diagonal of symmetric matrix
method	`levels`	`<-`	levels of factor
generic	`recodeLevels`		recode a factor to different levels
generic	`sortLevels`		sort the levels and recoce a factor
method	`is.factor`		if is factor
method	`is.ordered`		if is ordered (factor)
generic	`ramclass`		get ramclass
generic	`ramattribs`		get ramattribs
			Access functions
function	`get.ff`		get single ff element (currently `[[` is a shortcut)
function	`set.ff`		set single ff element (currently `[[<-` is a shortcut)
function	`getset.ff`		set single ff element and get old value in one access operation
function	`read.ff`		get vector of contiguous elements
function	`write.ff`		set vector of contiguous elements
function	`readwrite.ff`		set vector of contiguous elements and get old values in one access operation
method	`[`		get vector of indexed elements, uses HIP, see `hi`
method	`[<-`		set vector of indexed elements, uses HIP, see `hi`
generic	`swap`		set vector of indexed elements and get old values in one access operation
generic	`add`		(almost) unifies '+=' operation for ff and ram objects
generic	`bigsample`		sample from ff object
			Opening/Closing/Deleting
generic	`is.open`		check if ff is open
method	`open`		open ff object (is done automatically on access)
method	`close`		close ff object (releases C++ memory and protects against file deletion if `deleteIfOpen`) finalizer is used
generic	`delete`		deletes ff file (unconditionally)
generic	`deleteIfOpen`		deletes ff file if ff object is open (finalization method)
generic	`finalizer`	<-	get and set finalizer
generic	`finalize`		force finalization
			Other
function	`geterror.ff`		get error code
function	`geterrstr.ff`		get error message

ff options

Through options or getOption one can change and query global features of the ff package:

option	description	default
`fftempdir`	default directory for creating ff files	`tempdir`
`fffinalizer`	name of default finalizer	`deleteIfOpen`
`fffinonexit`	default for invoking finalizer on exit of R	`TRUE`
`ffpagesize`	default pagesize	`getdefaultpagesize`
`ffcaching`	caching scheme for the C++ backend	`'mmnoflush'`
`ffdrop`	default for the drop parameter in the ff subscript methods	TRUE
`ffbatchbytes`	default for the byte limit in batched/chunked processing	16MB

OS specific

The following table gives an overview of file size limits for common file systems (see https://en.wikipedia.org/wiki/Comparison_of_file_systems for further details):

File System	File size limit
FAT16	2GB
FAT32	4GB
NTFS	16GB
ext2/3/4	16GB to 2TB
ReiserFS	4GB (up to version 3.4) / 8TB (from version 3.5)
XFS	8EB
JFS	4PB
HFS	2GB
HFS Plus	16GB
USF1	4GB to 256TB
USF2	512GB to 32PB
UDF	16EB

Credits

Package Version 1.0

Daniel Adler	dadler@uni-goettingen.de
	R package design, C++ generic file vectors, Memory-Mapping, 64-bit Multi-Indexing adapter and Documentation, Platform ports
Oleg Nenadic	onenadi@uni-goettingen.de
	Index sequence packing, Documentation
Walter Zucchini	wzucchi@uni-goettingen.de
	Array Indexing, Sampling, Documentation
Christian Gläser	christian_glaeser@gmx.de
	Wrapper for biglm package

Package Version 2.0

Jens Oehlschlägel	Jens.Oehlschlaegel@truecluster.com
	R package redesign; Hybrid Index Preprocessing; transparent object creation and finalization; vmode design; virtualization and hybrid copying; arrays with dimorder and bydim; symmetric matrices; factors and POSIXct; virtual windows and transpose; new generics update, clone, swap, add, as.ff and as.ram; ffapply and collapsing functions. R-coding, C-coding and Rd-documentation.
Daniel Adler	dadler@uni-goettingen.de
	C++ generic file vectors, vmode implementation and low-level bit-packing/unpacking, arithmetic operations and NA handling, Memory-Mapping and backend caching. C++ coding and platform ports. R-code extensions for opening existing flat files readonly and shared.

Licence

Package under GPL-2, included C++ code released by Daniel Adler under the less restrictive ISCL

Details

The atomic data is stored in filename as a native encoded raw flat file on disk, OS specific limitations of the file system apply. The number of elements per ff object is limited to the integer indexing, i.e. .Machine$integer.max. Atomic objects created with ff are is.open, a C++ object is ready to access the file via memory-mapping. Currently the C++ backend provides two caching schemes: 'mmnoflush' let the OS decide when to flash memory mapped pages and 'mmeachflush' will flush memory mapped pages at each page swap per ff file. These minimal memory ressources can be released by closeing or deleteing the ff file. ff objects can be saved and loaded across R sessions. If the ff file still exists in the same location, it will be opened automatically at the first attempt to access its data. If the ff object is removed, at the next garbage collection (see gc) the ff object's finalizer is invoked. Raw data files can be made accessible as an ff object by explicitly given the filename and vmode but no size information (length or dim). The ff object will open the file and handle the data with respect to the given vmode. The close finalizer will close the ff file, the delete finalizer will delete the ff file. The default finalizer deleteIfOpen will delete open files and do nothing for closed files. If the default finalizer is used, two actions are needed to protect the ff file against deletion: create the file outside the standard 'fftempdir' and close the ff object before removing it or before quitting R. When R is exited through q, the finalizer will be invoked depending on the 'fffinonexit' option, furthermore the 'fftempdir' is unlinked.

Examples

Run this code

  message("make sure you understand the following ff options 
    before you start using the ff package!!")
  oldoptions <- options(fffinalizer="deleteIfOpen", fffinonexit="TRUE", fftempdir=tempdir())
  message("an integer vector")
  ff(1:12)                  
  message("a double vector of length 12")
  ff(0, 12)
  message("a 2-bit logical vector of length 12 (vmode='boolean' has 1 bit)")
  ff(vmode="logical", length=12)
  message("an integer matrix 3x4 (standard colwise physical layout)")
  ff(1:12, dim=c(3,4))
  message("an integer matrix 3x4 (rowwise physical layout, but filled in standard colwise order)")
  ff(1:12, dim=c(3,4), dimorder=c(2,1))
  message("an integer matrix 3x4 (standard colwise physical layout, but filled in rowwise order
aka matrix(, byrow=TRUE))")
  ff(1:12, dim=c(3,4), bydim=c(2,1))
  gc()
  options(oldoptions)

  if (ffxtensions()){
     message("a 26-dimensional boolean array using 1-bit representation
      (file size 8 MB compared to 256 MB int in ram)")
     a <- ff(vmode="boolean", dim=rep(2, 26))
     dimnames(a) <- dummy.dimnames(a)
     rm(a); gc()
  }

  if (FALSE) {

     message("This 2GB biglm example can take long, you might want to change
       the size in order to define a size appropriate for your computer")
     require(biglm)

     b <- 1000
     n <- 100000
     k <- 3
     memory.size(max = TRUE)
     system.time(
     x <- ff(vmode="double", dim=c(b*n,k), dimnames=list(NULL, LETTERS[1:k]))
     )
     memory.size(max = TRUE)
     system.time(
     ffrowapply({
        l <- i2 - i1 + 1
        z <- rnorm(l)
        x[i1:i2,] <- z + matrix(rnorm(l*k), l, k)
     }, X=x, VERBOSE=TRUE, BATCHSIZE=n)
     )
     memory.size(max = TRUE)

     form <- A ~ B + C
     first <- TRUE
     system.time(
     ffrowapply({
        if (first){
          first <- FALSE
          fit <- biglm(form, as.data.frame(x[i1:i2,,drop=FALSE], stringsAsFactors = TRUE))
        }else
          fit <- update(fit, as.data.frame(x[i1:i2,,drop=FALSE], stringsAsFactors = TRUE))
     }, X=x, VERBOSE=TRUE, BATCHSIZE=n)
     )
     memory.size(max = TRUE)
     first
     fit
     summary(fit)
     rm(x); gc()
  }

Run the code above in your browser using DataLab