shuf: use memory more efficiently when returning a subset

* gl/lib/randperm.c (randperm_new): When the number of items
to return H, is much smaller than the total number of items N,
use a hash to represent the sparse permutations of the set N.
This is currently enabled for N > 128K and N/H > 32.
* tests/misc/shuf: Ensure shuf can quickly return 2 numbers
from a large range.
* gl/modules/randperm: Depend on hash.
* NEWS: Mention the change.

diff --git a/NEWS b/NEWS
index 82ce53c..7a7f761 100644 (file)
--- a/NEWS
+++ b/NEWS
@@ -12,6 +12,11 @@ GNU coreutils NEWS                                    -*- outline -*-
   Note the use of single quotes, not double quotes.
   That creates files named xaa.xz, xab.xz and xac.xz.
 
+** Improvements
+
+  shuf outputs small subsets of large permutations much more efficiently.
+  For example `shuf -i1-$((2**32-1)) -n2` no longer exhausts memory.
+
 
 * Noteworthy changes in release 8.12 (2011-04-26) [stable]
 

diff --git a/gl/lib/randperm.c b/gl/lib/randperm.c
index 97c8d9a..26181b8 100644 (file)
--- a/gl/lib/randperm.c
+++ b/gl/lib/randperm.c
@@ -19,9 +19,11 @@
 
 #include <config.h>
 
+#include "hash.h"
 #include "randperm.h"
 
 #include <limits.h>
+#include <stdlib.h>
 
 #include "xalloc.h"
 
@@ -57,6 +59,94 @@ randperm_bound (size_t h, size_t n)
   return bound;
 }
 
+/* Swap elements I and J in array V.  */
+
+static void
+swap (size_t *v, size_t i, size_t j)
+{
+  size_t t = v[i];
+  v[i] = v[j];
+  v[j] = t;
+}
+
+/* Structures and functions for a sparse_map abstract data type that's
+   used to effectively swap elements I and J in array V like swap(),
+   but in a more memory efficient manner (when the number of permutations
+   performed is significantly less than the size of the input).  */
+
+struct sparse_ent_
+{
+   size_t index;
+   size_t val;
+};
+
+static size_t
+sparse_hash_ (void const *x, size_t table_size)
+{
+  struct sparse_ent_ const *ent = x;
+  return ent->index % table_size;
+}
+
+static bool
+sparse_cmp_ (void const *x, void const *y)
+{
+  struct sparse_ent_ const *ent1 = x;
+  struct sparse_ent_ const *ent2 = y;
+  return ent1->index == ent2->index;
+}
+
+typedef Hash_table sparse_map;
+
+/* Initialize the structure for the sparse map,
+   when a best guess as to the number of entries
+   specified with SIZE_HINT.  */
+
+static sparse_map *
+sparse_new (size_t size_hint)
+{
+  return hash_initialize (size_hint, NULL, sparse_hash_, sparse_cmp_, free);
+}
+
+/* Swap the values for I and J.  If a value is not already present
+   then assume it's equal to the index.  Update the value for
+   index I in array V.  */
+
+static void
+sparse_swap (sparse_map *sv, size_t* v, size_t i, size_t j)
+{
+  struct sparse_ent_ *v1 = hash_delete (sv, &(struct sparse_ent_) {i,0});
+  struct sparse_ent_ *v2 = hash_delete (sv, &(struct sparse_ent_) {j,0});
+
+  /* FIXME: reduce the frequency of these mallocs.  */
+  if (!v1)
+    {
+      v1 = xmalloc (sizeof *v1);
+      v1->index = v1->val = i;
+    }
+  if (!v2)
+    {
+      v2 = xmalloc (sizeof *v2);
+      v2->index = v2->val = j;
+    }
+
+  size_t t = v1->val;
+  v1->val = v2->val;
+  v2->val = t;
+  if (!hash_insert (sv, v1))
+    xalloc_die ();
+  if (!hash_insert (sv, v2))
+    xalloc_die ();
+
+  v[i] = v1->val;
+}
+
+static void
+sparse_free (sparse_map *sv)
+{
+  hash_free (sv);
+}
+
+
 /* From R, allocate and return a malloc'd array of the first H elements
    of a random permutation of N elements.  H must not exceed N.
    Return NULL if H is zero.  */
@@ -79,21 +169,66 @@ randperm_new (struct randint_source *r, size_t h, size_t n)
 
     default:
       {
+        /* The algorithm is essentially the same in both
+           the sparse and non sparse case.  In the sparse case we use
+           a hash to implement sparse storage for the set of n numbers
+           we're shuffling.  When to use the sparse method was
+           determined with the help of this script:
+
+           #!/bin/sh
+           for n in $(seq 2 32); do
+             for h in $(seq 2 32); do
+               test $h -gt $n && continue
+               for s in o n; do
+                 test $s = o && shuf=shuf || shuf=./shuf
+                 num=$(env time -f "$s:${h},${n} = %e,%M" \
+                       $shuf -i0-$((2**$n-2)) -n$((2**$h-2)) | wc -l)
+                 test $num = $((2**$h-2)) || echo "$s:${h},${n} = failed" >&2
+               done
+             done
+           done
+
+           This showed that if sparseness = n/h, then:
+
+           sparseness = 128 => .125 mem used, and about same speed
+           sparseness =  64 => .25  mem used, but 1.5 times slower
+           sparseness =  32 => .5   mem used, but 2 times slower
+
+           Also the memory usage was only significant when n > 128Ki
+        */
+        bool sparse = (n >= (128 * 1024)) && (n / h >= 32);
+
         size_t i;
+        sparse_map *sv;
 
-        v = xnmalloc (n, sizeof *v);
-        for (i = 0; i < n; i++)
-          v[i] = i;
+        if (sparse)
+          {
+            sv = sparse_new (h * 2);
+            if (sv == NULL)
+              xalloc_die ();
+            v = xnmalloc (h, sizeof *v);
+          }
+        else
+          {
+            sv = NULL; /* To placate GCC's -Wuninitialized.  */
+            v = xnmalloc (n, sizeof *v);
+            for (i = 0; i < n; i++)
+              v[i] = i;
+          }
 
         for (i = 0; i < h; i++)
           {
             size_t j = i + randint_choose (r, n - i);
-            size_t t = v[i];
-            v[i] = v[j];
-            v[j] = t;
+            if (sparse)
+              sparse_swap (sv, v, i, j);
+            else
+              swap (v, i, j);
           }
 
-        v = xnrealloc (v, h, sizeof *v);
+        if (sparse)
+          sparse_free (sv);
+        else
+          v = xnrealloc (v, h, sizeof *v);
       }
       break;
     }

diff --git a/gl/modules/randperm b/gl/modules/randperm
index 9cef782..daf9e32 100644 (file)
--- a/gl/modules/randperm
+++ b/gl/modules/randperm
@@ -8,6 +8,7 @@ lib/randperm.h
 Depends-on:
 randint
 xalloc
+hash
 
 configure.ac:
 

diff --git a/tests/misc/shuf b/tests/misc/shuf
index cdfe705..10d2858 100755 (executable)
--- a/tests/misc/shuf
+++ b/tests/misc/shuf
@@ -18,6 +18,7 @@
 
 . "${srcdir=.}/init.sh"; path_prepend_ ../src
 print_ver_ shuf
+getlimits_
 
 seq 100 > in || framework_failure
 
@@ -51,4 +52,9 @@ shuf --zero-terminated -i 1-1 > out || fail=1
 printf '1\0' > exp || framework_failure
 cmp out exp || { fail=1; echo "missing NUL terminator?" 1>&2; }
 
+# Ensure shuf -n operates efficiently for small n. Before coreutils-8.13
+# this would try to allocate $SIZE_MAX * sizeof(size_t)
+timeout 10 shuf -i1-$SIZE_MAX -n2 >/dev/null ||
+  { fail=1; echo "couldn't get a small subset" >&2; }
+
 Exit $fail