X-Git-Url: http://mj.ucw.cz/gitweb/?a=blobdiff_plain;f=lib%2Fsorter%2Fgovern.c;h=1211a81cfe55e912d87c1b9fab4bfb875681f4bd;hb=969ce845f3bf35bb7423bac705d46196bb863e72;hp=97942dee37bbc0d7604d3b2a6ed62f0e7fb631ca;hpb=baa9f9a3368c8d318b9711340727f822d8fc8a34;p=libucw.git

diff --git a/lib/sorter/govern.c b/lib/sorter/govern.c
index 97942dee..1211a81c 100644
--- a/lib/sorter/govern.c
+++ b/lib/sorter/govern.c
@@ -10,105 +10,48 @@
 #include "lib/lib.h"
 #include "lib/fastbuf.h"
 #include "lib/mempool.h"
+#include "lib/stkstring.h"
 #include "lib/sorter/common.h"
 
-void *
-sorter_alloc(struct sort_context *ctx, uns size)
-{
-  return mp_alloc_zero(ctx->pool, size);
-}
+#include <string.h>
+#include <sys/time.h>
+#include <time.h>
 
-struct sort_bucket *
-sbuck_new(struct sort_context *ctx)
-{
-  return sorter_alloc(ctx, sizeof(struct sort_bucket));
-}
+#define F_BSIZE(b) stk_fsize(sbuck_size(b))
 
-void
-sbuck_drop(struct sort_bucket *b)
+static u64
+sorter_clock(void)
 {
-  if (b)
-    {
-      if (b->n.prev)
-	clist_remove(&b->n);
-      bclose(b->fb);
-      bzero(b, sizeof(*b));
-    }
+  struct timeval tv;
+  gettimeofday(&tv, NULL);
+  return (u64)tv.tv_sec * 1000 + tv.tv_usec / 1000;
 }
 
-int
-sbuck_can_read(struct sort_bucket *b)
-{
-  return b && b->size;
-}
-
-struct fastbuf *
-sbuck_open_read(struct sort_bucket *b)
-{
-  /* FIXME: These functions should handle buckets with no fb and only name. */
-  ASSERT(b->fb);
-  return b->fb;
-}
-
-struct fastbuf *
-sbuck_open_write(struct sort_bucket *b)
-{
-  if (!b->fb)
-    b->fb = bopen_tmp(sorter_stream_bufsize);
-  return b->fb;
-}
-
-void
-sbuck_close_read(struct sort_bucket *b)
-{
-  if (!b)
-    return;
-  ASSERT(b->fb);
-  bclose(b->fb);
-  b->fb = NULL;
-}
-
-void
-sbuck_close_write(struct sort_bucket *b)
-{
-  if (b->fb)
-    {
-      b->size = btell(b->fb);
-      brewind(b->fb);
-    }
-}
-
-void
-sorter_alloc_buf(struct sort_context *ctx)
+static void
+sorter_start_timer(struct sort_context *ctx)
 {
-  if (ctx->big_buf)
-    return;
-  u64 bs = MAX(sorter_bufsize/2, 1);
-  bs = ALIGN_TO(bs, (u64)CPU_PAGE_SIZE);
-  ctx->big_buf = big_alloc(2*bs);
-  ctx->big_buf_size = 2*bs;
-  ctx->big_buf_half = ((byte*) ctx->big_buf) + bs;
-  ctx->big_buf_half_size = bs;
-  SORT_XTRACE("Allocated sorting buffer (%jd bytes)", (uintmax_t) bs);
+  ctx->start_time = sorter_clock();
 }
 
-void
-sorter_free_buf(struct sort_context *ctx)
+static uns
+sorter_speed(struct sort_context *ctx, u64 size)
 {
-  if (!ctx->big_buf)
-    return;
-  big_free(ctx->big_buf, ctx->big_buf_size);
-  ctx->big_buf = NULL;
-  SORT_XTRACE("Freed sorting buffer");
+  u64 stop_time = sorter_clock();
+  if (!size)
+    return 0;
+  if (stop_time <= ctx->start_time)
+    return -1;
+  return (uns)((double)size / (1<<20) * 1000 / (stop_time-ctx->start_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)
 {
-  /* FIXME: Mode with no presorting (mostly for debugging) */
   sorter_alloc_buf(ctx);
   if (in->flags & SBF_CUSTOM_PRESORT)
     {
-      struct fastbuf *f = sbuck_open_write(out);
+      struct fastbuf *f = sbuck_write(out);
+      out->runs++;
       return ctx->custom_presort(f, ctx->big_buf, ctx->big_buf_size);	// FIXME: out_only optimization?
     }
   return ctx->internal_sort(ctx, in, out, out_only);
@@ -117,6 +60,9 @@ static int sorter_presort(struct sort_context *ctx, struct sort_bucket *in, stru
 static inline struct sort_bucket *
 sbuck_join_to(struct sort_bucket *b)
 {
+  if (sorter_debug & SORT_DEBUG_NO_JOIN)
+    return NULL;
+
   struct sort_bucket *out = (struct sort_bucket *) b->n.prev;	// Such bucket is guaranteed to exist
   return (out->flags & SBF_FINAL) ? out : NULL;
 }
@@ -124,74 +70,154 @@ sbuck_join_to(struct sort_bucket *b)
 static void
 sorter_join(struct sort_bucket *b)
 {
-  struct sort_bucket *join = sbuck_join_to(b);
-  ASSERT(join);
+  struct sort_bucket *join = (struct sort_bucket *) b->n.prev;
+  ASSERT(join->flags & SBF_FINAL);
+  ASSERT(b->runs == 1);
 
-  // FIXME: What if the final bucket doesn't contain any file yet?
-
-  SORT_TRACE("Copying %jd bytes to output file", (uintmax_t) b->size);
-  struct fastbuf *src = sbuck_open_read(b);
-  struct fastbuf *dest = sbuck_open_write(join);
-  bbcopy(src, dest, ~0U);
-  sbuck_drop(b);
+  if (!sbuck_has_file(join))
+    {
+      // The final bucket doesn't have any file associated yet, so replace
+      // it with the new bucket.
+      SORT_XTRACE(2, "Replaced final bucket");
+      b->flags |= SBF_FINAL;
+      sbuck_drop(join);
+    }
+  else
+    {
+      SORT_TRACE("Copying to output file: %s", F_BSIZE(b));
+      struct fastbuf *src = sbuck_read(b);
+      struct fastbuf *dest = sbuck_write(join);
+      bbcopy(src, dest, ~0U);
+      sbuck_drop(b);
+    }
 }
 
 static void
 sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
 {
-  struct sort_bucket *ins[3], *outs[3];
+  struct sort_bucket *ins[3] = { NULL }, *outs[3] = { NULL };
+  cnode *list_pos = b->n.prev;
   struct sort_bucket *join = sbuck_join_to(b);
 
-  SORT_TRACE("Presorting");
-  ins[0] = sbuck_new(ctx);
-  sbuck_open_read(b);
-  if (!sorter_presort(ctx, b, ins[0], join ? : ins[0]))
+  if (!(sorter_debug & SORT_DEBUG_NO_PRESORT) || (b->flags & SBF_CUSTOM_PRESORT))
     {
-      if (join)
-	sbuck_drop(ins[0]);
-      else
-	clist_insert_after(&ins[0]->n, &b->n);
+      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_XTRACE(((b->flags & SBF_SOURCE) ? 1 : 2), "Sorted in memory");
+	  if (join)
+	    sbuck_drop(ins[0]);
+	  else
+	    clist_insert_after(&ins[0]->n, list_pos);
+	  sbuck_drop(b);
+	  return;
+	}
+
+      ins[1] = sbuck_new(ctx);
+      int i = 1;
+      while (sorter_presort(ctx, b, ins[i], ins[i]))
+	i = 1-i;
       sbuck_drop(b);
-      return;
+      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
+    {
+      SORT_XTRACE(2, "Presorting disabled");
+      ins[0] = b;
     }
 
-  ins[1] = sbuck_new(ctx);
-  ins[2] = NULL;
-  int i = 1;
-  while (sorter_presort(ctx, b, ins[i], ins[i]))
-    i = 1-i;
-  sbuck_close_read(b);
-  sbuck_close_write(ins[0]);
-  sbuck_close_write(ins[1]);
-
-  SORT_TRACE("Main sorting");
+  SORT_XTRACE(3, "Main sorting");
+  uns pass = 0;
   do {
-    if (ins[0]->runs == 1 && ins[1]->runs == 1 && join)	// FIXME: Debug switch for disabling joining optimizations
+    ++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
+	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("Pass done (joined final run)");
-	sbuck_drop(b);
+	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;
       }
     outs[0] = sbuck_new(ctx);
     outs[1] = sbuck_new(ctx);
     outs[2] = NULL;
     ctx->twoway_merge(ctx, ins, outs);
-    sbuck_close_write(outs[0]);
-    sbuck_close_write(outs[1]);
-    SORT_TRACE("Pass done (%d+%d runs, %jd+%jd bytes)", outs[0]->runs, outs[1]->runs, (uintmax_t) outs[0]->size, (uintmax_t) outs[1]->size);
+    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 *));
-  } while (ins[1]->size);
+  } while (sbuck_have(ins[1]));
 
   sbuck_drop(ins[1]);
-  clist_insert_after(&ins[0]->n, &b->n);
+  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_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);
 }
 
@@ -200,43 +226,47 @@ 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);
-  bin->flags = SBF_SOURCE;
+  bin->flags = SBF_SOURCE | SBF_OPEN_READ;
   if (ctx->custom_presort)
     bin->flags |= SBF_CUSTOM_PRESORT;
   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);
+  SORT_XTRACE(2, "Input size: %s", F_BSIZE(bin));
 
   // Create bucket for the output
   struct sort_bucket *bout = sbuck_new(ctx);
   bout->flags = SBF_FINAL;
-  bout->fb = ctx->out_fb;
+  if (bout->fb = ctx->out_fb)
+    bout->flags |= SBF_OPEN_WRITE;
   bout->ident = "out";
   bout->runs = 1;
   clist_add_head(&ctx->bucket_list, &bout->n);
 
   struct sort_bucket *b;
-  while (b = clist_next(&ctx->bucket_list, &bout->n))
+  uns bits;
+  while (bout = clist_head(&ctx->bucket_list), b = clist_next(&ctx->bucket_list, &bout->n))
     {
-      if (!b->size)
+      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);
     }
 
   sorter_free_buf(ctx);
-  sbuck_close_write(bout);
-  SORT_XTRACE("Final size: %jd", (uintmax_t) bout->size);
-  ctx->out_fb = sbuck_open_read(bout);
+  sbuck_write(bout);		// Force empty bucket to a file
+  SORT_XTRACE(2, "Final size: %s", F_BSIZE(bout));
+  ctx->out_fb = sbuck_read(bout);
 }