#include "lib/lib.h"
#include "lib/fastbuf.h"
#include "lib/mempool.h"
+#include "lib/stkstring.h"
#include "lib/sorter/common.h"
-struct sort_bucket *
-sorter_new_bucket(struct sort_context *ctx)
+#include <string.h>
+#include <sys/time.h>
+#include <time.h>
+
+#define F_BSIZE(b) stk_fsize(sbuck_size(b))
+
+static u64
+sorter_clock(void)
{
- return mp_alloc_zero(ctx->pool, sizeof(struct sort_bucket));
+ struct timeval tv;
+ gettimeofday(&tv, NULL);
+ return (u64)tv.tv_sec * 1000 + tv.tv_usec / 1000;
}
-struct fastbuf *
-sorter_open_read(struct sort_bucket *b)
+static void
+sorter_start_timer(struct sort_context *ctx)
{
- /* FIXME: These functions should handle buckets with no fb and only name. */
- ASSERT(b->fb);
- return b->fb;
+ ctx->start_time = sorter_clock();
}
-struct fastbuf *
-sorter_open_write(struct sort_bucket *b)
+static uns
+sorter_speed(struct sort_context *ctx, u64 size)
{
- if (!b->fb)
- b->fb = bopen_tmp(sorter_stream_bufsize);
- return b->fb;
+ 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));
}
-void
-sorter_close_read(struct sort_bucket *b)
+static int
+sorter_presort(struct sort_context *ctx, struct sort_bucket *in, struct sort_bucket *out, struct sort_bucket *out_only)
{
- if (!b)
- return;
- ASSERT(b->fb);
- bclose(b->fb);
- b->fb = NULL;
+ sorter_alloc_buf(ctx);
+ if (in->flags & SBF_CUSTOM_PRESORT)
+ {
+ 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);
}
-void
-sorter_close_write(struct sort_bucket *b)
+static inline struct sort_bucket *
+sbuck_join_to(struct sort_bucket *b)
{
- if (b->fb)
+ 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;
+}
+
+static void
+sorter_join(struct sort_bucket *b)
+{
+ struct sort_bucket *join = (struct sort_bucket *) b->n.prev;
+ ASSERT(join->flags & SBF_FINAL);
+ ASSERT(b->runs == 1);
+
+ 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
{
- b->size = btell(b->fb);
- brewind(b->fb);
+ 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);
}
- /* FIXME: Remove empty buckets from the list automatically? */
}
-void
-sorter_run(struct sort_context *ctx)
+static void
+sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
{
- ctx->pool = mp_new(4096);
- ASSERT(!ctx->custom_presort);
- ASSERT(!ctx->out_fb);
- clist_init(&ctx->bucket_list);
+ struct sort_bucket *ins[3] = { NULL }, *outs[3] = { NULL };
+ cnode *list_pos = b->n.prev;
+ struct sort_bucket *join = sbuck_join_to(b);
- /* FIXME: There should be a way how to detect size of the input file */
+ if (!(sorter_debug & SORT_DEBUG_NO_PRESORT) || (b->flags & SBF_CUSTOM_PRESORT))
+ {
+ 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;
+ }
- /* Trivial 2-way merge with no presorting (just a testing hack) */
- struct sort_bucket *bin = sorter_new_bucket(ctx);
- bin->flags = SBF_SOURCE;
- bin->fb = ctx->in_fb;
- bin->ident = "src";
- struct sort_bucket *ins[3], *outs[3];
- ins[0] = bin;
- ins[1] = NULL;
+ ins[1] = sbuck_new(ctx);
+ int i = 1;
+ while (sorter_presort(ctx, b, ins[i], ins[i]))
+ i = 1-i;
+ sbuck_drop(b);
+ 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;
+ }
+ SORT_XTRACE(3, "Main sorting");
+ uns pass = 0;
do {
- outs[0] = sorter_new_bucket(ctx);
- outs[1] = sorter_new_bucket(ctx);
+ ++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(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;
- log(L_DEBUG, "Pass...");
ctx->twoway_merge(ctx, ins, outs);
- log(L_DEBUG, "Done (%d+%d runs)", outs[0]->runs, outs[1]->runs);
- sorter_close_write(outs[0]);
- sorter_close_write(outs[1]);
+ 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]->fb);
+ } while (sbuck_have(ins[1]));
+
+ sbuck_drop(ins[1]);
+ 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);
+}
+
+void
+sorter_run(struct sort_context *ctx)
+{
+ ctx->pool = mp_new(4096);
+ clist_init(&ctx->bucket_list);
+ sorter_prepare_buf(ctx);
+
+ // Create bucket containing the source
+ struct sort_bucket *bin = sbuck_new(ctx);
+ 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 = 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;
+ 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;
+ 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);
+ }
- ctx->out_fb = sorter_open_read(ins[0]);
+ sorter_free_buf(ctx);
+ sbuck_write(bout); // Force empty bucket to a file
+ SORT_XTRACE(2, "Final size: %s", F_BSIZE(bout));
+ ctx->out_fb = sbuck_read(bout);
}
projects / libucw.git / blobdiff