#define F_BSIZE(b) stk_fsize(sbuck_size(b))
-static u64
-sorter_clock(void)
+static void
+sorter_start_timer(struct sort_context *ctx)
{
- struct timeval tv;
- gettimeofday(&tv, NULL);
- return (u64)tv.tv_sec * 1000 + tv.tv_usec / 1000;
+ init_timer(&ctx->start_time);
}
static void
-sorter_start_timer(struct sort_context *ctx)
+sorter_stop_timer(struct sort_context *ctx, uns *account_to)
{
- ctx->start_time = sorter_clock();
+ 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)
{
- u64 stop_time = sorter_clock();
if (!size)
return 0;
- if (stop_time <= ctx->start_time)
+ if (!ctx->last_pass_time)
return -1;
- return (uns)((double)size / (1<<20) * 1000 / (stop_time-ctx->start_time));
+ return (uns)((double)size / (1<<20) * 1000 / ctx->last_pass_time);
}
static int
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;
+ if (!(out->flags & SBF_FINAL))
+ return NULL;
+ ASSERT(out->runs == 1);
+ return out;
}
static void
ins[0] = sbuck_new(ctx);
if (!sorter_presort(ctx, b, ins[0], join ? : ins[0]))
{
+ 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]);
+ {
+ ASSERT(join->runs == 2);
+ join->runs--;
+ sbuck_drop(ins[0]);
+ }
else
clist_insert_after(&ins[0]->n, list_pos);
sbuck_drop(b);
while (sorter_presort(ctx, b, ins[i], ins[i]))
i = 1-i;
sbuck_drop(b);
+ 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]),
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]);
outs[1] = sbuck_new(ctx);
outs[2] = NULL;
ctx->twoway_merge(ctx, ins, outs);
+ 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]),
clist_insert_after(&ins[0]->n, list_pos);
}
+static void
+sorter_multiway(struct sort_context *ctx, struct sort_bucket *b)
+{
+ clist parts;
+ cnode *list_pos = b->n.prev;
+ struct sort_bucket *join = sbuck_join_to(b);
+
+ clist_init(&parts);
+ ASSERT(!(sorter_debug & SORT_DEBUG_NO_PRESORT));
+ // FIXME: What if the parts will be too small?
+ SORT_XTRACE(3, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
+ uns cont;
+ uns part_cnt = 0;
+ u64 total_size = 0;
+ sorter_start_timer(ctx);
+ do
+ {
+ struct sort_bucket *p = sbuck_new(ctx);
+ clist_add_tail(&parts, &p->n);
+ cont = sorter_presort(ctx, b, p, (!part_cnt && join) ? join : p);
+ part_cnt++;
+ total_size += sbuck_size(p);
+ }
+ while (cont);
+ sorter_stop_timer(ctx, &ctx->total_pre_time);
+
+ // FIXME: This is way too similar to the two-way case.
+ if (part_cnt == 1)
+ {
+ struct sort_bucket *p = clist_head(&parts);
+ SORT_XTRACE(((b->flags & SBF_SOURCE) ? 1 : 2), "Sorted in memory");
+ if (join)
+ {
+ ASSERT(join->runs == 2);
+ join->runs--;
+ sbuck_drop(p);
+ }
+ else
+ clist_insert_after(&p->n, list_pos);
+ sbuck_drop(b);
+ return;
+ }
+
+ SORT_TRACE("Multi-way presorting pass (%d parts, %s, %dMB/s)", part_cnt, stk_fsize(total_size), sorter_speed(ctx, total_size));
+ sbuck_drop(b);
+
+ uns max_ways = 16;
+ struct sort_bucket *ways[max_ways+1];
+ SORT_XTRACE(2, "Starting up to %d-way merge", max_ways);
+ for (;;)
+ {
+ uns n = 0;
+ struct sort_bucket *p;
+ while (n < max_ways && (p = clist_head(&parts)))
+ {
+ clist_remove(&p->n);
+ ways[n++] = p;
+ }
+ ways[n] = NULL;
+ ASSERT(n > 1);
+
+ struct sort_bucket *out;
+ out = sbuck_new(ctx); // FIXME: No joining so far
+ sorter_start_timer(ctx);
+ ctx->multiway_merge(ctx, ways, out);
+ sorter_stop_timer(ctx, &ctx->total_ext_time);
+
+ for (uns i=0; i<n; i++)
+ sbuck_drop(ways[i]);
+
+ if (clist_empty(&parts))
+ {
+ clist_insert_after(&out->n, list_pos);
+ SORT_TRACE("Multi-way merge completed (%s, %dMB/s)", F_BSIZE(out), sorter_speed(ctx, sbuck_size(out)));
+ return;
+ }
+ else
+ {
+ clist_add_tail(&parts, &out->n);
+ SORT_TRACE("Multi-way merge pass (%d ways, %s, %dMB/s)", n, F_BSIZE(out), sorter_speed(ctx, sbuck_size(out)));
+ }
+ }
+}
+
static uns
sorter_radix_bits(struct sort_context *ctx, struct sort_bucket *b)
{
- if (!b->hash_bits || !ctx->radix_split ||
+ if (!b->hash_bits || b->hash_bits < sorter_min_radix_bits ||
+ !ctx->radix_split ||
(b->flags & SBF_CUSTOM_PRESORT) ||
(sorter_debug & SORT_DEBUG_NO_RADIX))
return 0;
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 *));
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);
clist_init(&ctx->bucket_list);
sorter_prepare_buf(ctx);
- /* FIXME: Remember to test sorting of empty files */
-
// Create bucket containing the source
struct sort_bucket *bin = sbuck_new(ctx);
bin->flags = SBF_SOURCE | SBF_OPEN_READ;
else
bin->fb = ctx->in_fb;
bin->ident = "in";
- bin->size = ctx->in_size; /* FIXME: Sizes should be either sh_off_t or u64, not both; beware of ~0U */
+ 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));
+ SORT_XTRACE(2, "Input size: %s, %d hash bits", F_BSIZE(bin), bin->hash_bits);
// Create bucket for the output
struct sort_bucket *bout = sbuck_new(ctx);
sbuck_drop(b);
else if (b->runs == 1)
sorter_join(b);
+ else if (ctx->multiway_merge && !(sorter_debug & (SORT_DEBUG_NO_MULTIWAY | SORT_DEBUG_NO_PRESORT))) // FIXME: Just kidding...
+ sorter_multiway(ctx, b);
else if (bits = sorter_radix_bits(ctx, b))
sorter_radix(ctx, b, bits);
else
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);
}
projects / libucw.git / blobdiff