External sort
Sort a huge file too large to fit into memory. The algorithm consists in reading a large file to be sorted in chunks of data small enough to fit in main memory, sort each of the chunks, write them out to a temporary file, and finally combined the smaller subfiles into a single larger file. For more info see: https://en.wikipedia.org/wiki/External_sorting
The sorting algorithm can be any popular sort, like quicksort. For simplicity one can assume that the file consists of fixed length integers and that the sort function is less-than (<).
Go
This is a translation of the C++ code here which implements external sorting using a merge sort. In the interests of brevity, the extensive comments in the C++ version have been largely omitted.
A small test file consisting of random integers has been generated and sorted to demonstrate that the approach works. <lang go>package main
import (
"fmt" "io" "log" "math" "math/rand" "os" "time"
)
type MinHeapNode struct{ element, index int }
type MinHeap struct{ nodes []MinHeapNode }
func left(i int) int {
return (2*i + 1)
}
func right(i int) int {
return (2*i + 2)
}
func newMinHeap(nodes []MinHeapNode) *MinHeap {
mh := new(MinHeap)
mh.nodes = nodes
for i := (len(nodes) - 1) / 2; i >= 0; i-- {
mh.minHeapify(i)
}
return mh
}
func (mh *MinHeap) getMin() MinHeapNode {
return mh.nodes[0]
}
func (mh *MinHeap) replaceMin(x MinHeapNode) {
mh.nodes[0] = x mh.minHeapify(0)
}
func (mh *MinHeap) minHeapify(i int) {
l, r := left(i), right(i)
smallest := i
heapSize := len(mh.nodes)
if l < heapSize && mh.nodes[l].element < mh.nodes[i].element {
smallest = l
}
if r < heapSize && mh.nodes[r].element < mh.nodes[smallest].element {
smallest = r
}
if smallest != i {
mh.nodes[i], mh.nodes[smallest] = mh.nodes[smallest], mh.nodes[i]
mh.minHeapify(smallest)
}
}
func merge(arr []int, l, m, r int) {
n1, n2 := m-l+1, r-m
tl := make([]int, n1)
tr := make([]int, n2)
copy(tl, arr[l:])
copy(tr, arr[m+1:])
i, j, k := 0, 0, l
for i < n1 && j < n2 {
if tl[i] <= tr[j] {
arr[k] = tl[i]
k++
i++
} else {
arr[k] = tr[j]
k++
j++
}
}
for i < n1 {
arr[k] = tl[i]
k++
i++
}
for j < n2 {
arr[k] = tr[j]
k++
j++
}
}
func mergeSort(arr []int, l, r int) {
if l < r {
m := l + (r-l)/2
mergeSort(arr, l, m)
mergeSort(arr, m+1, r)
merge(arr, l, m, r)
}
}
// Merge k sorted files: es0,es1 etc. func mergeFiles(outputFile string, n, k int) {
in := make([]*os.File, k)
var err error
for i := 0; i < k; i++ {
fileName := fmt.Sprintf("es%d", i)
in[i], err = os.Open(fileName)
check(err)
}
out, err := os.Create(outputFile)
check(err)
nodes := make([]MinHeapNode, k)
i := 0
for ; i < k; i++ {
_, err = fmt.Fscanf(in[i], "%d", &nodes[i].element)
if err == io.EOF {
break
}
check(err)
nodes[i].index = i
}
hp := newMinHeap(nodes[:i])
count := 0
for count != i {
root := hp.getMin()
fmt.Fprintf(out, "%d ", root.element)
_, err = fmt.Fscanf(in[root.index], "%d", &root.element)
if err == io.EOF {
root.element = math.MaxInt32
count++
} else {
check(err)
}
hp.replaceMin(root)
}
for j := 0; j < k; j++ {
in[j].Close()
}
out.Close()
}
func check(err error) {
if err != nil {
log.Fatal(err)
}
}
// Create initial runs, divide them evenly amongst the output files // and then merge-sort them. func createInitialRuns(inputFile string, runSize, numWays int) {
in, err := os.Open(inputFile)
out := make([]*os.File, numWays)
for i := 0; i < numWays; i++ {
fileName := fmt.Sprintf("es%d", i) // es0, es1 etc.
out[i], err = os.Create(fileName)
check(err)
}
arr := make([]int, runSize)
moreInput := true
nextOutputFile := 0
var i int
for moreInput {
for i = 0; i < runSize; i++ {
_, err := fmt.Fscanf(in, "%d", &arr[i])
if err == io.EOF {
moreInput = false
break
}
check(err)
}
mergeSort(arr, 0, i-1)
for j := 0; j < i; j++ {
fmt.Fprintf(out[nextOutputFile], "%d ", arr[j])
}
nextOutputFile++
}
for j := 0; j < numWays; j++ {
out[j].Close()
}
in.Close()
}
func externalSort(inputFile, outputFile string, numWays, runSize int) {
createInitialRuns(inputFile, runSize, numWays) mergeFiles(outputFile, runSize, numWays)
}
func main() {
// Create a small test file of 40 random ints and split it into 4 files
// of 10 integers each.
numWays := 4
runSize := 10
inputFile := "input.txt"
outputFile := "output.txt"
in, err := os.Create(inputFile)
check(err)
rand.Seed(time.Now().UnixNano())
for i := 0; i < numWays*runSize; i++ {
fmt.Fprintf(in, "%d ", rand.Intn(math.MaxInt32))
}
in.Close()
externalSort(inputFile, outputFile, numWays, runSize)
// remove temporary files
for i := 0; i < numWays; i++ {
fileName := fmt.Sprintf("es%d", i)
err = os.Remove(fileName)
check(err)
}
}</lang>
- Output:
Contents of input.txt:
921996447 760852351 223421434 1245608832 745990119 1414811249 1947335121 762344474 588429291 993452626 2592794 491133923 1275871423 1152039534 649892156 278215570 595760601 1878223040 1267371451 2097209826 1409628494 1147072290 309824251 108477605 1705270413 1821354697 1703557665 473708588 110138202 1292465428 946557804 148800949 1471244316 1508853596 1306802817 1016358698 1661284048 527644251 546155704 337874167
Contents of output.txt:
2592794 108477605 110138202 148800949 223421434 278215570 309824251 337874167 473708588 491133923 527644251 546155704 588429291 595760601 649892156 745990119 760852351 762344474 921996447 946557804 993452626 1016358698 1147072290 1152039534 1245608832 1267371451 1275871423 1292465428 1306802817 1409628494 1414811249 1471244316 1508853596 1661284048 1703557665 1705270413 1821354697 1878223040 1947335121 2097209826
j
Untested on a memory mapped file. <lang J> NB. Apply an in-place sorting algorithm to a memory mapped file NB. in-place sort is translation of in-place python quicksort.
require 'jmf' JCHAR map_jmf_ 'DATA'; 'file.huge' NB. The noun DATA now refers to the memory mapped file. NB. Use: quicksort DATA
NB. use: quicksort DATA
quicksort=: 3 :'qsinternal 0 , <:@:# ARRAY=: y' NB. ARRAY is global
qsinternal=: 3 :0
'start stop'=. y
if. 0 < stop - start do.
'left right pivot'=. start, stop, start{ARRAY NB. pivot, left, right = array[start], start, stop
while. left <: right do. NB. while left <= right:
while. pivot > left { ARRAY do. NB. while array[left] < pivot:
left=. >: left
end.
while. pivot < right { ARRAY do. NB. while array[right] > pivot:
right=. <: right NB. right -= 1
end.
if. left <: right do. NB. if left <= right:
NB. mapped files work by reference, assignment not required, but for testing.
ARRAY=: (left, right) {`(|.@:[)`]} ARRAY NB. array[left], array[right] = array[right], array[left]
left=. >: left NB. left += 1 right=. <: right NB. right -= 1 end. end. qsinternal start , right NB. _quicksort(array, start, right) qsinternal left , stop NB. _quicksort(array, left, stop) end. i. 0 0 NB. verbs return the final noun
) </lang>
Demonstration the sorting works:
quicksort ?~10 ARRAY 0 1 2 3 4 5 6 7 8 9
Julia
<lang julia>intfile = open("/tmp/mmap.bin", "r+")
arr = Mmap.mmap(intfile, Vector{Int64}, (div(stat(intfile).size, 8))) # Int64 is 8 bytes
sort!(arr) </lang>
Perl
Simulate task by reading from 'DATA' handle and using tiny record limit. As written, works for any numeric input, but could define any kind of customized sorting. <lang perl>use strict; use warnings;
my $max = 4; # records per merge file my(@chunk,@tempf);
sub mysort ($$) { return $_[0] <=> $_[1] }
sub store {
my($a) = @_;
my $f = IO::File->new_tmpfile; # self-deleting after program exit
print $f sort mysort @$a;
seek $f, 0, 0 or warn "Oops: $!";
push(@tempf, { fh => $f, queued => scalar <$f> } );
}
- read input and create sorted temporary files
while () {
push @chunk, $_; store(\@chunk), @chunk = () if @chunk == $max;
} store(\@chunk) if @chunk;
- merge everything
while (1) {
my($lowest) = (sort { mysort($a->{queued}, $b->{queued}); } grep(defined $_->{queued}, @tempf) )[0];
last unless $lowest->{queued};
print $lowest->{queued};
$lowest->{queued} = $lowest->{fh}->getline();
}
__DATA__ 432 345 321 543 987 456 678 123 765 567 876 654 789 234</lang>
- Output:
123 234 321 345 432 456 543 567 654 678 765 789 876 987
Phix
Slight variation on Stream_Merge <lang Phix>include builtins/pqueue.e include builtins/pfile.e -- write_lines() - not [yet] documented
procedure add(integer fn, pq)
object line = gets(fn)
if line=-1 then
close(fn)
else
pq_add({fn,line}, pq)
end if
end procedure
procedure sort_files(sequence filenames)
for i=1 to length(filenames) do
sequence lines = get_text(filenames[i],GT_LF_STRIPPED),
sorted = sort(lines)
printf(1,"%s:%v => %v\n",{filenames[i],lines,sorted})
if write_lines(filenames[i],sorted)!=1 then ?9/0 end if
end for
end procedure
procedure merge_files(integer outfn, sequence filenames)
integer pq = pq_new()
for i=1 to length(filenames) do
add(open(filenames[i], "r"),pq)
end for
while not pq_empty(pq) do
{integer fn, string line} = pq_pop(pq)
puts(outfn,line)
add(fn, pq)
end while
pq_destroy(pq)
end procedure
procedure test()
integer nf = rand(5), -- number of files
lp = 3 -- lines per file
sequence filenames = {},
lines = shuffle(tagset(nf*lp))
for i=1 to nf do
string filename = sprintf("file%d.txt",i)
filenames = append(filenames,filename)
integer fn = open(filename,"w")
for l=1 to lp do
printf(fn,"Line %02d\n",lines[l])
end for
lines = lines[lp+1..$]
close(fn)
end for
printf(1,"sorting %d lines split over %d files\n",{nf*lp,nf})
sort_files(filenames)
integer outfn = 1 -- or open("results.txt","w")
merge_files(outfn,filenames)
-- close(outfn)
for i=1 to nf do
{} = delete_file(filenames[i])
end for
end procedure test()</lang>
- Output:
sorting 9 lines split over 3 files
file1.txt:{"Line 04","Line 01","Line 09"} => {"Line 01","Line 04","Line 09"}
file2.txt:{"Line 06","Line 07","Line 02"} => {"Line 02","Line 06","Line 07"}
file3.txt:{"Line 08","Line 03","Line 05"} => {"Line 03","Line 05","Line 08"}
Line 01
Line 02
Line 03
Line 04
Line 05
Line 06
Line 07
Line 08
Line 09
Python
A technique demonstrated with a short string character data. <lang python>
- ! /usr/bin/python3
$ # example session in bash $ python3 external_sort.py expect 123456789 memory size 1 passed memory size 2 passed memory size 3 passed memory size 4 passed memory size 5 passed memory size 6 passed memory size 7 passed memory size 8 passed memory size 9 passed memory size 10 passed memory size 11 passed
import io
def sort_large_file(n: int, source: open, sink: open, file_opener = open)->None:
approach:
break the source into files of size n
sort each of these files
merge these onto the sink
# store sorted chunks into files of size n
mergers = []
while True:
text = list(source.read(n))
if not len(text):
break;
text.sort()
merge_me = file_opener()
merge_me.write(.join(text))
mergers.append(merge_me)
merge_me.seek(0)
# merge onto sink
stack_tops = [f.read(1) for f in mergers]
while stack_tops:
c = min(stack_tops)
sink.write(c)
i = stack_tops.index(c)
t = mergers[i].read(1)
if t:
stack_tops[i] = t
else:
del stack_tops[i]
mergers[i].close()
del mergers[i] # __del__ method of file_opener should delete the file
def main():
test case
sort 6,7,8,9,2,5,3,4,1 with several memory sizes
# load test case into a file like object
input_file_too_large_for_memory = io.StringIO('678925341')
# generate the expected output
t = list(input_file_too_large_for_memory.read())
t.sort()
expect = .join(t)
print('expect', expect)
# attempt to sort with several memory sizes
for memory_size in range(1,12):
input_file_too_large_for_memory.seek(0)
output_file_too_large_for_memory = io.StringIO()
sort_large_file(memory_size, input_file_too_large_for_memory, output_file_too_large_for_memory, io.StringIO)
output_file_too_large_for_memory.seek(0)
assert(output_file_too_large_for_memory.read() == expect)
print('memory size {} passed'.format(memory_size))
if __name__ == '__main__':
example = main example()
</lang>
Raku
(formerly Perl 6) Borrowing from Stream_Merge here. Temporary files are automatically deleted when program is done, so no explicit clean-up required. <lang perl6>use File::Temp;
sub merge_streams ( @streams ) {
my @s = @streams.map({ hash( STREAM => $_, HEAD => .get ) }).grep({ .<HEAD>.defined });
return gather while @s {
my $h = @s.min: +*.<HEAD>;
take $h<HEAD>;
$h<HEAD> := $h<STREAM>.get
orelse @s .= grep( { $_ !=== $h } );
}
}
sub store (@values) {
my ($filename,$filehandle) = tempfile(:prefix('external-sort.'));
$filehandle.say: join "\n", @values.sort: +*;
$filename
}
- we're going to pretend that this is a long stream of input from stdin...
my (@chunk,@files); for (<43 2 45 32 15 4 3 -9 45 66 0 42 78 123 -11 76 55 87 -2 64 92 34>) {
@chunk.push: $_; @files.push: store(@chunk) and @chunk = () if @chunk.elems == 4; # limit of records per merge file
} @files.push: store(@chunk) if @chunk;
say join ' ', merge_streams @files».&open;</lang>
- Output:
-11 -9 -2 0 2 3 4 15 32 34 42 43 45 45 55 64 66 76 78 87 92 123
REXX
Programming note: the method used to generate the input file is to generate N records,
breaking up the records into sort work files of no more than 10 records (limit).
The sort work files are then sorted with an external sort, and then merged into one big file.
This particular example uses the DOS SORT and DEL commands. <lang rexx>/*REXX pgm reads a file, splits into smaller files, sorts 'em, combines into sorted file*/ parse arg FID n lim seed . /*obtain optional arguments from the CL*/ if FID== | FID=="," then FID= 'SORT_EXT.OUT' /*name of the output (sorted) file. */ if n== | n=="," then n= 500 /*number of records (rand #s) to gen. */ if lim== | lim=="," then lim= 10 /*number of records per SORTWORK file. */ if datatype(seed, 'W') then call random ,,seed /*Numeric? Then use it as a rand seed.*/ sWork = 'SORTWORK.' /*the filename of the SORTWORK files.*/ call gen n,lim /*generate SORTWORK.nnn files. */ call srt # /*sort records in all SORTWORK files.*/ call mrg /*merge records in the SORTWORK files.*/ exit /*stick a fork in it, we're all done. */ /*──────────────────────────────────────────────────────────────────────────────────────*/ mrg: procedure expose FID sWork; parse arg # /*#: the number of SORTWORK files. */
@.= copies('ff'x, 1e5) /*no value should be larger than this. */
call lineout FID, , 1 /*position the output file at rec # 1. */
do j=1 until @.j==@.; call rdr j /*read any number of SORTWORK files, */
end /*j*/ /*but initially just 1 record per file.*/
j=j-1 /*adj. J; read from a non─existent file*/
do forever; y=@.; z=0 /*find the lowest value for N values.*/
do k=1 for j /*traipse through the stemmed @ array.*/
if @.k==@. then call rdr k /*Not defined? Then read a file record*/
if @.k<<y then do; y=@.k; z=k; end /*Lowest so far? Then mark this as min.*/
end /*k*/ /* [↑] note use of << exact comparison*/
if z==0 then leave /*Any more records? No, close file. */
call lineout FID, @.z /*write the value to the output file. */
call rdr z /*re-populate a value from the # file. */
end /*forever*/ /*keep reading/merging until exhausted.*/
call lineout FID /*close the output file (just in case).*/
'DEL' sWORK"*" /*delete all the SORTWORK files. */
return
/*──────────────────────────────────────────────────────────────────────────────────────*/ gen: procedure expose #; parse arg m,siz; d= digits() /*for justify. */
# = 0 /*number of SORTWORK.nnn files so far*/
do j=1 for m; #= 1 + j % siz /*create workfile#*/
call lineout 'SORTWORK.'#, right(random(, 1e5), d) /*write rand #. */
end /*j*/
do k=1 for #; call lineout 'SORTWORK.'#; end /*close a workfile*/
return
/*──────────────────────────────────────────────────────────────────────────────────────*/ rdr: parse arg a; @.a=@.; f=sWork||a; if lines(f)\==0 then @.a= linein(f); return /*──────────────────────────────────────────────────────────────────────────────────────*/ srt: procedure expose sWork; parse arg #
do j=1 for #; fn= sWORK || j; 'SORT' fn "/O" fn; end /*j*/; return</lang>