2 * UCW Library -- Universal Sorter: Governing Routines
4 * (c) 2007 Martin Mares <mj@ucw.cz>
6 * This software may be freely distributed and used according to the terms
7 * of the GNU Lesser General Public License.
11 #include <ucw/fastbuf.h>
12 #include <ucw/mempool.h>
13 #include <ucw/stkstring.h>
15 #include <ucw/sorter/common.h>
21 #define F_BSIZE(b) stk_fsize(sbuck_size(b))
24 sorter_start_timer(struct sort_context *ctx)
26 init_timer(&ctx->start_time);
30 sorter_stop_timer(struct sort_context *ctx, uint *account_to)
32 ctx->last_pass_time = get_timer(&ctx->start_time);
33 *account_to += ctx->last_pass_time;
37 sorter_speed(struct sort_context *ctx, u64 size)
41 if (!ctx->last_pass_time)
43 return (uint)((double)size / (1<<20) * 1000 / ctx->last_pass_time);
47 sorter_presort(struct sort_context *ctx, struct sort_bucket *in, struct sort_bucket *out, struct sort_bucket *out_only)
49 sorter_alloc_buf(ctx);
50 if (in->flags & SBF_CUSTOM_PRESORT)
53 * The trick with automatic joining, which we use for the normal presorter,
54 * is not necessary with the custom presorter, because the custom presorter
55 * is never called in the middle of the sorted data.
57 struct fastbuf *f = sbuck_write(out);
59 return ctx->custom_presort(f, ctx->big_buf, ctx->big_buf_size);
61 return ctx->internal_sort(ctx, in, out, out_only);
64 static struct sort_bucket *
65 sbuck_join_to(struct sort_bucket *b, ucw_off_t *sizep)
67 if (sorter_debug & SORT_DEBUG_NO_JOIN)
70 struct sort_bucket *out = (struct sort_bucket *) b->n.prev; // Such bucket is guaranteed to exist
71 if (!(out->flags & SBF_FINAL))
73 ASSERT(out->runs == 1);
74 *sizep = sbuck_size(out);
79 sbuck_ins_or_join(struct sort_bucket *b, cnode *list_pos, struct sort_bucket *join, ucw_off_t join_size)
81 if (join && join->runs >= 2)
85 ASSERT(join->runs == 2);
87 return sbuck_size(join) - join_size;
91 clist_insert_after(&b->n, list_pos);
99 sorter_join(struct sort_bucket *b)
101 struct sort_bucket *join = (struct sort_bucket *) b->n.prev;
102 ASSERT(join->flags & SBF_FINAL);
103 ASSERT(b->runs == 1);
105 if (!sbuck_has_file(join))
107 // The final bucket doesn't have any file associated yet, so replace
108 // it with the new bucket.
109 SORT_XTRACE(3, "Replaced final bucket");
110 b->flags |= SBF_FINAL;
115 SORT_TRACE("Copying to output file: %s", F_BSIZE(b));
116 struct fastbuf *src = sbuck_read(b);
117 struct fastbuf *dest = sbuck_write(join);
118 bbcopy(src, dest, ~0U);
124 sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
126 struct sort_bucket *ins[3] = { NULL }, *outs[3] = { NULL };
127 cnode *list_pos = b->n.prev;
129 struct sort_bucket *join = sbuck_join_to(b, &join_size);
131 if (!(sorter_debug & SORT_DEBUG_NO_PRESORT) || (b->flags & SBF_CUSTOM_PRESORT))
133 SORT_XTRACE(3, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
134 sorter_start_timer(ctx);
135 ins[0] = sbuck_new(ctx);
136 if (!sorter_presort(ctx, b, ins[0], join ? : ins[0]))
138 sorter_stop_timer(ctx, &ctx->total_pre_time);
139 ucw_off_t size = sbuck_ins_or_join(ins[0], list_pos, join, join_size);
140 SORT_XTRACE(((b->flags & SBF_SOURCE) ? 1 : 3), "Sorted in memory (%s, %dMB/s)", stk_fsize(size), sorter_speed(ctx, size));
145 ins[1] = sbuck_new(ctx);
147 while (sorter_presort(ctx, b, ins[i], ins[i]))
150 sorter_stop_timer(ctx, &ctx->total_pre_time);
151 SORT_TRACE("Presorting pass (%d+%d runs, %s+%s, %dMB/s)",
152 ins[0]->runs, ins[1]->runs,
153 F_BSIZE(ins[0]), F_BSIZE(ins[1]),
154 sorter_speed(ctx, sbuck_size(ins[0]) + sbuck_size(ins[1])));
158 SORT_XTRACE(2, "Presorting disabled");
162 SORT_XTRACE(3, "Main sorting");
166 sorter_start_timer(ctx);
167 if (ins[0]->runs <= 1 && ins[1]->runs <= 1 && join)
169 // This is guaranteed to produce a single run, so join if possible
172 ctx->twoway_merge(ctx, ins, outs);
173 ucw_off_t size = sbuck_ins_or_join(NULL, NULL, join, join_size);
174 sorter_stop_timer(ctx, &ctx->total_ext_time);
175 SORT_TRACE("Mergesort pass %d (final run, %s, %dMB/s)", pass, stk_fsize(size), sorter_speed(ctx, size));
180 outs[0] = sbuck_new(ctx);
181 outs[1] = sbuck_new(ctx);
183 ctx->twoway_merge(ctx, ins, outs);
184 sorter_stop_timer(ctx, &ctx->total_ext_time);
185 SORT_TRACE("Mergesort pass %d (%d+%d runs, %s+%s, %dMB/s)", pass,
186 outs[0]->runs, outs[1]->runs,
187 F_BSIZE(outs[0]), F_BSIZE(outs[1]),
188 sorter_speed(ctx, sbuck_size(outs[0]) + sbuck_size(outs[1])));
191 memcpy(ins, outs, 3*sizeof(struct sort_bucket *));
192 } while (sbuck_have(ins[1]));
195 clist_insert_after(&ins[0]->n, list_pos);
199 sorter_multiway(struct sort_context *ctx, struct sort_bucket *b)
202 cnode *list_pos = b->n.prev;
204 struct sort_bucket *join = sbuck_join_to(b, &join_size);
205 uint trace_level = (b->flags & SBF_SOURCE) ? 1 : 3;
208 ASSERT(!(sorter_debug & SORT_DEBUG_NO_PRESORT));
209 SORT_XTRACE(3, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
213 sorter_start_timer(ctx);
216 struct sort_bucket *p = sbuck_new(ctx);
217 cont = sorter_presort(ctx, b, p, (!part_cnt && join) ? join : p);
221 clist_add_tail(&parts, &p->n);
222 total_size += sbuck_size(p);
229 sorter_stop_timer(ctx, &ctx->total_pre_time);
230 sorter_free_buf(ctx);
235 ucw_off_t size = sbuck_ins_or_join(clist_head(&parts), list_pos, (part_cnt ? NULL : join), join_size);
236 SORT_XTRACE(trace_level, "Sorted in memory (%s, %dMB/s)", stk_fsize(size), sorter_speed(ctx, size));
240 SORT_TRACE("Multi-way presorting pass (%d parts, %s, %dMB/s)", part_cnt, stk_fsize(total_size), sorter_speed(ctx, total_size));
242 uint max_ways = 1 << sorter_max_multiway_bits;
243 struct sort_bucket *ways[max_ways+1];
244 SORT_XTRACE(3, "Starting up to %d-way merge", max_ways);
248 struct sort_bucket *p;
249 while (n < max_ways && (p = clist_head(&parts)))
257 struct sort_bucket *out;
258 if (clist_empty(&parts) && join)
261 out = sbuck_new(ctx);
262 sorter_start_timer(ctx);
263 ctx->multiway_merge(ctx, ways, out);
264 sorter_stop_timer(ctx, &ctx->total_ext_time);
266 for (uint i=0; i<n; i++)
269 if (clist_empty(&parts))
271 ucw_off_t size = sbuck_ins_or_join((join ? NULL : out), list_pos, join, join_size);
272 SORT_TRACE("Multi-way merge completed (%d ways, %s, %dMB/s)", n, stk_fsize(size), sorter_speed(ctx, size));
278 clist_add_tail(&parts, &out->n);
279 SORT_TRACE("Multi-way merge pass (%d ways, %s, %dMB/s)", n, F_BSIZE(out), sorter_speed(ctx, sbuck_size(out)));
285 sorter_radix(struct sort_context *ctx, struct sort_bucket *b, uint bits)
287 // Add more bits if requested and allowed.
288 bits = MIN(bits + sorter_add_radix_bits, sorter_max_radix_bits);
290 uint nbuck = 1 << bits;
291 SORT_XTRACE(3, "Running radix split on %s with hash %d bits of %d (expecting %s buckets)",
292 F_BSIZE(b), bits, b->hash_bits, stk_fsize(sbuck_size(b) / nbuck));
293 sorter_free_buf(ctx);
294 sorter_start_timer(ctx);
296 struct sort_bucket **outs = alloca(nbuck * sizeof(struct sort_bucket *));
297 for (uint i=nbuck; i--; )
299 outs[i] = sbuck_new(ctx);
300 outs[i]->hash_bits = b->hash_bits - bits;
301 clist_insert_after(&outs[i]->n, &b->n);
304 ctx->radix_split(ctx, b, outs, b->hash_bits - bits, bits);
306 u64 min = ~(u64)0, max = 0, sum = 0;
307 for (uint i=0; i<nbuck; i++)
309 u64 s = sbuck_size(outs[i]);
314 sbuck_swap_out(outs[i]);
317 sorter_stop_timer(ctx, &ctx->total_ext_time);
318 SORT_TRACE("Radix split (%d buckets, %s min, %s max, %s avg, %dMB/s)", nbuck,
319 stk_fsize(min), stk_fsize(max), stk_fsize(sum / nbuck), sorter_speed(ctx, sum));
324 sorter_decide(struct sort_context *ctx, struct sort_bucket *b)
326 // Drop empty buckets
329 SORT_XTRACE(4, "Dropping empty bucket");
334 // How many bits of bucket size we have to reduce before it fits in the RAM?
335 // (this is insanely large if the input size is unknown, but it serves our purpose)
336 u64 insize = sbuck_size(b);
337 u64 mem = ctx->internal_estimate(ctx, b) * 0.8; // Magical factor accounting for various non-uniformities
339 while ((insize >> bits) > mem)
342 // Calculate the possibilities of radix splits
344 if (!ctx->radix_split ||
345 (b->flags & SBF_CUSTOM_PRESORT) ||
346 (sorter_debug & SORT_DEBUG_NO_RADIX))
350 radix_bits = MIN(bits, b->hash_bits);
351 radix_bits = MIN(radix_bits, sorter_max_radix_bits);
352 if (radix_bits < sorter_min_radix_bits)
356 // The same for multi-way merges
358 if (!ctx->multiway_merge ||
359 (sorter_debug & SORT_DEBUG_NO_MULTIWAY) ||
360 (sorter_debug & SORT_DEBUG_NO_PRESORT))
364 multiway_bits = MIN(bits, sorter_max_multiway_bits);
365 if (multiway_bits < sorter_min_multiway_bits)
369 SORT_XTRACE(3, "Decisions: size=%s max=%s runs=%d bits=%d hash=%d -> radix=%d multi=%d",
370 stk_fsize(insize), stk_fsize(mem), b->runs, bits, b->hash_bits,
371 radix_bits, multiway_bits);
373 // If the input already consists of a single run, just join it
375 return sorter_join(b);
377 // If everything fits in memory, the 2-way strategy will sort it in memory
379 return sorter_twoway(ctx, b);
381 // If we can reduce everything in one pass, do so and prefer radix splits
382 if (radix_bits == bits)
383 return sorter_radix(ctx, b, radix_bits);
384 if (multiway_bits == bits)
385 return sorter_multiway(ctx, b);
387 // Otherwise, reduce as much as possible and again prefer radix splits
389 return sorter_radix(ctx, b, radix_bits);
391 return sorter_multiway(ctx, b);
393 // Fall back to 2-way strategy if nothing else applies
394 return sorter_twoway(ctx, b);
398 sorter_run(struct sort_context *ctx)
400 ctx->pool = mp_new(4096);
401 clist_init(&ctx->bucket_list);
402 sorter_prepare_buf(ctx);
403 asort_start_threads(0);
405 // Create bucket containing the source
406 struct sort_bucket *bin = sbuck_new(ctx);
407 bin->flags = SBF_SOURCE | SBF_OPEN_READ;
408 if (ctx->custom_presort)
409 bin->flags |= SBF_CUSTOM_PRESORT;
411 bin->fb = ctx->in_fb;
413 bin->size = ctx->in_size;
414 bin->hash_bits = ctx->hash_bits;
415 clist_add_tail(&ctx->bucket_list, &bin->n);
416 SORT_XTRACE(2, "Input size: %s, %d hash bits", F_BSIZE(bin), bin->hash_bits);
417 ctx->fb_params = (bin->size < sorter_small_input) ? &sorter_small_fb_params : &sorter_fb_params;
419 // Create bucket for the output
420 struct sort_bucket *bout = sbuck_new(ctx);
421 bout->flags = SBF_FINAL;
422 if (bout->fb = ctx->out_fb)
423 bout->flags |= SBF_OPEN_WRITE;
426 clist_add_head(&ctx->bucket_list, &bout->n);
428 // Repeatedly sort buckets
429 struct sort_bucket *b;
430 while (bout = clist_head(&ctx->bucket_list), b = clist_next(&ctx->bucket_list, &bout->n))
431 sorter_decide(ctx, b);
433 asort_stop_threads();
434 sorter_free_buf(ctx);
435 sbuck_write(bout); // Force empty bucket to a file
436 SORT_XTRACE(2, "Final size: %s", F_BSIZE(bout));
437 SORT_XTRACE(2, "Final timings: %.3fs external sorting, %.3fs presorting, %.3fs internal sorting",
438 ctx->total_ext_time/1000., ctx->total_pre_time/1000., ctx->total_int_time/1000.);
439 ctx->out_fb = sbuck_read(bout);
440 mp_delete(ctx->pool);