Related
The following question is more related to design, rather than actual coding. I don't know if there's a technical term for such problem, so I'll proceed with an example.
I have some openCL code not optimized at all, and in the Kernel there's essentially a switch statement similar to the following
switch(const) {
case const_a : do_something_a(...); break;
case const_b : do_something_b(....); break;
... //etc
}
I cannot write the actual statement since is quite long. As a simple example consider the following switch statement:
int a;
switch(input):
case 13 : {a = 3; break;}
case 1 : {a = 7; break;}
case 23 : {a = 1; break;}
default : {...}
The question is... would it be better to change such switch with an expression like
a = (input == 13)*3 + (input == 1)*7 + (input == 23)
?
If it's not, is it possible to make it more efficient anyway?
You can assume input only takes values in the set of cases of the switch statement.
You've discovered an interesting question that GPU compilers wrestle with. The general advice is try not to branch. Tricks to make that possible are splitting kernels up (as suggested above) and preprocessor (program-time definitions). Research in GPU algorithm development basically works from this axiom.
Branching all over the place won't get great efficiency because of the inherent divergence (channel = work item within the SIMD thread/warp). Remember that all these channels must execute together. So in a switch where all are taking different paths everyone else goes along for the ride silently waiting for their "case" to execute. Now, if input is always always the same value, it can still be a win.
Another popular option is a table indirection.
kernel void foo(const int *t, ...)
...
a = tbl[input];
This case has a few problems too depending on hardware, inputs, and problem size.
Without more specific context, I can conjure up a case where any of these can run well or poorly.
Switching (or big if-then-else chains).
PROS: If all work items generally take the same path (input is mostly the same value), it's going to be efficient. You could also write an if-then-else chain putting the most common cases first. (On GPUs a switch is not necessarily as easy as an indirect jump since there are multiple work items and they may take different paths.)
CONS: Might generate lots of program code and could blow out the instruction cache. Branching all over the place can get a little costly depending on how many cases need to be evaluated. It might just be better to grind through the compute with the predicated code.
Predicated Code (Your (input == 13)*3 ... code).
PROS: This will probably generate smaller programs and stress the I$ less. (Lookup the OpenCL select function to see a more general approach for your case.)
CONS: We've basically punted and decided to evaluate every "case in the switch". If input is usually the same value, we're wasting time here.
Lookup-table based approaches (my example).
PROS: If the switch you are evaluating has a massive number of cases (branches), but can be indexed by integer you might be ahead to just use a lookup table. On some hardware this means a read from global memory (far far away). Other architectures have a dedicated constant cache, but I understand that a vector lookup will serialize (K cycles for each channel). So it might be only marginally better than the global memory table. However, the code table-lookup generated will be short (I$ friendly) and as the number of branches (case statements) grow this will win in the limit. This approach also deals well with uniform/scattered distributions of input's value.
CONS: The read from global memory (or serialized access from the constant cache) has a big latency even compared to branching. In some cases, to eliminate the extra memory traffic I've seen compilers convert lookup tables into if-then-else/switch chains. It's rare that we have 100 element case statements.
I am now inspired to go study this cutoff. :-)
Appending an element onto a million-element ArrayList has the cost of setting one reference now, and copying one reference in the future when the ArrayList must be resized.
As I understand it, appending an element onto a million-element PersistenVector must create a new path, which consists of 4 arrays of size 32. Which means more than 120 references have to be touched.
How does Clojure manage to keep the vector overhead to "2.5 times worse" or "4 times worse" (as opposed to "60 times worse"), which has been claimed in several Clojure videos I have seen recently? Has it something to do with caching or locality of reference or something I am not aware of?
Or is it somehow possible to build a vector internally with mutation and then turn it immutable before revealing it to the outside world?
I have tagged the question scala as well, since scala.collection.immutable.vector is basically the same thing, right?
Clojure's PersistentVector's have special tail buffer to enable efficient operation at the end of the vector. Only after this 32-element array is filled is it added to the rest of the tree. This keeps the amortized cost low. Here is one article on the implementation. The source is also worth a read.
Regarding, "is it somehow possible to build a vector internally with mutation and then turn it immutable before revealing it to the outside world?", yes! These are known as transients in Clojure, and are used for efficient batch changes.
Cannot tell about Clojure, but I can give some comments about Scala Vectors.
Persistent Scala vectors (scala.collection.immutable.Vectors) are much slower than an array buffer when it comes to appending. In fact, they are 10x slower than the List prepend operation. They are 2x slower than appending to Conc-trees, which we use in Parallel Collections.
But, Scala also has mutable vectors -- they're hidden in the class VectorBuilder. Appending to mutable vectors does not preserve the previous version of the vector, but mutates it in place by keeping the pointer to the rightmost leaf in the vector. So, yes -- keeping the vector mutable internally, and than returning an immutable reference is exactly what's done in Scala collections.
The VectorBuilder is slightly faster than the ArrayBuffer, because it needs to allocate its arrays only once, whereas ArrayBuffer needs to do it twice on average (because of growing). Conc.Buffers, which we use as parallel array combiners, are twice as fast compared to VectorBuilders.
Benchmarks are here. None of the benchmarks involve any boxing, they work with reference objects to avoid any bias:
comparison of Scala List, Vector and Conc
comparison of Scala ArrayBuffer, VectorBuilder and Conc.Buffer
More collections benchmarks here.
These tests were executed using ScalaMeter.
In the article written by Daniel Korzekwa, he said that the performance of following code:
list.map(e => e*2).filter(e => e>10)
is much worse than the iterative solution written using Java.
Can anyone explain why? And what is the best solution for such code in Scala (I hope it's not a Java iterative version which is Scala-fied)?
The reason that particular code is slow is because it's working on primitives but it's using generic operations, so the primitives have to be boxed. (This could be improved if List and its ancestors were specialized.) This will probably slow things down by a factor of 5 or so.
Also, algorithmically, those operations are somewhat expensive, because you make a whole list, and then make a whole new list throwing a few components of the intermediate list away. If you did it in one swoop, then you'd be better off. You could do something like:
list collect (case e if (e*2>10) => e*2)
but what if the calculation e*2 is really expensive? Then you could
(List[Int]() /: list)((ls,e) => { val x = e*2; if (x>10) x :: ls else ls }
except that this would appear backwards. (You could reverse it if need be, but that requires creating a new list, which again isn't ideal algorithmically.)
Of course, you have the same sort of algorithmic problems in Java if you're using a singly linked list--your new list will end up backwards, or you have to create it twice, first in reverse and then forwards, or you have to build it with (non-tail) recursion (which is easy in Scala, but inadvisable for this sort of thing in either language since you'll exhaust the stack), or you have to create a mutable list and then pretend afterwards that it's not mutable. (Which, incidentally, you can do in Scala also--see mutable.LinkedList.)
Basically it's traversing a list twice. Once for multiplying every element with two. And then another time to filter the results.
Think of it as one while loop producing a LinkedList with the intermediate results. And then another loop applying the filter to produce the final results.
This should be faster:
list.view.map(e => e * 2).filter(e => e > 10).force
The solution lies mostly with JVM. Though Scala has a workaround in the figure of #specialization, that increases the size of any specialized class hugely, and only solves half the problem -- the other half being the creation of temporary objects.
The JVM actually does a good job optimizing a lot of it, or the performance would be even more terrible, but Java does not require the optimizations that Scala does, so JVM does not provide them. I expect that to change to some extent with the introduction of SAM not-real-closures in Java.
But, in the end, it comes down to balancing the needs. The same while loop that Java and Scala do so much faster than Scala's function equivalent can be done faster yet in C. Yet, despite what the microbenchmarks tell us, people use Java.
Scala approach is much more abstract and generic. Therefore it is hard to optimize every single case.
I could imagine that HotSpot JIT compiler might apply stream- and loop-fusion to the code in the future if it sees that the immediate results are not used.
Additionally the Java code just does much more.
If you really just want to mutate over a datastructure, consider transform.
It looks a bit like map but doesn't create a new collection, e. g.:
val array = Array(1,2,3,4,5,6,7,8,9,10).transform(_ * 2)
// array is now WrappedArray(2, 4, 6, 8, 10, 12, 14, 16, 18, 20)
I really hope some additional in-place operations will be added ion the future...
To avoid traversing the list twice, I think the for syntax is a nice option here:
val list2 = for(v <- list1; e = v * 2; if e > 10) yield e
Rex Kerr correctly states the major problem: Operating on immutable lists, the stated piece of code creates intermediate lists in memory. Note that this is not necessarily slower than equivalent Java code; you just never use immutable datastructures in Java.
Wilfried Springer has a nice, Scala idomatic solution. Using view, no (manipulated) copies of the whole list are created.
Note that using view might not always be ideal. For example, if your first call is filter that is expected to throw away most of the list, is might be worthwhile to create the shorter version explicitly and use view only after that in order to improve memory locality for later iterations.
list.filter(e => e*2>10).map(e => e*2)
This attempt reduces first the List. So the second traversing is on less elements.
In one of my first attempts to create functional code, I ran into a performance issue.
I started with a common task - multiply the elements of two arrays and sum up the results:
var first:Array[Float] ...
var second:Array[Float] ...
var sum=0f;
for (ix<-0 until first.length)
sum += first(ix) * second(ix);
Here is how I reformed the work:
sum = first.zip(second).map{ case (a,b) => a*b }.reduceLeft(_+_)
When I benchmarked the two approaches, the second method takes 40 times as long to complete!
Why does the second method take so much longer? How can I reform the work to be both speed efficient and use functional programming style?
The main reasons why these two examples are so different in speed are:
the faster one doesn't use any generics, so it doesn't face boxing/unboxing.
the faster one doesn't create temporary collections and, thus, avoids extra memory copies.
Let's consider the slower one by parts. First:
first.zip(second)
That creates a new array, an array of Tuple2. It will copy all elements from both arrays into Tuple2 objects, and then copy a reference to each of these objects into a third array. Now, notice that Tuple2 is parameterized, so it can't store Float directly. Instead, new instances of java.lang.Float are created for each number, the numbers are stored in them, and then a reference for each of them is stored into the Tuple2.
map{ case (a,b) => a*b }
Now a fourth array is created. To compute the values of these elements, it needs to read the reference to the tuple from the third array, read the reference to the java.lang.Float stored in them, read the numbers, multiply, create a new java.lang.Float to store the result, and then pass this reference back, which will be de-referenced again to be stored in the array (arrays are not type-erased).
We are not finished, though. Here's the next part:
reduceLeft(_+_)
That one is relatively harmless, except that it still do boxing/unboxing and java.lang.Float creation at iteration, since reduceLeft receives a Function2, which is parameterized.
Scala 2.8 introduces a feature called specialization which will get rid of a lot of these boxing/unboxing. But let's consider alternative faster versions. We could, for instance, do map and reduceLeft in a single step:
sum = first.zip(second).foldLeft(0f) { case (a, (b, c)) => a + b * c }
We could use view (Scala 2.8) or projection (Scala 2.7) to avoid creating intermediary collections altogether:
sum = first.view.zip(second).map{ case (a,b) => a*b }.reduceLeft(_+_)
This last one doesn't save much, actually, so I think the non-strictness if being "lost" pretty fast (ie, one of these methods is strict even in a view). There's also an alternative way of zipping that is non-strict (ie, avoids some intermediary results) by default:
sum = (first,second).zipped.map{ case (a,b) => a*b }.reduceLeft(_+_)
This gives much better result that the former. Better than the foldLeft one, though not by much. Unfortunately, we can't combined zipped with foldLeft because the former doesn't support the latter.
The last one is the fastest I could get. Faster than that, only with specialization. Now, Function2 happens to be specialized, but for Int, Long and Double. The other primitives were left out, as specialization increases code size rather dramatically for each primitive. On my tests, though Double is actually taking longer. That might be a result of it being twice the size, or it might be something I'm doing wrong.
So, in the end, the problem is a combination of factors, including producing intermediary copies of elements, and the way Java (JVM) handles primitives and generics. A similar code in Haskell using supercompilation would be equal to anything short of assembler. On the JVM, you have to be aware of the trade-offs and be prepared to optimize critical code.
I did some variations of this with Scala 2.8. The loop version is as you write but the
functional version is slightly different:
(xs, ys).zipped map (_ * _) reduceLeft(_ + _)
I ran with Double instead of Float, because currently specialization only kicks in for Double. I then tested with arrays and vectors as the carrier type. Furthermore, I tested Boxed variants which work on java.lang.Double's instead of primitive Doubles to measure
the effect of primitive type boxing and unboxing. Here is what I got (running Java 1.6_10 server VM, Scala 2.8 RC1, 5 runs per test).
loopArray 461 437 436 437 435
reduceArray 6573 6544 6718 6828 6554
loopVector 5877 5773 5775 5791 5657
reduceVector 5064 4880 4844 4828 4926
loopArrayBoxed 2627 2551 2569 2537 2546
reduceArrayBoxed 4809 4434 4496 4434 4365
loopVectorBoxed 7577 7450 7456 7463 7432
reduceVectorBoxed 5116 4903 5006 4957 5122
The first thing to notice is that by far the biggest difference is between primitive array loops and primitive array functional reduce. It's about a factor of 15 instead of the 40 you have seen, which reflects improvements in Scala 2.8 over 2.7. Still, primitive array loops are the fastest of all tests whereas primitive array reduces are the slowest. The reason is that primitive Java arrays and generic operations are just not a very good fit. Accessing elements of primitive Java arrays from generic functions requires a lot of boxing/unboxing and sometimes even requires reflection. Future versions of Scala will specialize the Array class and then we should see some improvement. But right now that's what it is.
If you go from arrays to vectors, you notice several things. First, the reduce version is now faster than the imperative loop! This is because vector reduce can make use of efficient bulk operations. Second, vector reduce is faster than array reduce, which illustrates the inherent overhead that arrays of primitive types pose for generic higher-order functions.
If you eliminate the overhead of boxing/unboxing by working only with boxed java.lang.Double values, the picture changes. Now reduce over arrays is a bit less than 2 times slower than looping, instead of the 15 times difference before. That more closely approximates the inherent overhead of the three loops with intermediate data structures instead of the fused loop of the imperative version. Looping over vectors is now by far the slowest solution, whereas reducing over vectors is a little bit slower than reducing over arrays.
So the overall answer is: it depends. If you have tight loops over arrays of primitive values, nothing beats an imperative loop. And there's no problem writing the loops because they are neither longer nor less comprehensible than the functional versions. In all other situations, the FP solution looks competitive.
This is a microbenchmark, and it depends on how the compiler optimizes you code. You have 3 loops composed here,
zip . map . fold
Now, I'm fairly sure the Scala compiler cannot fuse those three loops into a single loop, and the underlying data type is strict, so each (.) corresponds to an intermediate array being created. The imperative/mutable solution would reuse the buffer each time, avoiding copies.
Now, an understanding of what composing those three functions means is key to understanding performance in a functional programming language -- and indeed, in Haskell, those three loops will be optimized into a single loop that reuses an underlying buffer -- but Scala cannot do that.
There are benefits to sticking to the combinator approach, however -- by distinguishing those three functions, it will be easier to parallelize the code (replace map with parMap etc). In fact, given the right array type, (such as a parallel array) a sufficiently smart compiler will be able to automatically parallelize your code, yielding more performance wins.
So, in summary:
naive translations may have unexpected copies and inefficiences
clever FP compilers remove this overhead (but Scala can't yet)
sticking to the high level approach pays off if you want to retarget your code, e.g. to parallelize it
Don Stewart has a fine answer, but it might not be obvious how going from one loop to three creates a factor of 40 slowdown. I'll add to his answer that Scala compiles to JVM bytecodes, and not only does the Scala compiler not fuse the three loops into one, but the Scala compiler is almost certainly allocating all the intermediate arrays. Notoriously, implementations of the JVM are not designed to handle the allocation rates required by functional languages. Allocation is a significant cost in functional programs, and that's one the loop-fusion transformations that Don Stewart and his colleagues have implemented for Haskell are so powerful: they eliminate lots of allocations. When you don't have those transformations, plus you're using an expensive allocator such as is found on a typical JVM, that's where the big slowdown comes from.
Scala is a great vehicle for experimenting with the expressive power of an unusual mix of language ideas: classes, mixins, modules, functions, and so on. But it's a relatively young research language, and it runs on the JVM, so it's unreasonable to expect great performance except on the kind of code that JVMs are good at. If you want to experiment with the mix of language ideas that Scala offers, great—it's a really interesting design—but don't expect the same performance on pure functional code that you'd get with a mature compiler for a functional language, like GHC or MLton.
Is scala functional programming slower than traditional coding?
Not necessarily. Stuff to do with first-class functions, pattern matching, and currying need not be especially slow. But with Scala, more than with other implementations of other functional languages, you really have to watch out for allocations—they can be very expensive.
The Scala collections library is fully generic, and the operations provided are chosen for maximum capability, not maximum speed. So, yes, if you use a functional paradigm with Scala without paying attention (especially if you are using primitive data types), your code will take longer to run (in most cases) than if you use an imperative/iterative paradigm without paying attention.
That said, you can easily create non-generic functional operations that perform quickly for your desired task. In the case of working with pairs of floats, we might do the following:
class FastFloatOps(a: Array[Float]) {
def fastMapOnto(f: Float => Float) = {
var i = 0
while (i < a.length) { a(i) = f(a(i)); i += 1 }
this
}
def fastMapWith(b: Array[Float])(f: (Float,Float) => Float) = {
val len = a.length min b.length
val c = new Array[Float](len)
var i = 0
while (i < len) { c(i) = f(a(i),b(i)); i += 1 }
c
}
def fastReduce(f: (Float,Float) => Float) = {
if (a.length==0) Float.NaN
else {
var r = a(0)
var i = 1
while (i < a.length) { r = f(r,a(i)); i += 1 }
r
}
}
}
implicit def farray2fastfarray(a: Array[Float]) = new FastFloatOps(a)
and then these operations will be much faster. (Faster still if you use Double and 2.8.RC1, because then the functions (Double,Double)=>Double will be specialized, not generic; if you're using something earlier, you can create your own abstract class F { def f(a: Float) : Float } and then call with new F { def f(a: Float) = a*a } instead of (a: Float) => a*a.)
Anyway, the point is that it's not the functional style that makes functional coding in Scala slow, it's that the library is designed with maximum power/flexibility in mind, not maximum speed. This is sensible, since each person's speed requirements are typically subtly different, so it's hard to cover everyone supremely well. But if it's something you're doing more than just a little, you can write your own stuff where the performance penalty for a functional style is extremely small.
I am not an expert Scala programmer, so there is probably a more efficient method, but what about something like this. This can be tail call optimized, so performance should be OK.
def multiply_and_sum(l1:List[Int], l2:List[Int], sum:Int):Int = {
if (l1 != Nil && l2 != Nil) {
multiply_and_sum(l1.tail, l2.tail, sum + (l1.head * l2.head))
}
else {
sum
}
}
val first = Array(1,2,3,4,5)
val second = Array(6,7,8,9,10)
multiply_and_sum(first.toList, second.toList, 0) //Returns: 130
To answer the question in the title: Simple functional constructs may be slower than imperative on the JVM.
But, if we consider only simple constructs, then we might as well throw out all modern languages and stick with C or assembler. If you look a the programming language shootout, C always wins.
So why choose a modern language? Because it lets you express a cleaner design. Cleaner design leads to performance gains in the overall operation of the application. Even if some low-level methods can be slower. One of my favorite examples is the performance of BuildR vs. Maven. BuildR is written in Ruby, an interpreted, slow, language. Maven is written in Java. A build in BuildR is twice as fast as Maven. This is due mostly to the design of BuildR which is lightweight compared with that of Maven.
Your functional solution is slow because it is generating unnecessary temporary data structures. Removing these is known as deforesting and it is easily done in strict functional languages by rolling your anonymous functions into a single anonymous function and using a single aggregator. For example, your solution written in F# using zip, map and reduce:
let dot xs ys = Array.zip xs ys |> Array.map (fun (x, y) -> x * y) -> Array.reduce ( * )
may be rewritten using fold2 so as to avoid all temporary data structures:
let dot xs ys = Array.fold2 (fun t x y -> t + x * y) 0.0 xs ys
This is a lot faster and the same transformation can be done in Scala and other strict functional languages. In F#, you can also define the fold2 as inline in order to have the higher-order function inlined with its functional argument whereupon you recover the optimal performance of the imperative loop.
Here is dbyrnes solution with arrays (assuming Arrays are to be used) and just iterating over the index:
def multiplyAndSum (l1: Array[Int], l2: Array[Int]) : Int =
{
def productSum (idx: Int, sum: Int) : Int =
if (idx < l1.length)
productSum (idx + 1, sum + (l1(idx) * l2(idx))) else
sum
if (l2.length == l1.length)
productSum (0, 0) else
error ("lengths don't fit " + l1.length + " != " + l2.length)
}
val first = (1 to 500).map (_ * 1.1) toArray
val second = (11 to 510).map (_ * 1.2) toArray
def loopi (n: Int) = (1 to n).foreach (dummy => multiplyAndSum (first, second))
println (timed (loopi (100*1000)))
That needs about 1/40 of the time of the list-approach. I don't have 2.8 installed, so you have to test #tailrec yourself. :)
Does anyone know the answer and/or have an opinion about this?
Since tuples would normally not be very large, I would assume it would make more sense to use structs than classes for these. What say you?
Microsoft made all tuple types reference types in the interests of simplicity.
I personally think this was a mistake. Tuples with more than 4 fields are very unusual and should be replaced with a more typeful alternative anyway (such as a record type in F#) so only small tuples are of practical interest. My own benchmarks showed that unboxed tuples up to 512 bytes could still be faster than boxed tuples.
Although memory efficiency is one concern, I believe the dominant issue is the overhead of the .NET garbage collector. Allocation and collection are very expensive on .NET because its garbage collector has not been very heavily optimized (e.g. compared to the JVM). Moreover, the default .NET GC (workstation) has not yet been parallelized. Consequently, parallel programs that use tuples grind to a halt as all cores contend for the shared garbage collector, destroying scalability. This is not only the dominant concern but, AFAIK, was completely neglected by Microsoft when they examined this problem.
Another concern is virtual dispatch. Reference types support subtypes and, therefore, their members are typically invoked via virtual dispatch. In contrast, value types cannot support subtypes so member invocation is entirely unambiguous and can always be performed as a direct function call. Virtual dispatch is hugely expensive on modern hardware because the CPU cannot predict where the program counter will end up. The JVM goes to great lengths to optimize virtual dispatch but .NET does not. However, .NET does provide an escape from virtual dispatch in the form of value types. So representing tuples as value types could, again, have dramatically improved performance here. For example, calling GetHashCode on a 2-tuple a million times takes 0.17s but calling it on an equivalent struct takes only 0.008s, i.e. the value type is 20× faster than the reference type.
A real situation where these performance problems with tuples commonly arises is in the use of tuples as keys in dictionaries. I actually stumbled upon this thread by following a link from the Stack Overflow question F# runs my algorithm slower than Python! where the author's F# program turned out to be slower than his Python precisely because he was using boxed tuples. Manually unboxing using a hand-written struct type makes his F# program several times faster, and faster than Python. These issues would never had arisen if tuples were represented by value types and not reference types to begin with...
The reason is most likely because only the smaller tuples would make sense as value types since they would have a small memory footprint. The larger tuples (i.e. the ones with more properties) would actually suffer in performance since they would be larger than 16 bytes.
Rather than have some tuples be value types and others be reference types and force developers to know which are which I would imagine the folks at Microsoft thought making them all reference types was simpler.
Ah, suspicions confirmed! Please see Building Tuple:
The first major decision was whether
to treat tuples either as a reference
or value type. Since they are
immutable any time you want to change
the values of a tuple, you have to
create a new one. If they are
reference types, this means there can
be lots of garbage generated if you
are changing elements in a tuple in a
tight loop. F# tuples were reference
types, but there was a feeling from
the team that they could realize a
performance improvement if two, and
perhaps three, element tuples were
value types instead. Some teams that
had created internal tuples had used
value instead of reference types,
because their scenarios were very
sensitive to creating lots of managed
objects. They found that using a value
type gave them better performance. In
our first draft of the tuple
specification, we kept the two-,
three-, and four-element tuples as
value types, with the rest being
reference types. However, during a
design meeting that included
representatives from other languages
it was decided that this "split"
design would be confusing, due to the
slightly different semantics between
the two types. Consistency in behavior
and design was determined to be of
higher priority than potential
performance increases. Based on this
input, we changed the design so that
all tuples are reference types,
although we asked the F# team to do
some performance investigation to see
if it experienced a speedup when using
a value type for some sizes of tuples.
It had a good way to test this, since
its compiler, written in F#, was a
good example of a large program that
used tuples in a variety of scenarios.
In the end, the F# team found that it
did not get a performance improvement
when some tuples were value types
instead of reference types. This made
us feel better about our decision to
use reference types for tuple.
If the .NET System.Tuple<...> types were defined as structs, they would not be scalable. For instance, a ternary tuple of long integers currently scales as follows:
type Tuple3 = System.Tuple<int64, int64, int64>
type Tuple33 = System.Tuple<Tuple3, Tuple3, Tuple3>
sizeof<Tuple3> // Gets 4
sizeof<Tuple33> // Gets 4
If the ternary tuple were defined as a struct, the result would be as follows (based on a test example I implemented):
sizeof<Tuple3> // Would get 32
sizeof<Tuple33> // Would get 104
As tuples have built-in syntax support in F#, and they are used extremely often in this language, "struct" tuples would pose F# programmers at risk of writing inefficient programs without even being aware of it. It would happen so easily:
let t3 = 1L, 2L, 3L
let t33 = t3, t3, t3
In my opinion, "struct" tuples would cause a high probability of creating significant inefficiencies in everyday programming. On the other hand, the currently existing "class" tuples also cause certain inefficiencies, as mentioned by #Jon. However, I think that the product of "occurrence probability" times "potential damage" would be much higher with structs than it currently is with classes. Therefore, the current implementation is the lesser evil.
Ideally, there would be both "class" tuples and "struct" tuples, both with syntactic support in F#!
Edit (2017-10-07)
Struct tuples are now fully supported as follows:
Built into mscorlib (.NET >= 4.7) as System.ValueTuple
Available as NuGet for other versions
Syntactic support in C# >= 7
Syntactic support in F# >= 4.1
For 2-tuples, you can still always use the KeyValuePair<TKey,TValue> from earlier versions of the Common Type System. It's a value type.
A minor clarification to the Matt Ellis article would be that the difference in use semantics between reference and value types is only "slight" when immutability is in effect (which, of course, would be the case here). Nevertheless, I think it would have been best in the BCL design not to introduce the confusion of having Tuple cross over to a reference type at some threshold.
I don't know but if you have ever used F# Tuples are part of the language. If I made a .dll and returned a type of Tuples it be nice to have a type to put that in. I suspect now that F# is part of the language (.Net 4) some modifications to CLR were made to accommodate some common structures in F#
From http://en.wikibooks.org/wiki/F_Sharp_Programming/Tuples_and_Records
let scalarMultiply (s : float) (a, b, c) = (a * s, b * s, c * s);;
val scalarMultiply : float -> float * float * float -> float * float * float
scalarMultiply 5.0 (6.0, 10.0, 20.0);;
val it : float * float * float = (30.0, 50.0, 100.0)