typo

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

diff --git a/lib/sorter/govern.c b/lib/sorter/govern.c
index d8006e423ea7118f4d46a70d358517133a56a3c1..9f525e60c67c22b695328d873ba43cf0fa711306 100644 (file)
--- a/lib/sorter/govern.c
+++ b/lib/sorter/govern.c
@@ -10,9 +10,37 @@
 #include "lib/lib.h"
 #include "lib/fastbuf.h"
 #include "lib/mempool.h"
 #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 "lib/sorter/common.h"
 
 #include <string.h>

+#include <sys/time.h>
+#include <time.h>
+
+#define F_BSIZE(b) stk_fsize(sbuck_size(b))
+
+static void
+sorter_start_timer(struct sort_context *ctx)
+{
+  init_timer(&ctx->start_time);
+}
+
+static void
+sorter_stop_timer(struct sort_context *ctx, uns *account_to)
+{
+  ctx->last_pass_time = get_timer(&ctx->start_time);
+  *account_to += ctx->last_pass_time;
+}
+
+static uns
+sorter_speed(struct sort_context *ctx, u64 size)
+{
+  if (!size)
+    return 0;
+  if (!ctx->last_pass_time)
+    return -1;
+  return (uns)((double)size / (1<<20) * 1000 / ctx->last_pass_time);
+}

 
 static int
 sorter_presort(struct sort_context *ctx, struct sort_bucket *in, struct sort_bucket *out, struct sort_bucket *out_only)
 
 static int
 sorter_presort(struct sort_context *ctx, struct sort_bucket *in, struct sort_bucket *out, struct sort_bucket *out_only)


@@ -71,11 +99,13 @@ sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
 
   if (!(sorter_debug & SORT_DEBUG_NO_PRESORT) || (b->flags & SBF_CUSTOM_PRESORT))
     {
 
   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]))
        {
       ins[0] = sbuck_new(ctx);
       if (!sorter_presort(ctx, b, ins[0], join ? : ins[0]))
        {

-         SORT_TRACE("Sorted in memory");
+         sorter_stop_timer(ctx, &ctx->total_pre_time);
+         SORT_XTRACE(((b->flags & SBF_SOURCE) ? 1 : 2), "Sorted in memory");

          if (join)
            sbuck_drop(ins[0]);
          else
          if (join)
            sbuck_drop(ins[0]);
          else


@@ -89,7 +119,11 @@ sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
       while (sorter_presort(ctx, b, ins[i], ins[i]))
        i = 1-i;
       sbuck_drop(b);
       while (sorter_presort(ctx, b, ins[i], ins[i]))
        i = 1-i;
       sbuck_drop(b);

-      SORT_TRACE("Presorting pass (%d+%d runs, %s+%s)", ins[0]->runs, ins[1]->runs, F_BSIZE(ins[0]), F_BSIZE(ins[1]));
+      sorter_stop_timer(ctx, &ctx->total_pre_time);
+      SORT_TRACE("Presorting pass (%d+%d runs, %s+%s, %dMB/s)",
+                ins[0]->runs, ins[1]->runs,
+                F_BSIZE(ins[0]), F_BSIZE(ins[1]),
+                sorter_speed(ctx, sbuck_size(ins[0]) + sbuck_size(ins[1])));

     }
   else
     {
     }
   else
     {


@@ -97,19 +131,23 @@ sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
       ins[0] = b;
     }
 
       ins[0] = b;
     }
 

-  SORT_XTRACE(2, "Main sorting");
+  SORT_XTRACE(3, "Main sorting");

   uns pass = 0;
   do {
     ++pass;
   uns pass = 0;
   do {
     ++pass;

+    sorter_start_timer(ctx);

     if (ins[0]->runs == 1 && ins[1]->runs == 1 && join)
       {
        // This is guaranteed to produce a single run, so join if possible
     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);
        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)", pass, F_BSIZE(outs[0]));
+       ASSERT(join->runs == 2);
+       join->runs--;
+       join_size = sbuck_size(join) - join_size;
+       sorter_stop_timer(ctx, &ctx->total_ext_time);
+       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;
        sbuck_drop(ins[0]);
        sbuck_drop(ins[1]);
        return;


@@ -118,7 +156,11 @@ sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
     outs[1] = sbuck_new(ctx);
     outs[2] = NULL;
     ctx->twoway_merge(ctx, ins, outs);
     outs[1] = sbuck_new(ctx);
     outs[2] = NULL;
     ctx->twoway_merge(ctx, ins, outs);

-    SORT_TRACE("Mergesort pass %d (%d+%d runs, %s+%s)", pass, outs[0]->runs, outs[1]->runs, F_BSIZE(outs[0]), F_BSIZE(outs[1]));
+    sorter_stop_timer(ctx, &ctx->total_ext_time);
+    SORT_TRACE("Mergesort pass %d (%d+%d runs, %s+%s, %dMB/s)", pass,
+              outs[0]->runs, outs[1]->runs,
+              F_BSIZE(outs[0]), F_BSIZE(outs[1]),
+              sorter_speed(ctx, sbuck_size(outs[0]) + sbuck_size(outs[1])));

     sbuck_drop(ins[0]);
     sbuck_drop(ins[1]);
     memcpy(ins, outs, 3*sizeof(struct sort_bucket *));
     sbuck_drop(ins[0]);
     sbuck_drop(ins[1]);
     memcpy(ins, outs, 3*sizeof(struct sort_bucket *));


@@ -128,14 +170,67 @@ sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
   clist_insert_after(&ins[0]->n, list_pos);
 }
 
   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]);
+    }
+
+  sorter_stop_timer(ctx, &ctx->total_ext_time);
+  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);
 void
 sorter_run(struct sort_context *ctx)
 {
   ctx->pool = mp_new(4096);
   clist_init(&ctx->bucket_list);

-
-  /* FIXME: There should be a way how to detect size of the input file */
-  /* FIXME: Remember to test sorting of empty files */
+  sorter_prepare_buf(ctx);

 
   // Create bucket containing the source
   struct sort_bucket *bin = sbuck_new(ctx);
 
   // Create bucket containing the source
   struct sort_bucket *bin = sbuck_new(ctx);


@@ -145,9 +240,10 @@ sorter_run(struct sort_context *ctx)
   else
     bin->fb = ctx->in_fb;
   bin->ident = "in";
   else
     bin->fb = ctx->in_fb;
   bin->ident = "in";

-  bin->size = ~(u64)0;
+  bin->size = ctx->in_size;

   bin->hash_bits = ctx->hash_bits;
   clist_add_tail(&ctx->bucket_list, &bin->n);
   bin->hash_bits = ctx->hash_bits;
   clist_add_tail(&ctx->bucket_list, &bin->n);

+  SORT_XTRACE(2, "Input size: %s", F_BSIZE(bin));

 
   // Create bucket for the output
   struct sort_bucket *bout = sbuck_new(ctx);
 
   // Create bucket for the output
   struct sort_bucket *bout = sbuck_new(ctx);


@@ -159,12 +255,16 @@ sorter_run(struct sort_context *ctx)
   clist_add_head(&ctx->bucket_list, &bout->n);
 
   struct sort_bucket *b;
   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))
     {
   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);
       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);
     }
       else
        sorter_twoway(ctx, b);
     }


@@ -172,5 +272,7 @@ sorter_run(struct sort_context *ctx)
   sorter_free_buf(ctx);
   sbuck_write(bout);           // Force empty bucket to a file
   SORT_XTRACE(2, "Final size: %s", F_BSIZE(bout));
   sorter_free_buf(ctx);
   sbuck_write(bout);           // Force empty bucket to a file
   SORT_XTRACE(2, "Final size: %s", F_BSIZE(bout));

+  SORT_XTRACE(2, "Final timings: %.3fs external sorting, %.3fs presorting, %.3fs internal sorting",
+             ctx->total_ext_time/1000., ctx->total_pre_time/1000., ctx->total_int_time/1000.);

   ctx->out_fb = sbuck_read(bout);
 }
   ctx->out_fb = sbuck_read(bout);
 }