Skip to content

SLOW split() function

11 messages · jim holtman, Dennis Murphy, William Dunlap +4 more

ivo welch
# 
dear R experts: ?apologies for all my speed and memory questions. ?I
have a bet with my coauthors that I can make R reasonably efficient
through R-appropriate programming techniques.  this is not just for
kicks, but for work.  for benchmarking, my [3 year old] Mac Pro has
2.8GHz Xeons, 16GB of RAM, and R 2.13.1.

right now, it seems that 'split()' is why I am losing my bet. ?(split
is an integral component of *apply() and by(), so I need split() to be
fast.  its resulting list can then be fed, e.g., to mclapply().)  I
made up an example to illustrate my ills:

    library(data.table)
    N <- 1000
    T <- N*10
    d <- data.table(data.frame( key= rep(1:T, rep(N,T)), val=rnorm(N*T) ))
    setkey(d, "key"); gc() ## force a garbage collection
    cat("N=", N, ". ?Size of d=", object.size(d)/1024/1024, "MB\n")
    print(system.time( s<-split(d, d$key) ))

My ordered input data table (or data frame; doesn't make a difference)
is 114MB in size.  it takes about a second to create.  split() only
needs to reshape it.  this simple operation takes almost 5 minutes on
my computer.

with a data set that is larger, this explodes further.

am I doing something wrong?  is there an alternative to split()?

sincerely,

/iaw

----
Ivo Welch (ivo.welch at gmail.com)
jim holtman
# 
instead of spliting the entire dataframe, split the indices and then use these to access your data: try 

system.time(s <- split(seq(nrow(d)), d$key))

this should be faster and less memory intensive.  you can then use the indices to access the subset:

result <- lapply(s, function(.indx){
    doSomething <- sum(d$someCol[.indx])
})

Sent from my iPad
On Oct 10, 2011, at 21:01, ivo welch <ivo.welch at gmail.com> wrote:

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Dennis Murphy
    

    

      
      #
    

  


  

      
I tried this:

library(data.table)
    N <- 1000
    T <- N*10
d <- data.table(gp= rep(1:T, rep(N,T)), val=rnorm(N*T), key = 'gp')
dim(d)
[1] 10000000        2

# On my humble 8Gb system,

            

      
      
      
    

        
user  system elapsed
   4.15    0.09    4.27

I wouldn't be surprised if there were a much faster way to do this
operation in data.table since split() is a data frame operation. This
is about as fast as Jim Holtman's suggestion:

system.time(s <- split(seq_len(nrow(d)), d$gp))
   user  system elapsed
   4.15    0.09    4.29

HTH,
Dennis

        
On Mon, Oct 10, 2011 at 6:01 PM, ivo welch <ivo.welch at gmail.com> wrote:

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      William Dunlap
    

    

      
      #
    

  


  

      
The following avoids the overhead of data.frame methods
(and assumes the data.frame doesn't include matrices
or other data.frames) and relies on split(vector,factor)
quickly splitting a vector into a list of vectors.
For a 10^6 row by 10 column data.frame split in 10^5
groups this took 14.1 seconds while split took 658.7 s.
Both returned the same thing.

Perhaps something based on this idea would help your
parallelized by().

mysplit.data.frame <-
function (x, f, drop = FALSE, ...)
{
    f <- as.factor(f)
    tmp <- lapply(x, function(xi) split(xi, f, drop = drop, ...))
    rn <- split(rownames(x), f, drop = drop, ...)
    tmp <- unlist(unname(tmp), recursive = FALSE)
    tmp <- split(tmp, factor(names(tmp), levels = unique(names(tmp))))
    tmp <- lapply(setNames(seq_along(tmp), names(tmp)), function(i) {
        t <- tmp[[i]]
        names(t) <- names(x)
        attr(t, "row.names") <- rn[[i]]
        class(t) <- "data.frame"
        t
    })
    tmp
} 

Bill Dunlap
Spotfire, TIBCO Software
wdunlap tibco.com

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

  
  


  


      

    

    

      
      

        


  

        

      ivo welch
    

    

      
      #
    

  


  

      
thank you, everyone.  this was very helpful to my specific task and
understanding.  for the benefit of future googlers, I thought I would
post some experiments and results here.

ultimately, I need to do a by() on an irregular matrix, and I now know
how to speed up by() on a single-core, and then again on a multi-core
machine.

library(data.table)
N <- 1000*1000
d <- data.table(data.frame( key= as.integer(runif(N, min=1,
max=N/10)), x=rnorm(N), y=rnorm(N) ))  # irregular
setkey(d, "key"); gc() ## sort and force a garbage collection


cat("N=", N, ".  Size of d=", object.size(d)/1024/1024, "MB\n")

cat("\nStandard by() Function:\n")
print(system.time( all.1 <- by( d, d$key, function(d) coef(lm(y ~ x, data=d)))))


cat("\n\nPreSplit Function [aka Jim H]\n\t(a) Splitting Operation:\n")
print(system.time(si <- split(seq(nrow(d)), d$key)))
cat("\n\t(b) Regressions:\n")
print(system.time(all.2 <- lapply(si, function(.indx) {
coef(lm(d$y[.indx] ~ d$x[.indx])) })))
print(system.time(all.2b <- lapply(si, function(.indx) { coef(lm(y ~
x, data=d[.indx,])) })))

cat("\n\nNaive Split Data Frame\n\t(a) Splitting Operation:\n")
print(system.time(ds <- split(d, d$key)))
cat("\n\t(b) Regressions:\n")
print(system.time(all.3a <- lapply(ds, function(ds) { coef(lm(ds$y ~ ds$x)) })))
print(system.time(all.3b <- lapply(ds, function(ds) { coef(lm(y ~ x,
data=ds)) })))

the first and the last ways (all.1 and all.3) are "naive" ways of
doing this, and take about 400-500 seconds on a Mac Air, core i5.
Jim's suggestion (all.2) cuts this roughly into half by speeding up
the split to take almost no time.

and now,

library(multicore)
print(system.time(all.4 <- mclapply(si, function(.indx) { coef(lm(y ~
x, data=d[.indx,])) })))

on my dual-core (quad-thread) i5, all four pseudo cores become busy,
and the time roughly halves again from 230 seconds to 120 seconds.


maybe the by() function should use Jim's approach, and multicore
should provide mcby().  of course, knowing how to do this myself fast
now by hand, this is not so important for me.  but it may help some
other novices.

thanks again everybody.

regards,

/iaw

----
Ivo Welch (ivo.welch at gmail.com)

        
On Mon, Oct 10, 2011 at 9:31 PM, William Dunlap <wdunlap at tibco.com> wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Joshua Wiley
    

    

      
      #
    

  


  

      
As another followup, given that you are doing numerous regression
models and (I presume) working with finance/stock data that is
strictly numeric (no need for special contrast coding, etc.), you can
substantially reduce the time spent estimating the coefficients.  A
simple way is to use lm.fit directly instead of lm.  For lm.fit, you
pass the y and x (design) matrices directly.  This skips a good deal
of overhead.  Here is one naive way, I imagine more speedups could be
gained by incorporating the intercept (1 vector) into d instead of
cbind()ing it.  The catch it that lm.fit requires matrices, not data
tables, so what you gain may be lost in having to do an extra
conversion.  In any case, here are the times on my system for the two
options (note I used N = 1000 * 100 because I am presently on a
glorified netbook).

            

      
      
      
    

        
+ x, data=d[.indx,])) })))
   user  system elapsed
  69.00    0.00   69.56

            

      
      
      
    

        
user  system elapsed
  37.83    0.03   38.36

the column names for the coeficients will not be the same as from lm,
but the estimates should be identical.  While this is not recommended
in typical usage, in an application like regressions on rolling time
windows, etc. where you know the data are not changing, I think it
makes sense to bypass the clever determine your data and best methods
to use, and go straight to passing the design matrix.  Since you do
not need residuals, variances, etc. it may be possible to speed this
up even more, perhaps bypassing dqrls altogether.

Cheers,

Josh

        
On Mon, Oct 10, 2011 at 9:56 PM, ivo welch <ivo.welch at gmail.com> wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      ivo welch
    

    

      
      #
    

  


  

      
thanks, josh.  in my posting example, I did not need anything except
coefficients.  (when this is the case, I usually do not even use
lm.fit, but I eliminate all missing obs first and then use solve
crossprod(y,cbind(1,x)) crossprod(cbind(1,x)).)  this is pretty fast.)

alas, I will need to figure how to get coef standard errors faster in
this case.  summary.lm() is really slow.

regards,

/iaw
----
Ivo Welch (ivo.welch at gmail.com)
http://www.ivo-welch.info/
J. Fred Weston Professor of Finance
Anderson School at UCLA, C519

        
On Mon, Oct 10, 2011 at 11:30 PM, Joshua Wiley <jwiley.psych at gmail.com> wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Joshua Wiley
    

    

      
      #
    

  


  

      
I do not know if stripping down functions is generally recommended,
but it is not too difficult to do if you know that you can make
assumptions.  Here is an example (I also found a fast way to convert
the data table to a matrix, again if some assumptions can be made).
Using the stripped down function, you can get coefficients and
standard errors in less time than you can get just coefficients using
default lm.  It is hugely less flexible.

Cheers,

Josh

##########################################
library(data.table)

## stripped down lm and summary.lm (for standard errors)
minimal.lm <- function(y, x) {
  dims <- dim(x)
  x <- unlist(x, FALSE, FALSE)
  dim(x) <- dims
  obj <- lm.fit(x = x, y = y)
  resvar <- sum(obj$residuals^2)/obj$df.residual
  p <- obj$rank
  R <- .Call("La_chol2inv", x = obj$qr$qr[1L:p, 1L:p, drop = FALSE],
size = p, PACKAGE = "base")
  m <- min(dim(R))
  d <- c(R)[1L + 0L:(m - 1L) * (dim(R)[1L] + 1L)]
  se <- sqrt(d * resvar)
  cbind(coef = obj$coefficients, se)
}

N <- 1000*100
d <- data.table(data.frame( key= as.integer(runif(N, min=1,
max=N/10)), x=rnorm(N), y=rnorm(N) ))  # irregular
## add intercept column
d$int <- 1L
setkey(d, "key"); gc() ## sort and force a garbage collection

cat("N=", N, ".  Size of d=", object.size(d)/1024/1024, "MB\n")
print(system.time(si <- split(seq(nrow(d)), d$key)))

cat("\n\t(b) Regressions:\n")
## using lm
print(system.time(all.2b <- lapply(si, function(.indx) { coef(lm(y ~
x, data=d[.indx,])) })))
## using minimal.lm---faster and gives standard errors
print(system.time(all.2c <- lapply(si, function(.indx) { minimal.lm(y
= d[.indx, y], x = d[.indx, list(int, x)]) })))

#### Timings on my system ####

            

      
      
      
    

        
+ x, data=d[.indx,])) })))
   user  system elapsed
  67.87    0.01   68.46

            

      
      
      
    

        
user  system elapsed
  47.72    0.00   48.00

######################################################################

        
On Tue, Oct 11, 2011 at 8:56 AM, ivo welch <ivo.welch at gmail.com> wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Thomas Lumley
    

    

      
      #
    

  


  

      
On Wed, Oct 12, 2011 at 4:56 AM, ivo welch <ivo.welch at gmail.com> wrote:

            

      
      
      
    

        
solve(cbind(1,x), y) should be even faster, and more numerically stable,
 [and less likely to make certain people want to cast you into the
outer darkness, where there is SAS and gnashing of teeth]

            

      
      
      
    

        
The code from summary.lm that actually computes the standard errors is
fairly efficient; you could extract that.

   -thomas

  
    

      
      
    

  


  


      

    

    
1 day later

    

      
      

        


  

        

      Matt Dowle
    

    

      
      #
    

  


  

      
Using Josh's nice example, with data.table's built-in 'by' (optimised
grouping) yields a 6 times speedup (100 seconds down to 15 on
my netbook).

            

      
      
      
    

        
+ x, data=d[.indx,])) })) 
   user  system elapsed 
144.501   0.300 145.525

            

      
      
      
    

        
+ = d[.indx, y], x = d[.indx, list(int, x)]) })) 
   user  system elapsed 
100.819   0.084 101.552

            

      
      
      
    

        
user  system elapsed 
 15.269   0.012  15.323   # 6 times faster

            

      
      
      
    

        
$`1`
        coef        se
x1 0.5152438 0.6277254
x2 0.5621320 0.5754560

$`2`
        coef       se
x1 0.2228235 0.312918
x2 0.3312261 0.261529

$`3`
         coef        se
x1 -0.1972439 0.4674000
x2 -0.1674313 0.4479957

$`4`
          coef        se
x1 -0.13915746 0.2729158
x2 -0.03409833 0.2212416

$`5`
           coef        se
x1  0.007969786 0.2389103
x2 -0.083776526 0.2046823

$`6`
          coef        se
x1 -0.58576454 0.5677619
x2 -0.07249539 0.5009013

            

      
      
      
    

        
key       coef        V2
[1,]   1  0.5152438 0.6277254
[2,]   1  0.5621320 0.5754560
[3,]   2  0.2228235 0.3129180
[4,]   2  0.3312261 0.2615290
[5,]   3 -0.1972439 0.4674000
[6,]   3 -0.1674313 0.4479957

            

      
      
      
    

        
function(y, x) { 
  obj <- lm.fit(x = x, y = y) 
  resvar <- sum(obj$residuals^2)/obj$df.residual 
  p <- obj$rank 
  R <- .Call("La_chol2inv", x = obj$qr$qr[1L:p, 1L:p, drop = FALSE], 
size = p, PACKAGE = "base") 
  m <- min(dim(R)) 
  d <- c(R)[1L + 0L:(m - 1L) * (dim(R)[1L] + 1L)] 
  se <- sqrt(d * resvar) 
  list(coef = obj$coefficients, se) 
}

            

      
      
      
    

        
--
View this message in context: http://r.789695.n4.nabble.com/SLOW-split-function-tp3892349p3900851.html
Sent from the R help mailing list archive at Nabble.com.

  
  


  


      

    

    

      
      

        


  

        

      Joshua Wiley
    

    

      
      #
    

  


  

      
Very nice!  I am quite impressed at how flexible data.table is.

        
On Thu, Oct 13, 2011 at 1:05 AM, Matthew Dowle <mdowle at mdowle.plus.com> wrote: