Dividing Scala iterators leads to GCoverhead/JavaHeapSpace problems - scala

I am processing large data with Scala, so memory and time is an even more important companion than it usually is to me. I am trying to increase the speed of some evaluation by subdividing the initial Iterator[String] obtained by getLines on a large source file in order to make some subevaluation in parallel and merge the results. I do this by recursively slice-ing the iterator into two halfs and recalling the recursive function on each subiterator.
Now, I am wondering why I get GCoverhead or JavaHeapSpace exception, although the "critical" elements are only evaluated once before the recursion step (in order to get the size of the iterator), but in my opinion not in the recursion step, because slice returns an iterator again (which is non-strict by implementation). The following (reduced!) code will fail applied on a ~15g file before concatenating the sublists.
I use .duplicatein each step. I looked up the api, the doc of .duplicate says "The implementation may allocate temporary storage for elements iterated by one iterator but not yet by the other.", but no element has been iterated yet. Could someone give me a hint what is going wrong there and how to solve this problem? Thank you so much!
type itType = Iterator[String]
def src = io.Source.fromFile(args(0)).getLines
// recursively divide into equal size blocks in divide&conquer fashion
def getSubItsDC(it: itType, depth: Int = 4) = {
println("Getting length of file..")
val totalSize = src.length
println(totalSize)
def rec(it_rec: itType = it, depth_rec: Int = depth, size: Int = totalSize):
List[itType] = depth_rec match {
case n if n > 0 =>
println(n)
val (it1, it2) = it_rec.duplicate
val newSize = size/2
rec(it1 slice (0,newSize), n-1, newSize) ++
rec(it2 slice (newSize,size), n-1, newSize)
case n if n == 0 => List(it_rec)
}
println("Starting recursion..")
rec()
}
getSubItsDC(src)
In the REPL the code runs equally fast with arbitrary size of iterators (when hard coding the totalSize), thus I assumed correct lazyness.

I think you might be better off using the itr grouped size to get an Iterator[Iterator[String]] (a GroupedIterator):
scala> val itr = (1 to 100000000).iterator grouped 1000000
itr: Iterator[Int]#GroupedIterator[Int] = non-empty iterator
This will allow you to chunk the processing of parts of your file.
Why your solution uses too much memory
Duplicating an Iterator is obviously an operation which means that the Iterator may have to cache its computed values. For example:
scala> val itr = (1 to 100000000).iterator
itr: Iterator[Int] = non-empty iterator
scala> itr filter (_ % 10000000 == 0) foreach println
10000000
....
100000000
But when I take a duplicate:
scala> val (a, b) = (1 to 100000000).iterator.duplicate
a: Iterator[Int] = non-empty iterator
b: Iterator[Int] = non-empty iterator
scala> a filter (_ % 10000000 == 0) foreach println
//oh dear, garbage collecting
Exception in thread "main" java.lang.OutOfMemoryError: GC overhead limit exceeded
In this example, as I run through a, in order that b be a duplicate, the elements that a has iterated over but which b has not, need to be cached

Related

How to write an efficient groupBy-size filter in Scala, can be approximate

Given a List[Int] in Scala, I wish to get the Set[Int] of all Ints which appear at least thresh times. I can do this using groupBy or foldLeft, then filter. For example:
val thresh = 3
val myList = List(1,2,3,2,1,4,3,2,1)
myList.foldLeft(Map[Int,Int]()){case(m, i) => m + (i -> (m.getOrElse(i, 0) + 1))}.filter(_._2 >= thresh).keys
will give Set(1,2).
Now suppose the List[Int] is very large. How large it's hard to say but in any case this seems wasteful as I don't care about each of the Ints frequencies, and I only care if they're at least thresh. Once it passed thresh there's no need to check anymore, just add the Int to the Set[Int].
The question is: can I do this more efficiently for a very large List[Int],
a) if I need a true, accurate result (no room for mistakes)
b) if the result can be approximate, e.g. by using some Hashing trick or Bloom Filters, where Set[Int] might include some false-positives, or whether {the frequency of an Int > thresh} isn't really a Boolean but a Double in [0-1].
First of all, you can't do better than O(N), as you need to check each element of your initial array at least once. You current approach is O(N), presuming that operations with IntMap are effectively constant.
Now what you can try in order to increase efficiency:
update map only when current counter value is less or equal to threshold. This will eliminate huge number of most expensive operations — map updates
try faster map instead of IntMap. If you know that values of the initial List are in fixed range, you can use Array instead of IntMap (index as the key). Another possible option will be mutable HashMap with sufficient initail capacity. As my benchmark shows it actually makes significant difference
As #ixx proposed, after incrementing value in the map, check whether it's equal to 3 and in this case add it immediately to result list. This will save you one linear traversing (appears to be not that significant for large input)
I don't see how any approximate solution can be faster (only if you ignore some elements at random). Otherwise it will still be O(N).
Update
I created microbenchmark to measure the actual performance of different implementations. For sufficiently large input and output Ixx's suggestion regarding immediately adding elements to result list doesn't produce significant improvement. However similar approach could be used to eliminate unnecessary Map updates (which appears to be the most expensive operation).
Results of benchmarks (avg run times on 1000000 elems with pre-warming):
Authors solution:
447 ms
Ixx solution:
412 ms
Ixx solution2 (eliminated excessive map writes):
150 ms
My solution:
57 ms
My solution involves using mutable HashMap instead of immutable IntMap and includes all other possible optimizations.
Ixx's updated solution:
val tuple = (Map[Int, Int](), List[Int]())
val res = myList.foldLeft(tuple) {
case ((m, s), i) =>
val count = m.getOrElse(i, 0) + 1
(if (count <= 3) m + (i -> count) else m, if (count == thresh) i :: s else s)
}
My solution:
val map = new mutable.HashMap[Int, Int]()
val res = new ListBuffer[Int]
myList.foreach {
i =>
val c = map.getOrElse(i, 0) + 1
if (c == thresh) {
res += i
}
if (c <= thresh) {
map(i) = c
}
}
The full microbenchmark source is available here.
You could use the foldleft to collect the matching items, like this:
val tuple = (Map[Int,Int](), List[Int]())
myList.foldLeft(tuple) {
case((m, s), i) => {
val count = (m.getOrElse(i, 0) + 1)
(m + (i -> count), if (count == thresh) i :: s else s)
}
}
I could measure a performance improvement of about 40% with a small list, so it's definitely an improvement...
Edited to use List and prepend, which takes constant time (see comments).
If by "more efficiently" you mean the space efficiency (in extreme case when the list is infinite), there's a probabilistic data structure called Count Min Sketch to estimate the frequency of items inside it. Then you can discard those with frequency below your threshold.
There's a Scala implementation from Algebird library.
You can change your foldLeft example a bit using a mutable.Set that is build incrementally and at the same time used as filter for iterating over your Seq by using withFilter. However, because I'm using withFilteri cannot use foldLeft and have to make do with foreach and a mutable map:
import scala.collection.mutable
def getItems[A](in: Seq[A], threshold: Int): Set[A] = {
val counts: mutable.Map[A, Int] = mutable.Map.empty
val result: mutable.Set[A] = mutable.Set.empty
in.withFilter(!result(_)).foreach { x =>
counts.update(x, counts.getOrElse(x, 0) + 1)
if (counts(x) >= threshold) {
result += x
}
}
result.toSet
}
So, this would discard items that have already been added to the result set while running through the Seq the first time, because withFilterfilters the Seqin the appended function (map, flatMap, foreach) rather than returning a filtered Seq.
EDIT:
I changed my solution to not use Seq.count, which was stupid, as Aivean correctly pointed out.
Using Aiveans microbench I can see that it is still slightly slower than his approach, but still better than the authors first approach.
Authors solution
377
Ixx solution:
399
Ixx solution2 (eliminated excessive map writes):
110
Sascha Kolbergs solution:
72
Aivean solution:
54

Scala, Erastothenes: Is there a straightforward way to replace a stream with an iteration?

I wrote a function that generates primes indefinitely (wikipedia: incremental sieve of Erastothenes) usings streams. It returns a stream, but it also merges streams of prime multiples internally to mark upcoming composites. The definition is concise, functional, elegant and easy to understand, if I do say so myself:
def primes(): Stream[Int] = {
def merge(a: Stream[Int], b: Stream[Int]): Stream[Int] = {
def next = a.head min b.head
Stream.cons(next, merge(if (a.head == next) a.tail else a,
if (b.head == next) b.tail else b))
}
def test(n: Int, compositeStream: Stream[Int]): Stream[Int] = {
if (n == compositeStream.head) test(n+1, compositeStream.tail)
else Stream.cons(n, test(n+1, merge(compositeStream, Stream.from(n*n, n))))
}
test(2, Stream.from(4, 2))
}
But, I get a "java.lang.OutOfMemoryError: GC overhead limit exceeded" when I try to generate the 1000th prime.
I have an alternative solution that returns an iterator over primes and uses a priority queue of tuples (multiple, prime used to generate multiple) internally to mark upcoming composites. It works well, but it takes about twice as much code, and I basically had to restart from scratch:
import scala.collection.mutable.PriorityQueue
def primes(): Iterator[Int] = {
// Tuple (composite, prime) is used to generate a primes multiples
object CompositeGeneratorOrdering extends Ordering[(Long, Int)] {
def compare(a: (Long, Int), b: (Long, Int)) = b._1 compare a._1
}
var n = 2;
val composites = PriorityQueue(((n*n).toLong, n))(CompositeGeneratorOrdering)
def advance = {
while (n == composites.head._1) { // n is composite
while (n == composites.head._1) { // duplicate composites
val (multiple, prime) = composites.dequeue
composites.enqueue((multiple + prime, prime))
}
n += 1
}
assert(n < composites.head._1)
val prime = n
n += 1
composites.enqueue((prime.toLong * prime.toLong, prime))
prime
}
Iterator.continually(advance)
}
Is there a straightforward way to translate the code with streams to code with iterators? Or is there a simple way to make my first attempt more memory efficient?
It's easier to think in terms of streams; I'd rather start that way, then tweak my code if necessary.
I guess it's a bug in current Stream implementation.
primes().drop(999).head works fine:
primes().drop(999).head
// Int = 7919
You'll get OutOfMemoryError with stored Stream like this:
val prs = primes()
prs.drop(999).head
// Exception in thread "main" java.lang.OutOfMemoryError: GC overhead limit exceeded
The problem here with class Cons implementation: it contains not only calculated tail, but also a function to calculate this tail. Even when the tail is calculated and function is not needed any more!
In this case functions are extremely heavy, so you'll get OutOfMemoryError even with 1000 functions stored.
We have to drop that functions somehow.
Intuitive fix is failed:
val prs = primes().iterator.toStream
prs.drop(999).head
// Exception in thread "main" java.lang.OutOfMemoryError: GC overhead limit exceeded
With iterator on Stream you'll get StreamIterator, with StreamIterator#toStream you'll get initial heavy Stream.
Workaround
So we have to convert it manually:
def toNewStream[T](i: Iterator[T]): Stream[T] =
if (i.hasNext) Stream.cons(i.next, toNewStream(i))
else Stream.empty
val prs = toNewStream(primes().iterator)
// Stream[Int] = Stream(2, ?)
prs.drop(999).head
// Int = 7919
In your first code, you should postpone the merging until the square of a prime is seen amongst the candidates. This will drastically reduce the number of streams in use, radically improving your memory usage issues. To get the 1000th prime, 7919, we only need to consider primes not above its square root, 88. That's just 23 primes/streams of their multiples, instead of 999 (22, if we ignore the evens from the outset). For the 10,000th prime, it's the difference between having 9999 streams of multiples and just 66. And for the 100,000th, only 189 are needed.
The trick is to separate the primes being consumed from the primes being produced, via a recursive invocation:
def primes(): Stream[Int] = {
def merge(a: Stream[Int], b: Stream[Int]): Stream[Int] = {
def next = a.head min b.head
Stream.cons(next, merge(if (a.head == next) a.tail else a,
if (b.head == next) b.tail else b))
}
def test(n: Int, q: Int,
compositeStream: Stream[Int],
primesStream: Stream[Int]): Stream[Int] = {
if (n == q) test(n+2, primesStream.tail.head*primesStream.tail.head,
merge(compositeStream,
Stream.from(q, 2*primesStream.head).tail),
primesStream.tail)
else if (n == compositeStream.head) test(n+2, q, compositeStream.tail,
primesStream)
else Stream.cons(n, test(n+2, q, compositeStream, primesStream))
}
Stream.cons(2, Stream.cons(3, Stream.cons(5,
test(7, 25, Stream.from(9, 6), primes().tail.tail))))
}
As an added bonus, there's no need to store the squares of primes as Longs. This will also be much faster and have better algorithmic complexity (time and space) as it avoids doing a lot of superfluous work. Ideone testing shows it runs at about ~ n1.5..1.6 empirical orders of growth in producing up to n = 80,000 primes.
There's still an algorithmic problem here: the structure that is created here is still a linear left-leaning structure (((mults_of_2 + mults_of_3) + mults_of_5) + ...), with more frequently-producing streams situated deeper inside it (so the numbers have more levels to percolate through, going up). The right-leaning structure should be better, mults_of_2 + (mults_of_3 + (mults_of_5 + ...)). Making it a tree should bring a real improvement in time complexity (pushing it down typically to about ~ n1.2..1.25). For a related discussion, see this haskellwiki page.
The "real" imperative sieve of Eratosthenes usually runs at around ~ n1.1 (in n primes produced) and an optimal trial division sieve at ~ n1.40..1.45. Your original code runs at about cubic time, or worse. Using imperative mutable array is usually the fastest, working by segments (a.k.a. the segmented sieve of Eratosthenes).
In the context of your second code, this is how it is achieved in Python.
Is there a straightforward way to translate the code with streams to code with iterators? Or is there a simple way to make my first attempt more memory efficient?
#Will Ness has given you an improved answer using Streams and given reasons why your code is taking so much memory and time as in adding streams early and a left-leaning linear structure, but no one has completely answered the second (or perhaps main) part of your question as to can a true incremental Sieve of Eratosthenes be implemented with Iterator's.
First, we should properly credit this right-leaning algorithm of which your first code is a crude (left-leaning) example (since it prematurely adds all prime composite streams to the merge operations), which is due to Richard Bird as in the Epilogue of Melissa E. O'Neill's definitive paper on incremental Sieve's of Eratosthenes.
Second, no, it isn't really possible to substitute Iterator's for Stream's in this algorithm as it depends on moving through a stream without restarting the stream, and although one can access the head of an iterator (the current position), using the next value (skipping over the head) to generate the rest of the iteration as a stream requires building a completely new iterator at a terrible cost in memory and time. However, we can use an Iterator to output the results of the sequence of primes in order to minimize memory use and make it easy to use iterator higher order functions, as you will see in my code below.
Now Will Ness has walked you though the principles of postponing adding prime composite streams to the calculations until they are needed, which works well when one is storing these in a structure such as a Priority Queue or a HashMap and was even missed in the O'Neill paper, but for the Richard Bird algorithm this is not necessary as future stream values will not be accessed until needed so are not stored if the Streams are being properly lazily built (as is lazily and left-leaning). In fact, this algorithm doesn't even need the memorization and overheads of a full Stream as each composite number culling sequence only moves forward without reference to any past primes other than one needs a separate source of the base primes, which can be supplied by a recursive call of the same algorithm.
For ready reference, let's list the Haskell code of the Richard Bird algorithms as follows:
primes = 2:([3..] ‘minus‘ composites)
where
composites = union [multiples p | p <− primes]
multiples n = map (n*) [n..]
(x:xs) ‘minus‘ (y:ys)
| x < y = x:(xs ‘minus‘ (y:ys))
| x == y = xs ‘minus‘ ys
| x > y = (x:xs) ‘minus‘ ys
union = foldr merge []
where
merge (x:xs) ys = x:merge’ xs ys
merge’ (x:xs) (y:ys)
| x < y = x:merge’ xs (y:ys)
| x == y = x:merge’ xs ys
| x > y = y:merge’ (x:xs) ys
In the following code I have simplified the 'minus' function (called "minusStrtAt") as we don't need to build a completely new stream but can incorporate the composite subtraction operation with the generation of the original (in my case odds only) sequence. I have also simplified the "union" function (renaming it as "mrgMltpls")
The stream operations are implemented as a non memoizing generic Co Inductive Stream (CIS) as a generic class where the first field of the class is the value of the current position of the stream and the second is a thunk (a zero argument function that returns the next value of the stream through embedded closure arguments to another function).
def primes(): Iterator[Long] = {
// generic class as a Co Inductive Stream element
class CIS[A](val v: A, val cont: () => CIS[A])
def mltpls(p: Long): CIS[Long] = {
var px2 = p * 2
def nxtmltpl(cmpst: Long): CIS[Long] =
new CIS(cmpst, () => nxtmltpl(cmpst + px2))
nxtmltpl(p * p)
}
def allMltpls(mps: CIS[Long]): CIS[CIS[Long]] =
new CIS(mltpls(mps.v), () => allMltpls(mps.cont()))
def merge(a: CIS[Long], b: CIS[Long]): CIS[Long] =
if (a.v < b.v) new CIS(a.v, () => merge(a.cont(), b))
else if (a.v > b.v) new CIS(b.v, () => merge(a, b.cont()))
else new CIS(b.v, () => merge(a.cont(), b.cont()))
def mrgMltpls(mlps: CIS[CIS[Long]]): CIS[Long] =
new CIS(mlps.v.v, () => merge(mlps.v.cont(), mrgMltpls(mlps.cont())))
def minusStrtAt(n: Long, cmpsts: CIS[Long]): CIS[Long] =
if (n < cmpsts.v) new CIS(n, () => minusStrtAt(n + 2, cmpsts))
else minusStrtAt(n + 2, cmpsts.cont())
// the following are recursive, where cmpsts uses oddPrms and
// oddPrms uses a delayed version of cmpsts in order to avoid a race
// as oddPrms will already have a first value when cmpsts is called to generate the second
def cmpsts(): CIS[Long] = mrgMltpls(allMltpls(oddPrms()))
def oddPrms(): CIS[Long] = new CIS(3, () => minusStrtAt(5L, cmpsts()))
Iterator.iterate(new CIS(2L, () => oddPrms()))
{(cis: CIS[Long]) => cis.cont()}
.map {(cis: CIS[Long]) => cis.v}
}
The above code generates the 100,000th prime (1299709) on ideone in about 1.3 seconds with about a 0.36 second overhead and has an empirical computational complexity to 600,000 primes of about 1.43. The memory use is negligible above that used by the program code.
The above code could be implemented using the built-in Scala Streams, but there is a performance and memory use overhead (of a constant factor) that this algorithm does not require. Using Streams would mean that one could use them directly without the extra Iterator generation code, but as this is used only for final output of the sequence, it doesn't cost much.
To implement some basic tree folding as Will Ness has suggested, one only needs to add a "pairs" function and hook it into the "mrgMltpls" function:
def primes(): Iterator[Long] = {
// generic class as a Co Inductive Stream element
class CIS[A](val v: A, val cont: () => CIS[A])
def mltpls(p: Long): CIS[Long] = {
var px2 = p * 2
def nxtmltpl(cmpst: Long): CIS[Long] =
new CIS(cmpst, () => nxtmltpl(cmpst + px2))
nxtmltpl(p * p)
}
def allMltpls(mps: CIS[Long]): CIS[CIS[Long]] =
new CIS(mltpls(mps.v), () => allMltpls(mps.cont()))
def merge(a: CIS[Long], b: CIS[Long]): CIS[Long] =
if (a.v < b.v) new CIS(a.v, () => merge(a.cont(), b))
else if (a.v > b.v) new CIS(b.v, () => merge(a, b.cont()))
else new CIS(b.v, () => merge(a.cont(), b.cont()))
def pairs(mltplss: CIS[CIS[Long]]): CIS[CIS[Long]] = {
val tl = mltplss.cont()
new CIS(merge(mltplss.v, tl.v), () => pairs(tl.cont()))
}
def mrgMltpls(mlps: CIS[CIS[Long]]): CIS[Long] =
new CIS(mlps.v.v, () => merge(mlps.v.cont(), mrgMltpls(pairs(mlps.cont()))))
def minusStrtAt(n: Long, cmpsts: CIS[Long]): CIS[Long] =
if (n < cmpsts.v) new CIS(n, () => minusStrtAt(n + 2, cmpsts))
else minusStrtAt(n + 2, cmpsts.cont())
// the following are recursive, where cmpsts uses oddPrms and
// oddPrms uses a delayed version of cmpsts in order to avoid a race
// as oddPrms will already have a first value when cmpsts is called to generate the second
def cmpsts(): CIS[Long] = mrgMltpls(allMltpls(oddPrms()))
def oddPrms(): CIS[Long] = new CIS(3, () => minusStrtAt(5L, cmpsts()))
Iterator.iterate(new CIS(2L, () => oddPrms()))
{(cis: CIS[Long]) => cis.cont()}
.map {(cis: CIS[Long]) => cis.v}
}
The above code generates the 100,000th prime (1299709) on ideone in about 0.75 seconds with about a 0.37 second overhead and has an empirical computational complexity to the 1,000,000th prime (15485863) of about 1.09 (5.13 seconds). The memory use is negligible above that used by the program code.
Note that the above codes are completely functional in that there is no mutable state used whatsoever, but that the Bird algorithm (or even the tree folding version) isn't as fast as using a Priority Queue or HashMap for larger ranges as the number of operations to handle the tree merging has a higher computational complexity than the log n overhead of the Priority Queue or the linear (amortized) performance of a HashMap (although there is a large constant factor overhead to handle the hashing so that advantage isn't really seen until some truly large ranges are used).
The reason that these codes use so little memory is that the CIS streams are formulated with no permanent reference to the start of the streams so that the streams are garbage collected as they are used, leaving only the minimal number of base prime composite sequence place holders, which as Will Ness has explained is very small - only 546 base prime composite number streams for generating the first million primes up to 15485863, each placeholder only taking a few 10's of bytes (eight for the Long number, eight for the 64-bit function reference, with another couple of eight bytes for the pointer to the closure arguments and another few bytes for function and class overheads, for a total per stream placeholder of perhaps 40 bytes, or a total of not much more than 20 Kilobytes for generating the sequence for a million primes).
If you just want an infinite stream of primes, this is the most elegant way in my opinion:
def primes = {
def sieve(from : Stream[Int]): Stream[Int] = from.head #:: sieve(from.tail.filter(_ % from.head != 0))
sieve(Stream.from(2))
}

What is the fastest way to subtract two arrays in scala

I have two arrays (that i have pulled out of a matrix (Array[Array[Int]]) and I need to subtract one from the other.
At the moment I am using this method however, when I profile it, it is the bottleneck.
def subRows(a: Array[Int], b: Array[Int], sizeHint: Int): Array[Int] = {
val l: Array[Int] = new Array(sizeHint)
var i = 0
while (i < sizeHint) {
l(i) = a(i) - b(i)
i += 1
}
l
}
I need to do this billions of times so any improvement in speed is a plus.
I have tried using a List instead of an Array to collect the differences and it is MUCH faster but I lose all benefit when I convert it back to an Array.
I did modify the downstream code to take a List to see if that would help but I need to access the contents of the list out of order so again there is loss of any gains there.
It seems like any conversion of one type to another is expensive and I am wondering if there is some way to use a map etc. that might be faster.
Is there a better way?
EDIT
Not sure what I did the first time!?
So the code I used to test it was this:
def subRowsArray(a: Array[Int], b: Array[Int], sizeHint: Int): Array[Int] = {
val l: Array[Int] = new Array(sizeHint)
var i = 0
while (i < sizeHint) {
l(i) = a(i) - b(i)
i += 1
}
l
}
def subRowsList(a: Array[Int], b: Array[Int], sizeHint: Int): List[Int] = {
var l: List[Int] = Nil
var i = 0
while (i < sizeHint) {
l = a(i) - b(i) :: l
i += 1
}
l
}
val a = Array.fill(100, 100)(scala.util.Random.nextInt(2))
val loops = 30000 * 10000
def runArray = for (i <- 1 to loops) subRowsArray(a(scala.util.Random.nextInt(100)), a(scala.util.Random.nextInt(100)), 100)
def runList = for (i <- 1 to loops) subRowsList(a(scala.util.Random.nextInt(100)), a(scala.util.Random.nextInt(100)), 100)
def optTimer(f: => Unit) = {
val s = System.currentTimeMillis
f
System.currentTimeMillis - s
}
The results I thought I got the first time I did this were the exact opposite... I must have misread or mixed up the methods.
My apologies for asking a bad question.
That code is the fastest you can manage single-threaded using a standard JVM. If you think List is faster, you're either fooling yourself or not actually telling us what you're doing. Putting an Int into List requires two object creations: one to create the list element, and one to box the integer. Object creations take about 10x longer than an array access. So it's really not a winning proposition to do it any other way.
If you really, really need to go faster, and must stay with a single thread, you should probably switch to C++ or the like and explicitly use SSE instructions. See this question, for example.
If you really, really need to go faster and can use multiple threads, then the easiest is to package up a chunk of work like this (i.e. a sensible number of pairs of vectors that need to be subtracted--probably at least a few million elements per chunk) into a list as long as the number of processors on your machine, and then call list.par.map(yourSubtractionRoutineThatActsOnTheChunkOfWork).
Finally, if you can be destructive,
a(i) -= b(i)
in the inner loop is, of course, faster. Likewise, if you can reuse space (e.g. with System.arraycopy), you're better off than if you have to keep allocating it. But that changes the interface from what you've shown.
You can use Scalameter to try a benchmark the two implementations which requires at least JRE 7 update 4 and Scala 2.10 to be run. I used scala 2.10 RC2.
Compile with scalac -cp scalameter_2.10-0.2.jar RangeBenchmark.scala.
Run with scala -cp scalameter_2.10-0.2.jar:. RangeBenchmark.
Here's the code I used:
import org.scalameter.api._
object RangeBenchmark extends PerformanceTest.Microbenchmark {
val limit = 100
val a = new Array[Int](limit)
val b = new Array[Int](limit)
val array: Array[Int] = new Array(limit)
var list: List[Int] = Nil
val ranges = for {
size <- Gen.single("size")(limit)
} yield 0 until size
measure method "subRowsArray" in {
using(ranges) curve("Range") in {
var i = 0
while (i < limit) {
array(i) = a(i) - b(i)
i += 1
}
r => array
}
}
measure method "subRowsList" in {
using(ranges) curve("Range") in {
var i = 0
while (i < limit) {
list = a(i) - b(i) :: list
i += 1
}
r => list
}
}
}
Here's the results:
::Benchmark subRowsArray::
Parameters(size -> 100): 8.26E-4
::Benchmark subRowsList::
Parameters(size -> 100): 7.94E-4
You can draw your own conclusions. :)
The stack blew up on larger values of limit. I'll guess it's because it's measuring the performance many times.

Why stream fold operation throws Out of memory exception?

I have following simple code
def fib(i:Long,j:Long):Stream[Long] = i #:: fib(j, i+j)
(0l /: fib(1,1).take(10000000)) (_+_)
And it throws OutOfMemmoryError exception.
I can not understand why, because I think all the parts use constant memmory i.e. lazy evaluation streams and foldLeft...
Those code also don't work
fib(1,1).take(10000000).sum or max, min e.t.c.
How to correctly implement infinite streams and do iterative operations upon it?
Scala version: 2.9.0
Also scala javadoc said, that foldLeft operation is memmory safe for streams
/** Stream specialization of foldLeft which allows GC to collect
* along the way.
*/
#tailrec
override final def foldLeft[B](z: B)(op: (B, A) => B): B = {
if (this.isEmpty) z
else tail.foldLeft(op(z, head))(op)
}
EDIT:
Implementation with iterators still not useful, since it throws ${domainName} exception
def fib(i:Long,j:Long): Iterator[Long] = Iterator(i) ++ fib(j, i + j)
How to define correctly infinite stream/iterator in Scala?
EDIT2:
I don't care about int overflow, I just want to understand how to create infinite stream/iterator etc in scala without side effects .
The reason to use Stream instead of Iterator is so that you don't have to calculate all the small terms in the series over again. But this means that you need to store ten million stream nodes. These are pretty large, unfortunately, so that could be enough to overflow the default memory. The only realistic way to overcome this is to start with more memory (e.g. scala -J-Xmx2G). (Also, note that you're going to overflow Long by an enormous margin; the Fibonacci series increases pretty quickly.)
P.S. The iterator implementation I have in mind is completely different; you don't build it out of concatenated singleton Iterators:
def fib(i: Long, j: Long) = Iterator.iterate((i,j)){ case (a,b) => (b,a+b) }.map(_._1)
Now when you fold, past results can be discarded.
The OutOfMemoryError happens indenpendently from the fact that you use Stream. As Rex Kerr mentioned above, Stream -- unlike Iterator -- stores everything in memory. The difference with List is that the elements of Stream are calculated lazily, but once you reach 10000000, there will be 10000000 elements, just like List.
Try with new Array[Int](10000000), you will have the same problem.
To calculate the fibonacci number as above you may want to use different approach. You can take into account the fact that you only need to have two numbers, instead of the whole fibonacci numbers discovered so far.
For example:
scala> def fib(i:Long,j:Long): Iterator[Long] = Iterator(i) ++ fib(j, i + j)
fib: (i: Long,j: Long)Iterator[Long]
And to get, for example, the index of the first fibonacci number exceeding 1000000:
scala> fib(1, 1).indexWhere(_ > 1000000)
res12: Int = 30
Edit: I added the following lines to cope with the StackOverflow
If you really want to work with 1 millionth fibonacci number, the iterator definition above will not work either for StackOverflowError. The following is the best I have in mind at the moment:
class FibIterator extends Iterator[BigDecimal] {
var i: BigDecimal = 1
var j: BigDecimal = 1
def next = {val temp = i
i = i + j
j = temp
j }
def hasNext = true
}
scala> new FibIterator().take(1000000).foldLeft(0:BigDecimal)(_ + _)
res49: BigDecimal = 82742358764415552005488531917024390424162251704439978804028473661823057748584031
0652444660067860068576582339667553466723534958196114093963106431270812950808725232290398073106383520
9370070837993419439389400053162345760603732435980206131237515815087375786729469542122086546698588361
1918333940290120089979292470743729680266332315132001038214604422938050077278662240891771323175496710
6543809955073045938575199742538064756142664237279428808177636434609546136862690895665103636058513818
5599492335097606599062280930533577747023889877591518250849190138449610994983754112730003192861138966
1418736269315695488126272680440194742866966916767696600932919528743675517065891097024715258730309025
7920682881137637647091134870921415447854373518256370737719553266719856028732647721347048627996967...
#yura's problem:
def fib(i:Long,j:Long):Stream[Long] = i #:: fib(j, i+j)
(0l /: fib(1,1).take(10000000)) (_+_)
besides using a Long which can't possibly hold the Fibonacci of 10,000,000, it does work. That is, if the foldLeft is written as:
fib(1,1).take(10000000).foldLeft(0L)(_+_)
Looking at the Streams.scala source, foldLeft() is clearly designed for Garbage Collection, but /: is not def'd.
The other answers alluded to another problem. The Fibonacci of 10 million is a big number and if BigInt is used, instead of just overflowing like with a Long, absolutely enormous numbers are being added to each over and over again.
Since Stream.foldLeft is optimized for GC it does look like the way to solve for really big Fibonacci numbers, rather than using a zip or tail recursion.
// Fibonacci using BigInt
def fib(i:BigInt,j:BigInt):Stream[BigInt] = i #:: fib(j, i+j)
fib(1,0).take(10000000).foldLeft(BigInt("0"))(_+_)
Results of the above code: 10,000,000 is a 8-figure number. How many figures in fib(10000000)? 2,089,877
fib(1,1).take(10000000) is the "this" of the method /:, it is likely that the JVM will consider the reference alive as long as the method runs, even if in this case, it might get rid of it.
So you keep a reference on the head of the stream all along, hence on the whole stream as you build it to 10M elements.
You could just use recursion, which is about as simple:
def fibSum(terms: Int, i: Long = 1, j: Long = 1, total: Long = 2): Long = {
if (terms == 2) total
else fibSum(terms - 1, j, i + j, total + i + j)
}
With this, you can "fold" a billion elements in only a couple of seconds, but as Rex points out, summing the Fibbonaci sequence overflows Long very quickly.
If you really wanted to know the answer to your original problem and don't mind sacrificing some accuracy you could do this:
def fibSum(terms: Int, i: Double = 1, j: Double = 1, tot: Double = 2,
exp: Int = 0): String = {
if (terms == 2) "%.6f".format(tot) + " E+" + exp
else {
val (i1, j1, tot1, exp1) =
if (tot + i + j > 10) (i/10, j/10, tot/10, exp + 1)
else (i, j, tot, exp)
fibSum(terms - 1, j1, i1 + j1, tot1 + i1 + j1, exp1)
}
}
scala> fibSum(10000000)
res54: String = 2.957945 E+2089876

Incrementing the for loop (loop variable) in scala by power of 5

I had asked this question on Javaranch, but couldn't get a response there. So posting it here as well:
I have this particular requirement where the increment in the loop variable is to be done by multiplying it with 5 after each iteration. In Java we could implement it this way:
for(int i=1;i<100;i=i*5){}
In scala I was trying the following code-
var j=1
for(i<-1.to(100).by(scala.math.pow(5,j).toInt))
{
println(i+" "+j)
j=j+1
}
But its printing the following output:
1 1
6 2
11 3
16 4
21 5
26 6
31 7
36 8
....
....
Its incrementing by 5 always. So how do I got about actually multiplying the increment by 5 instead of adding it.
Let's first explain the problem. This code:
var j=1
for(i<-1.to(100).by(scala.math.pow(5,j).toInt))
{
println(i+" "+j)
j=j+1
}
is equivalent to this:
var j = 1
val range: Range = Predef.intWrapper(1).to(100)
val increment: Int = scala.math.pow(5, j).toInt
val byRange: Range = range.by(increment)
byRange.foreach {
println(i+" "+j)
j=j+1
}
So, by the time you get to mutate j, increment and byRange have already been computed. And Range is an immutable object -- you can't change it. Even if you produced new ranges while you did the foreach, the object doing the foreach would still be the same.
Now, to the solution. Simply put, Range is not adequate for your needs. You want a geometric progression, not an arithmetic one. To me (and pretty much everyone else answering, it seems), the natural solution would be to use a Stream or Iterator created with iterate, which computes the next value based on the previous one.
for(i <- Iterator.iterate(1)(_ * 5) takeWhile (_ < 100)) {
println(i)
}
EDIT: About Stream vs Iterator
Stream and Iterator are very different data structures, that share the property of being non-strict. This property is what enables iterate to even exist, since this method is creating an infinite collection1, from which takeWhile will create a new2 collection which is finite. Let's see here:
val s1 = Stream.iterate(1)(_ * 5) // s1 is infinite
val s2 = s1.takeWhile(_ < 100) // s2 is finite
val i1 = Iterator.iterate(1)(_ * 5) // i1 is infinite
val i2 = i1.takeWhile(_ < 100) // i2 is finite
These infinite collections are possible because the collection is not pre-computed. On a List, all elements inside the list are actually stored somewhere by the time the list has been created. On the above examples, however, only the first element of each collection is known in advance. All others will only be computed if and when required.
As I mentioned, though, these are very different collections in other respects. Stream is an immutable data structure. For instance, you can print the contents of s2 as many times as you wish, and it will show the same output every time. On the other hand, Iterator is a mutable data structure. Once you used a value, that value will be forever gone. Print the contents of i2 twice, and it will be empty the second time around:
scala> s2 foreach println
1
5
25
scala> s2 foreach println
1
5
25
scala> i2 foreach println
1
5
25
scala> i2 foreach println
scala>
Stream, on the other hand, is a lazy collection. Once a value has been computed, it will stay computed, instead of being discarded or recomputed every time. See below one example of that behavior in action:
scala> val s2 = s1.takeWhile(_ < 100) // s2 is finite
s2: scala.collection.immutable.Stream[Int] = Stream(1, ?)
scala> println(s2)
Stream(1, ?)
scala> s2 foreach println
1
5
25
scala> println(s2)
Stream(1, 5, 25)
So Stream can actually fill up the memory if one is not careful, whereas Iterator occupies constant space. On the other hand, one can be surprised by Iterator, because of its side effects.
(1) As a matter of fact, Iterator is not a collection at all, even though it shares a lot of the methods provided by collections. On the other hand, from the problem description you gave, you are not really interested in having a collection of numbers, just in iterating through them.
(2) Actually, though takeWhile will create a new Iterator on Scala 2.8.0, this new iterator will still be linked to the old one, and changes in one have side effects on the other. This is subject to discussion, and they might end up being truly independent in the future.
In a more functional style:
scala> Stream.iterate(1)(i => i * 5).takeWhile(i => i < 100).toList
res0: List[Int] = List(1, 5, 25)
And with more syntactic sugar:
scala> Stream.iterate(1)(_ * 5).takeWhile(_ < 100).toList
res1: List[Int] = List(1, 5, 25)
Maybe a simple while-loop would do?
var i=1;
while (i < 100)
{
println(i);
i*=5;
}
or if you want to also print the number of iterations
var i=1;
var j=1;
while (i < 100)
{
println(j + " : " + i);
i*=5;
j+=1;
}
it seems you guys likes functional so how about a recursive solution?
#tailrec def quints(n:Int): Unit = {
println(n);
if (n*5<100) quints(n*5);
}
Update: Thanks for spotting the error... it should of course be power, not multiply:
Annoyingly, there doesn't seem to be an integer pow function in the standard library!
Try this:
def pow5(i:Int) = math.pow(5,i).toInt
Iterator from 1 map pow5 takeWhile (100>=) toList
Or if you want to use it in-place:
Iterator from 1 map pow5 takeWhile (100>=) foreach {
j => println("number:" + j)
}
and with the indices:
val iter = Iterator from 1 map pow5 takeWhile (100>=)
iter.zipWithIndex foreach { case (j, i) => println(i + " = " + j) }
(0 to 2).map (math.pow (5, _).toInt).zipWithIndex
res25: scala.collection.immutable.IndexedSeq[(Int, Int)] = Vector((1,0), (5,1), (25,2))
produces a Vector, with i,j in reversed order.