#include "lib/lib.h"
#include "lib/fastbuf.h"
#include "lib/mempool.h"
+#include "lib/stkstring.h"
#include "lib/sorter/common.h"
-#include <fcntl.h>
+#include <string.h>
+#include <sys/time.h>
+#include <time.h>
-void *
-sorter_alloc(struct sort_context *ctx, uns size)
-{
- return mp_alloc_zero(ctx->pool, size);
-}
-
-struct sort_bucket *
-sbuck_new(struct sort_context *ctx)
-{
- struct sort_bucket *b = sorter_alloc(ctx, sizeof(struct sort_bucket));
- b->ctx = ctx;
- return b;
-}
-
-void
-sbuck_drop(struct sort_bucket *b)
-{
- if (b)
- {
- ASSERT(!(b->flags & SBF_DESTROYED));
- if (b->n.prev)
- clist_remove(&b->n);
- bclose(b->fb);
- bzero(b, sizeof(*b));
- b->flags = SBF_DESTROYED;
- }
-}
-
-sh_off_t
-sbuck_size(struct sort_bucket *b)
-{
- if ((b->flags & SBF_OPEN_WRITE) && !(b->flags & SBF_SWAPPED_OUT))
- return btell(b->fb);
- else
- return b->size;
-}
-
-int
-sbuck_have(struct sort_bucket *b)
-{
- return b && sbuck_size(b);
-}
+#define F_BSIZE(b) stk_fsize(sbuck_size(b))
-static int
-sbuck_has_file(struct sort_bucket *b)
+static u64
+sorter_clock(void)
{
- return (b->fb || (b->flags & SBF_SWAPPED_OUT));
+ struct timeval tv;
+ gettimeofday(&tv, NULL);
+ return (u64)tv.tv_sec * 1000 + tv.tv_usec / 1000;
}
static void
-sbuck_swap_in(struct sort_bucket *b)
-{
- if (b->flags & SBF_SWAPPED_OUT)
- {
- b->fb = bopen(b->filename, O_RDWR, sorter_stream_bufsize);
- if (b->flags & SBF_OPEN_WRITE)
- bseek(b->fb, 0, SEEK_END);
- bconfig(b->fb, BCONFIG_IS_TEMP_FILE, 1);
- b->flags &= ~SBF_SWAPPED_OUT;
- SORT_XTRACE("Swapped in %s", b->filename);
- }
-}
-
-struct fastbuf *
-sbuck_read(struct sort_bucket *b)
+sorter_start_timer(struct sort_context *ctx)
{
- sbuck_swap_in(b);
- if (b->flags & SBF_OPEN_READ)
- return b->fb;
- else if (b->flags & SBF_OPEN_WRITE)
- {
- b->size = btell(b->fb);
- b->flags = (b->flags & ~SBF_OPEN_WRITE) | SBF_OPEN_READ;
- brewind(b->fb);
- return b->fb;
- }
- else
- ASSERT(0);
+ ctx->start_time = sorter_clock();
}
-struct fastbuf *
-sbuck_write(struct sort_bucket *b)
+static uns
+sorter_speed(struct sort_context *ctx, u64 size)
{
- sbuck_swap_in(b);
- if (b->flags & SBF_OPEN_WRITE)
- ASSERT(b->fb);
- else
- {
- ASSERT(!(b->flags & (SBF_OPEN_READ | SBF_DESTROYED)));
- b->fb = bopen_tmp(sorter_stream_bufsize);
- b->flags |= SBF_OPEN_WRITE;
- b->filename = mp_strdup(b->ctx->pool, b->fb->name);
- }
- 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
-sbuck_swap_out(struct sort_bucket *b)
-{
- if ((b->flags & (SBF_OPEN_READ | SBF_OPEN_WRITE)) && b->fb)
- {
- if (b->flags & SBF_OPEN_WRITE)
- b->size = btell(b->fb);
- bconfig(b->fb, BCONFIG_IS_TEMP_FILE, 0);
- bclose(b->fb);
- b->fb = NULL;
- b->flags |= SBF_SWAPPED_OUT;
- SORT_XTRACE("Swapped out %s", b->filename);
- }
-}
-
-void
-sorter_alloc_buf(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);
-}
-
-void
-sorter_free_buf(struct sort_context *ctx)
-{
- if (!ctx->big_buf)
- return;
- big_free(ctx->big_buf, ctx->big_buf_size);
- ctx->big_buf = NULL;
- SORT_XTRACE("Freed sorting buffer");
-}
-
-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_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);
if (!sbuck_has_file(join))
{
- // The final bucket doesn't have any fb associated yet, so replace
+ // The final bucket doesn't have any file associated yet, so replace
// it with the new bucket.
- SORT_XTRACE("Replaced final bucket");
+ SORT_XTRACE(2, "Replaced final bucket");
b->flags |= SBF_FINAL;
sbuck_drop(join);
}
else
{
- SORT_TRACE("Copying %jd bytes to output file", (uintmax_t) sbuck_size(b));
+ 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);
if (!(sorter_debug & SORT_DEBUG_NO_PRESORT) || (b->flags & SBF_CUSTOM_PRESORT))
{
- SORT_TRACE("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_XTRACE("Sorted in memory");
+ SORT_XTRACE(((b->flags & SBF_SOURCE) ? 1 : 2), "Sorted in memory");
if (join)
sbuck_drop(ins[0]);
else
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_TRACE("Skipped presorting");
+ SORT_XTRACE(2, "Presorting disabled");
ins[0] = b;
}
- SORT_TRACE("Main sorting");
+ SORT_XTRACE(3, "Main sorting");
+ 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
+ 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)");
+ 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[1] = sbuck_new(ctx);
outs[2] = NULL;
ctx->twoway_merge(ctx, ins, outs);
- SORT_TRACE("Pass done (%d+%d runs, %jd+%jd bytes)", outs[0]->runs, outs[1]->runs, (uintmax_t) sbuck_size(outs[0]), (uintmax_t) sbuck_size(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 *));
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);
-
- /* 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);
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;
+ 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);
}
sorter_free_buf(ctx);
- SORT_XTRACE("Final size: %jd", (uintmax_t) sbuck_size(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);
}
projects / libucw.git / blobdiff