[datatable-help] speeding up perception
Thanks for the replies and info. An attempt at fast
assign is now committed to data.table v1.6.3 on
R-Forge. From NEWS :
o Fast update is now implemented, FR#200.
DT[i,j]<-value is now handled by data.table in C rather
than falling through to data.frame methods.
Thanks to Ivo Welch for raising speed issues on r-devel,
to Simon Urbanek for the suggestion, and Luke Tierney and
Simon for information on R internals.
[<- syntax still incurs one working copy of the whole
table (as of R 2.13.0) due to R's [<- dispatch mechanism
copying to `*tmp*`, so, for ultimate speed and brevity,
'within' syntax is now available as follows.
o A new 'within' argument has been added to [.data.table,
by default TRUE. It is very similar to the within()
function in base R. If an assignment appears in j, it
assigns to the column of DT, by reference; e.g.,
DT[i,colname<-value]
This syntax makes no copies of any part of memory at all.
> m = matrix(1,nrow=100000,ncol=100)
> DF = as.data.frame(m)
> DT = as.data.table(m)
> system.time(for (i in 1:1000) DF[1,1] <- 3)
user system elapsed
287.730 323.196 613.453
> system.time(for (i in 1:1000) DT[1,V1 <- 3])
user system elapsed
1.152 0.004 1.161 # 528 times faster
Please note :
*******************************************************
** Within syntax is presently highly experimental. **
*******************************************************
http://datatable.r-forge.r-project.org/
On Wed, 2011-07-06 at 09:08 -0500, luke-tierney at uiowa.edu wrote:
On Wed, 6 Jul 2011, Simon Urbanek wrote:
Interesting, and I stand corrected:
x = data.frame(a=1:n,b=1:n) .Internal(inspect(x))
@103511c00 19 VECSXP g0c2 [OBJ,NAM(2),ATT] (len=2, tl=0) @102c7b000 13 INTSXP g0c7 [] (len=100000, tl=0) 1,2,3,4,5,... @102af3000 13 INTSXP g0c7 [] (len=100000, tl=0) 1,2,3,4,5,...
x[1,1]=42L .Internal(inspect(x))
@10349c720 19 VECSXP g0c2 [OBJ,NAM(2),ATT] (len=2, tl=0) @102c19000 13 INTSXP g0c7 [] (len=100000, tl=0) 42,2,3,4,5,... @102b55000 13 INTSXP g0c7 [] (len=100000, tl=0) 1,2,3,4,5,...
x[[1]][1]=42L .Internal(inspect(x))
@103511a78 19 VECSXP g1c2 [OBJ,MARK,NAM(2),ATT] (len=2, tl=0) @102e65000 13 INTSXP g0c7 [] (len=100000, tl=0) 42,2,3,4,5,... @101f14000 13 INTSXP g1c7 [MARK] (len=100000, tl=0) 1,2,3,4,5,...
x[[1]][1]=42L .Internal(inspect(x))
@10349c800 19 VECSXP g0c2 [OBJ,NAM(2),ATT] (len=2, tl=0) @102a2f000 13 INTSXP g0c7 [] (len=100000, tl=0) 42,2,3,4,5,... @102ec7000 13 INTSXP g0c7 [] (len=100000, tl=0) 1,2,3,4,5,... I have R to release ;) so I won't be looking into this right now, but it's something worth investigating ... Since all the inner contents have NAMED=0 I would not expect any duplication to be needed, but apparently becomes so is at some point ...
The internals assume in various places that deep copies are made (one of the reasons NAMED setings are not propagated to sub-sturcture). The main issues are avoiding cycles and that there is no easy way to check for sharing. There may be some circumstances in which a shallow copy would be OK but making sure it would be in all cases is probably more trouble than it is worth at this point. (I've tried this in the past in a few cases and always had to back off.) Best, luke
Cheers, Simon On Jul 6, 2011, at 4:36 AM, Matthew Dowle wrote:
On Tue, 2011-07-05 at 21:11 -0400, Simon Urbanek wrote:
No subassignment function satisfies that condition, because you can always call them directly. However, that doesn't stop the default method from making that assumption, so I'm not sure it's an issue. David, Just to clarify - the data frame content is not copied, we are talking about the vector holding columns.
If it is just the vector holding the columns that is copied (and not the columns themselves), why does n make a difference in this test (on R 2.13.0)?
n = 1000 x = data.frame(a=1:n,b=1:n) system.time(for (i in 1:1000) x[1,1] <- 42L)
user system elapsed 0.628 0.000 0.628
n = 100000 x = data.frame(a=1:n,b=1:n) # still 2 columns, but longer columns system.time(for (i in 1:1000) x[1,1] <- 42L)
user system elapsed 20.145 1.232 21.455
With $<- :
n = 1000 x = data.frame(a=1:n,b=1:n) system.time(for (i in 1:1000) x$a[1] <- 42L)
user system elapsed 0.304 0.000 0.307
n = 100000 x = data.frame(a=1:n,b=1:n) system.time(for (i in 1:1000) x$a[1] <- 42L)
user system elapsed 37.586 0.388 38.161
If it's because the 1st column needs to be copied (only) because that's the one being assigned to (in this test), that magnitude of slow down doesn't seem consistent with the time of a vector copy of the 1st column :
n=100000 v = 1:n system.time(for (i in 1:1000) v[1] <- 42L)
user system elapsed 0.016 0.000 0.017
system.time(for (i in 1:1000) {v2=v;v2[1] <- 42L})
user system elapsed 1.816 1.076 2.900 Finally, increasing the number of columns, again only the 1st is assigned to :
n=100000 x = data.frame(rep(list(1:n),100)) dim(x)
[1] 100000 100
system.time(for (i in 1:1000) x[1,1] <- 42L)
user system elapsed 167.974 50.903 219.711
Cheers, Simon Sent from my iPhone On Jul 5, 2011, at 9:01 PM, David Winsemius <dwinsemius at comcast.net> wrote:
On Jul 5, 2011, at 7:18 PM, <luke-tierney at uiowa.edu> <luke-tierney at uiowa.edu> wrote:
On Tue, 5 Jul 2011, Simon Urbanek wrote:
On Jul 5, 2011, at 2:08 PM, Matthew Dowle wrote:
Simon (and all), I've tried to make assignment as fast as calling `[<-.data.table` directly, for user convenience. Profiling shows (IIUC) that it isn't dispatch, but x being copied. Is there a way to prevent '[<-' from copying x?
Good point, and conceptually, no. It's a subassignment after all - see R-lang 3.4.4 - it is equivalent to `*tmp*` <- x x <- `[<-`(`*tmp*`, i, j, value) rm(`*tmp*`) so there is always a copy involved. Now, a conceptual copy doesn't mean real copy in R since R tries to keep the pass-by-value illusion while passing references in cases where it knows that modifications cannot occur and/or they are safe. The default subassign method uses that feature which means it can afford to not duplicate if there is only one reference -- then it's safe to not duplicate as we are replacing that only existing reference. And in the case of a matrix, that will be true at the latest from the second subassignment on. Unfortunately the method dispatch (AFAICS) introduces one more reference in the dispatch chain so there will always be two references so duplication is necessary. Since we have only 0 / 1 / 2+ information on the references, we can't distinguish whether the second reference is due to the dispatch or due to the passed object having more than one reference, so we have to duplicate in any case. That is unfortunate, and I don't see a way around (unless we handle subassignment methods is some special way).
I don't believe dispatch is bumping NAMED (and a quick experiment seems to confirm this though I don't guarantee I did that right). The issue is that a replacement function implemented as a closure, which is the only option for a package, will always see NAMED on the object to be modified as 2 (because the value is obtained by forcing the argument promise) and so any R level assignments will duplicate. This also isn't really an issue of imprecise reference counting -- there really are (at least) two legitimate references -- one though the argument and one through the caller's environment. It would be good it we could come up with a way for packages to be able to define replacement functions that do not duplicate in cases where we really don't want them to, but this would require coming up with some sort of protocol, minimally involving an efficient way to detect whether a replacement funciton is being called in a replacement context or directly.
Would "$<-" always satisfy that condition. It would be big help to me if it could be designed to avoid duplication the rest of the data.frame. --
There are some replacement functions that use C code to cheat, but these may create problems if called directly, so I won't advertise them. Best, luke
Cheers, Simon
-- Luke Tierney Statistics and Actuarial Science Ralph E. Wareham Professor of Mathematical Sciences University of Iowa Phone: 319-335-3386 Department of Statistics and Fax: 319-335-3017 Actuarial Science 241 Schaeffer Hall email: luke at stat.uiowa.edu Iowa City, IA 52242 WWW: http://www.stat.uiowa.edu______________________________________________ R-devel at r-project.org mailing list
David Winsemius, MD West Hartford, CT
--
Luke Tierney
Statistics and Actuarial Science
Ralph E. Wareham Professor of Mathematical Sciences
University of Iowa Phone: 319-335-3386
Department of Statistics and Fax: 319-335-3017
Actuarial Science
241 Schaeffer Hall email: luke at stat.uiowa.edu
Iowa City, IA 52242 WWW: http://www.stat.uiowa.edu