Really deallocate the big_buf when radix-splitting.

[libucw.git] / lib / sorter / govern.c

diff --git a/lib/sorter/govern.c b/lib/sorter/govern.c
index 98b9e9d45da0486c0745cabf9d2074f8833114a3..2a8bea0b27da7330c18f7b4f65d91fadcb96bc53 100644 (file)
--- a/lib/sorter/govern.c
+++ b/lib/sorter/govern.c
@@ -10,12 +10,15 @@
 #include "lib/lib.h"
 #include "lib/fastbuf.h"
 #include "lib/mempool.h"
+#include "lib/stkstring.h"
 #include "lib/sorter/common.h"
 
 #include <string.h>
 #include <sys/time.h>
 #include <time.h>
 
+#define F_BSIZE(b) stk_fsize(sbuck_size(b))
+
 static u64
 sorter_clock(void)
 {
@@ -98,12 +101,12 @@ sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
 
   if (!(sorter_debug & SORT_DEBUG_NO_PRESORT) || (b->flags & SBF_CUSTOM_PRESORT))
     {
-      SORT_XTRACE(2, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
+      SORT_XTRACE(3, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
       sorter_start_timer(ctx);
       ins[0] = sbuck_new(ctx);
       if (!sorter_presort(ctx, b, ins[0], join ? : ins[0]))
        {
-         SORT_TRACE("Sorted in memory");
+         SORT_XTRACE(((b->flags & SBF_SOURCE) ? 1 : 2), "Sorted in memory");
          if (join)
            sbuck_drop(ins[0]);
          else
@@ -128,7 +131,7 @@ sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
       ins[0] = b;
     }
 
-  SORT_XTRACE(2, "Main sorting");
+  SORT_XTRACE(3, "Main sorting");
   uns pass = 0;
   do {
     ++pass;
@@ -136,12 +139,14 @@ sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
     if (ins[0]->runs == 1 && ins[1]->runs == 1 && join)
       {
        // This is guaranteed to produce a single run, so join if possible
+       sh_off_t join_size = sbuck_size(join);
        outs[0] = join;
        outs[1] = NULL;
        ctx->twoway_merge(ctx, ins, outs);
-       ASSERT(outs[0]->runs == 2);
-       outs[0]->runs--;
-       SORT_TRACE("Mergesort pass %d (final run, %s, %dMB/s)", pass, F_BSIZE(outs[0]), sorter_speed(ctx, sbuck_size(outs[0])));
+       ASSERT(join->runs == 2);
+       join->runs--;
+       join_size = sbuck_size(join) - join_size;
+       SORT_TRACE("Mergesort pass %d (final run, %s, %dMB/s)", pass, stk_fsize(join_size), sorter_speed(ctx, join_size));
        sbuck_drop(ins[0]);
        sbuck_drop(ins[1]);
        return;
@@ -163,13 +168,66 @@ sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
   clist_insert_after(&ins[0]->n, list_pos);
 }
 
+static uns
+sorter_radix_bits(struct sort_context *ctx, struct sort_bucket *b)
+{
+  if (!b->hash_bits || !ctx->radix_split ||
+      (b->flags & SBF_CUSTOM_PRESORT) ||
+      (sorter_debug & SORT_DEBUG_NO_RADIX))
+    return 0;
+
+  u64 in = sbuck_size(b);
+  u64 mem = ctx->internal_estimate(ctx, b) * 0.8;      // FIXME: Magical factor for hash non-uniformity
+  if (in <= mem)
+    return 0;
+
+  uns n = sorter_min_radix_bits;
+  while (n < sorter_max_radix_bits && n < b->hash_bits && (in >> n) > mem)
+    n++;
+  return n;
+}
+
+static void
+sorter_radix(struct sort_context *ctx, struct sort_bucket *b, uns bits)
+{
+  uns nbuck = 1 << bits;
+  SORT_XTRACE(2, "Running radix split on %s with hash %d bits of %d (expecting %s buckets)",
+             F_BSIZE(b), bits, b->hash_bits, stk_fsize(sbuck_size(b) / nbuck));
+  sorter_free_buf(ctx);
+  sorter_start_timer(ctx);
+
+  struct sort_bucket **outs = alloca(nbuck * sizeof(struct sort_bucket *));
+  for (uns i=nbuck; i--; )
+    {
+      outs[i] = sbuck_new(ctx);
+      outs[i]->hash_bits = b->hash_bits - bits;
+      clist_insert_after(&outs[i]->n, &b->n);
+    }
+
+  ctx->radix_split(ctx, b, outs, b->hash_bits - bits, bits);
+
+  u64 min = ~(u64)0, max = 0, sum = 0;
+  for (uns i=0; i<nbuck; i++)
+    {
+      u64 s = sbuck_size(outs[i]);
+      min = MIN(min, s);
+      max = MAX(max, s);
+      sum += s;
+      if (nbuck > 4)
+       sbuck_swap_out(outs[i]);
+    }
+
+  SORT_TRACE("Radix split (%d buckets, %s min, %s max, %s avg, %dMB/s)", nbuck,
+            stk_fsize(min), stk_fsize(max), stk_fsize(sum / nbuck), sorter_speed(ctx, sum));
+  sbuck_drop(b);
+}
+
 void
 sorter_run(struct sort_context *ctx)
 {
   ctx->pool = mp_new(4096);
   clist_init(&ctx->bucket_list);
-
-  /* FIXME: Remember to test sorting of empty files */
+  sorter_prepare_buf(ctx);
 
   // Create bucket containing the source
   struct sort_bucket *bin = sbuck_new(ctx);
@@ -182,7 +240,7 @@ sorter_run(struct sort_context *ctx)
   bin->size = ctx->in_size;
   bin->hash_bits = ctx->hash_bits;
   clist_add_tail(&ctx->bucket_list, &bin->n);
-  SORT_XTRACE(2, "Input size: %s", (ctx->in_size == ~(u64)0 ? (byte*)"unknown" : F_BSIZE(bin)));
+  SORT_XTRACE(2, "Input size: %s", F_BSIZE(bin));
 
   // Create bucket for the output
   struct sort_bucket *bout = sbuck_new(ctx);
@@ -194,12 +252,16 @@ sorter_run(struct sort_context *ctx)
   clist_add_head(&ctx->bucket_list, &bout->n);
 
   struct sort_bucket *b;
+  uns bits;
   while (bout = clist_head(&ctx->bucket_list), b = clist_next(&ctx->bucket_list, &bout->n))
     {
+      SORT_XTRACE(2, "Next block: %s, %d hash bits", F_BSIZE(b), b->hash_bits);
       if (!sbuck_have(b))
        sbuck_drop(b);
       else if (b->runs == 1)
        sorter_join(b);
+      else if (bits = sorter_radix_bits(ctx, b))
+       sorter_radix(ctx, b, bits);
       else
        sorter_twoway(ctx, b);
     }