Professors sometime throw out provocative statements to ensure intellectual debate. I have had almost 1500+ hits in less than 2 days ( and I am glad I am on wordpress.com , my old beloved server would have crashed))
The remarks from Ross Ihaka, covered before and also at Xian’s blog at
Note most of his remarks are techie- and only a single line refers to Revlution Analytics.
Other senior members of community (read- professors are silent, though brobably some thought may have been ignited behind scenes)
http://xianblog.wordpress.com/2010/09/06/insane/comment-page-4/#comments
Ross Ihaka Says:
September 12, 2010 at 1:23 pm
Since (something like) my name has been taken in vain here, let me
chip in.
I’ve been worried for some time that R isn’t going to provide the base
that we’re going to need for statistical computation in the
future. (It may well be that the future is already upon us.) There
are certainly efficiency problems (speed and memory use), but there
are more fundamental issues too. Some of these were inherited from S
and some are peculiar to R.
One of the worst problems is scoping. Consider the following little
gem.
f =
function() {
if (runif(1) > .5)
x = 10
x
}
The x being returned by this function is randomly local or global.
There are other examples where variables alternate between local and
non-local throughout the body of a function. No sensible language
would allow this. It’s ugly and it makes optimisation really
difficult. This isn’t the only problem, even weirder things happen
because of interactions between scoping and lazy evaluation.
In light of this, I’ve come to the conclusion that rather than
“fixing” R, it would be much more productive to simply start over and
build something better. I think the best you could hope for by fixing
the efficiency problems in R would be to boost performance by a small
multiple, or perhaps as much as an order of magnitude. This probably
isn’t enough to justify the effort (Luke Tierney has been working on R
compilation for over a decade now).
To try to get an idea of how much speedup is possible, a number of us
have been carrying out some experiments to see how much better we
could do with something new. Based on prototyping we’ve been doing at
Auckland, it looks like it should be straightforward to get two orders
of magnitude speedup over R, at least for those computations which are
currently bottle-necked. There are a couple of ways to make this
happen.
First, scalar computations in R are very slow. This in part because
the R interpreter is very slow, but also because there are a no scalar
types. By introducing scalars and using compilation it looks like its
possible to get a speedup by a factor of several hundred for scalar
computations. This is important because it means that many ghastly
uses of array operations and the apply functions could be replaced by
simple loops. The cost of these improvements is that scope
declarations become mandatory and (optional) type declarations are
necessary to help the compiler.
As a side-effect of compilation and the use of type-hinting it should
be possible to eliminate dispatch overhead for certain (sealed)
classes (scalars and arrays in particular). This won’t bring huge
benefits across the board, but it will mean that you won’t have to do
foreign language calls to get efficiency.
A second big problem is that computations on aggregates (data frames
in particular) run at glacial rates. This is entirely down to
unnecessary copying because of the call-by-value semantics.
Preserving call-by-value semantics while eliminating the extra copying
is hard. The best we can probably do is to take a conservative
approach. R already tries to avoid copying where it can, but fails in
an epic fashion. The alternative is to abandon call-by-value and move
to reference semantics. Again, prototyping indicates that several
hundredfold speedup is possible (for data frames in particular).
The changes in semantics mentioned above mean that the new language
will not be R. However, it won’t be all that far from R and it should
be easy to port R code to the new system, perhaps using some form of
automatic translation.
If we’re smart about building the new system, it should be possible to
make use of multi-cores and parallelism. Adding this to the mix might just
make it possible to get a three order-of-magnitude performance boost
with just a fraction of the memory that R uses. I think it’s something
really worth putting some effort into.
I also think one other change is necessary. The license will need to a
better job of protecting work donated to the commons than GPL2 seems
to have done. I’m not willing to have any more of my work purloined by
the likes of Revolution Analytics, so I’ll be looking for better
protection from the license (and being a lot more careful about who I
work with).
The discussion spilled over to Stack Overflow as well
n the past week I’ve been following a discussion where Ross Ihaka wrote (here ):
I’ve been worried for some time that R isn’t going to provide the base that we’re going to need for statistical computation in the future. (It may well be that the future is already upon us.) There are certainly efficiency problems (speed and memory use), but there are more fundamental issues too. Some of these were inherited from S and some are peculiar to R.
He then continued explaining. This discussion started from this post, and was then followed by commentshere, here, here, here, here, here and maybe some more places I don’t know of.
We all know the problem now.
R can be improved substantially in terms of speed.
For some solutions, here are the patches by Radford-
http://www.cs.toronto.edu/~radford/speed-patches-doc
patch-dollar
Speeds up access to lists, pairlists, and environments using the
$ operator. The speedup comes mainly from avoiding the overhead of
calling DispatchOrEval if there are no complexities, from passing
on the field to extract as a symbol, or a name, or both, as available,
and then converting only as necessary, from simplifying and inlining
the pstrmatch procedure, and from not translating string multiple
times.
Relevant timing test script: test-dollar.r
This test shows about a 40% decrease in the time needed to extract
elements of lists and environments.
Changes unrelated to speed improvement:
A small error-reporting bug is fixed, illustrated by the following
output with r52822:
> options(warnPartialMatchDollar=TRUE)
> pl <- pairlist(abc=1,def=2)
> pl$ab
[1] 1
Warning message:
In pl$ab : partial match of 'ab' to ''
Some code is changed at the end of R_subset3_dflt because it seems
to be more correct, as discussed in code comments.
patch-evalList
Speeds up a large number of operations by avoiding allocation of
an extra CONS cell in the procedures for evaluating argument lists.
Relevant timing test scripts: all of them, but will look at test-em.r
On test-em.r, the speedup from this patch is about 5%.
patch-fast-base
Speeds up lookup of symbols defined in the base environment, by
flagging symbols that have a base environment definition recorded
in the global cache. This allows the definition to be retrieved
quickly without looking in the hash table.
Relevant timing test scripts: all of them, but will look at test-em.r
On test-em.r, the speedup from this patch is about 3%.
Issue: This patch uses the "spare" bit for the flag. This bit is
misnamed, since it is already used elsewhere (for closures). It is
possible that one of the "gp" bits should be used instead. The
"gp" bits should really be divided up for faster access, and so that
their present use is apparent in the code.
In case this use of the "spare" bit proves unwise, the patch code is
conditional on FAST_BASE_CACHE_LOOKUP being defined at the start of
envir.r.
patch-fast-spec
Speeds up lookup of function symbols that begin with a character
other than a letter or ".", by allowing fast bypass of non-global
environments that do not contain (and have never contained) symbols
of this sort. Since it is expected that only functions will be
given names of this sort, the check is done only in findFun, though
it could also be done in findVar.
Relevant timing test scripts: all of them, but will look at test-em.r
On test-em.r, the speedup from this patch is about 8%.
Issue: This patch uses the "spare" bit to flag environments known
to not have symbols starting with a special character. See remarks
on patch-fast-base.
In case this use of the "spare" bit proves unwise, the patch code is
conditional on FAST_SPEC_BYPASS being defined at the start of envir.r.
patch-for
Speeds up for loops by not allocating new space for the loop
variable every iteration, unless necessary.
Relevant timing test script: test-for.r
This test shows a speedup of about 5%.
Change unrelated to speed improvement:
Fixes what I consider to be a bug, in which the loop clobbers a
global variable, as demonstrated by the following output with r52822:
> i <- 99
> f <- function () for (i in 1:3) { print(i); if (i==2) rm(i); }
> f()
[1] 1
[1] 2
[1] 3
> print(i)
[1] 3
patch-matprod
Speeds up matrix products, including vector dot products. The
speed issue here is that the R code checks for any NAs, and
does the multiply in the matprod procedure (in array.c) if so,
since BLAS isn't trusted with NAs. If this check takes about
as long as just doing the multiply in matprod, calling a BLAS
routine makes no sense.
Relevant time test script: test-matprod.r
With no external BLAS, this patch speeds up long vector-vector
products by a factor of about six, matrix-vector products by a
factor of about three, and some matrix-matrix products by a
factor of about two.
Issue: The matrix multiply code in matprod using an LDOUBLE
(long double) variable to accumulate sums, for improved accuracy.
On a SPARC system I tested on, operations on long doubles are
vastly slower than on doubles, so that the patch produces a
large slowdown rather than an improvement. This is also an issue
for the "sum" function, which also uses an LDOUBLE to accumulate
the sum. Perhaps an ordinarly double should be used in these
places, or perhaps the configuration script should define LDOUBLE
as double on architectures where long doubles are extraordinarily
slow.
Due to this issue, not defining MATPROD_CAN_BE_DONE_HERE at the
start of array.c will disable this patch.
patch-parens
Speeds up parentheses by making "(" a special operator whose
argument is not evaluated, thereby bypassing the overhead of
evalList. Also slightly speeds up curly brackets by inlining
a function that is stylistically better inline anyway.
Relevant test script: test-parens.r
In the parens part of test-parens.r, the speedup is about 9%.
patch-protect
Speeds up numerous operations by making PROTECT, UNPROTECT, etc.
be mostly macros in the files in src/main. This takes effect
only for files that include Defn.h after defining the symbol
USE_FAST_PROTECT_MACROS. With these macros, code of the form
v = PROTECT(...) must be replaced by PROTECT(v = ...).
Relevant timing test scripts: all of them, but will look at test-em.r
On test-em.r, the speedup from this patch is about 9%.
patch-save-alloc
Speeds up some binary and unary arithmetic operations by, when
possible, using the space holding one of the operands to hold
the result, rather than allocating new space. Though primarily
a speed improvement, for very long vectors avoiding this allocation
could avoid running out of space.
Relevant test script: test-complex-expr.r
On this test, the speedup is about 5% for scalar operands and about
8% for vector operands.
Issues: There are some tricky issues with attributes, but I think
I got them right. This patch relies on NAMED being set correctly
in the rest of the code. In case it isn't, the patch can be disabled
by not defining AVOID_ALLOC_IF_POSSIBLE at the top of arithmetic.c.
patch-square
Speeds up a^2 when a is a long vector by not checking for the
special case of an exponent of 2 over and over again for every
vector element.
Relevant test script: test-square.r
The time for squaring a long vector is reduced in this test by a
factor of more than five.
patch-sum-prod
Speeds up the "sum" and "prod" functions by not checking for NA
when na.rm=FALSE, and other detailed code improvements.
Relevant test script: test-sum-prod.r
For sum, the improvement is about a factor of 2.5 when na.rm=FALSE,
and about 10% when na.rm=TRUE.
Issue: See the discussion of patch-matprod regarding LDOUBLE.
There is no change regarding this issue due to this patch, however.
patch-transpose
Speeds up the transpose operation (the "t" function) from detailed
code improvements.
Relevant test script: test-transpose.r
The improvement for 200x60 matrices is about a factor of two.
There is little or no improvement for long row or column vectors.
patch-vec-arith
Speeds up arithmetic on vectors of the same length, or when on
vector is of length one. This is done with detailed code improvements.
Relevant test script: test-vec-arith.r
On long vectors, the +, -, and * operators are sped up by about
20% when operands are the same length or one operand is of length one.
Rather mysteriously, when the operands are not length one or the
same length, there is about a 20% increase in time required, though
this may be due to some strange C optimizer peculiarity or some
strange cache effect, since the C code for this is the same as before,
with negligible additional overhead getting to it. Regardless, this
case is much less common than equal lengths or length one.
There is little change for the / operator, which is much slower than
+, -, or *.
patch-vec-subset
Speeds up extraction of subsets of vectors or matrices (eg, v[10:20]
or M[1:10,101:110]). This is done with detailed code improvements.
Relevant test script: test-vec-subset.r
There are lots of tests in this script. The most dramatic improvement
is for extracting many rows and columns of a large array, where the
improvement is by about a factor of four. Extracting many rows from
one column of a matrix is sped up by about 30%.
Changes unrelated to speed improvement:
Fixes two latent bugs where the code incorrectly refers to NA_LOGICAL
when NA_INTEGER is appropriate and where LOGICAL and INTEGER types
are treated as interchangeable. These cause no problems at the moment,
but would if representations were changed.
patch-subscript
(Formerly part of patch-vec-subset) This patch also speeds up
extraction, and also replacement, of subsets of vectors or
matrices, but focuses on the creation of the indexes rather than
the copy operations. Often avoids a duplication (see below) and
eliminates a second scan of the subscript vector for zero
subscripts, folding it into a previous scan at no additional cost.
Relevant test script: test-vec-subset.r
Speeds up some operations with scalar or short vector indexes by
about 10%. Speeds up subscripting with a longer vector of
positive indexes by about 20%.
Issues: The current code duplicates a vector of indexes when it
seems unnecessary. Duplication is for two reasons: to handle
the situation where the index vector is itself being modified in
a replace operation, and so that any attributes can be removed, which
is helpful only for string subscripts, given how the routine to handle
them returns information via an attribute. Duplication for the
second reasons can easily be avoided, so I avoided it. The first
reason for duplication is sometimes valid, but can usually be avoided
by first only doing it if the subscript is to be used for replacement
rather than extraction, and second only doing it if the NAMED field
for the subscript isn't zero.
I also removed two layers of procedure call overhead (passing seven
arguments, so not trivial) that seemed to be doing nothing. Probably
it used to do something, but no longer does, but if instead it is
preparation for some future use, then removing it might be a mistake.
Software problems are best solved by writing code or patches in my opinion rather than discussing endlessly
Some other solutions to a BETTERRRR R
1) Complete Code Design Review
2) Version 3 - Tuneup
3) Better Documentation
4) Suing Revolution Analytics for the code - Hand over da code pardner
DECISION STATS 