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 "lib/fastbuf.h"
12 #include "lib/mempool.h"
13 #include "lib/stkstring.h"
14 #include "lib/sorter/common.h"
20 #define F_BSIZE(b) stk_fsize(sbuck_size(b))
23 sorter_start_timer(struct sort_context *ctx)
25 init_timer(&ctx->start_time);
29 sorter_stop_timer(struct sort_context *ctx, uns *account_to)
31 ctx->last_pass_time = get_timer(&ctx->start_time);
32 *account_to += ctx->last_pass_time;
36 sorter_speed(struct sort_context *ctx, u64 size)
40 if (!ctx->last_pass_time)
42 return (uns)((double)size / (1<<20) * 1000 / ctx->last_pass_time);
46 sorter_presort(struct sort_context *ctx, struct sort_bucket *in, struct sort_bucket *out, struct sort_bucket *out_only)
48 sorter_alloc_buf(ctx);
49 if (in->flags & SBF_CUSTOM_PRESORT)
52 * The trick with automatic joining, which we use for the normal presorter,
53 * is not necessary with the custom presorter, because the custom presorter
54 * is never called in the middle of the sorted data.
56 struct fastbuf *f = sbuck_write(out);
58 return ctx->custom_presort(f, ctx->big_buf, ctx->big_buf_size);
60 return ctx->internal_sort(ctx, in, out, out_only);
63 static struct sort_bucket *
64 sbuck_join_to(struct sort_bucket *b, sh_off_t *sizep)
66 if (sorter_debug & SORT_DEBUG_NO_JOIN)
69 struct sort_bucket *out = (struct sort_bucket *) b->n.prev; // Such bucket is guaranteed to exist
70 if (!(out->flags & SBF_FINAL))
72 ASSERT(out->runs == 1);
73 *sizep = sbuck_size(out);
78 sbuck_ins_or_join(struct sort_bucket *b, cnode *list_pos, struct sort_bucket *join, sh_off_t join_size)
80 if (join && join->runs >= 2)
84 ASSERT(join->runs == 2);
86 return sbuck_size(join) - join_size;
90 clist_insert_after(&b->n, list_pos);
98 sorter_join(struct sort_bucket *b)
100 struct sort_bucket *join = (struct sort_bucket *) b->n.prev;
101 ASSERT(join->flags & SBF_FINAL);
102 ASSERT(b->runs == 1);
104 if (!sbuck_has_file(join))
106 // The final bucket doesn't have any file associated yet, so replace
107 // it with the new bucket.
108 SORT_XTRACE(3, "Replaced final bucket");
109 b->flags |= SBF_FINAL;
114 SORT_TRACE("Copying to output file: %s", F_BSIZE(b));
115 struct fastbuf *src = sbuck_read(b);
116 struct fastbuf *dest = sbuck_write(join);
117 bbcopy(src, dest, ~0U);
123 sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
125 struct sort_bucket *ins[3] = { NULL }, *outs[3] = { NULL };
126 cnode *list_pos = b->n.prev;
128 struct sort_bucket *join = sbuck_join_to(b, &join_size);
130 if (!(sorter_debug & SORT_DEBUG_NO_PRESORT) || (b->flags & SBF_CUSTOM_PRESORT))
132 SORT_XTRACE(3, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
133 sorter_start_timer(ctx);
134 ins[0] = sbuck_new(ctx);
135 if (!sorter_presort(ctx, b, ins[0], join ? : ins[0]))
137 sorter_stop_timer(ctx, &ctx->total_pre_time);
138 sh_off_t size = sbuck_ins_or_join(ins[0], list_pos, join, join_size);
139 SORT_XTRACE(((b->flags & SBF_SOURCE) ? 1 : 3), "Sorted in memory (%s, %dMB/s)", stk_fsize(size), sorter_speed(ctx, size));
144 ins[1] = sbuck_new(ctx);
146 while (sorter_presort(ctx, b, ins[i], ins[i]))
149 sorter_stop_timer(ctx, &ctx->total_pre_time);
150 SORT_TRACE("Presorting pass (%d+%d runs, %s+%s, %dMB/s)",
151 ins[0]->runs, ins[1]->runs,
152 F_BSIZE(ins[0]), F_BSIZE(ins[1]),
153 sorter_speed(ctx, sbuck_size(ins[0]) + sbuck_size(ins[1])));
157 SORT_XTRACE(2, "Presorting disabled");
161 SORT_XTRACE(3, "Main sorting");
165 sorter_start_timer(ctx);
166 if (ins[0]->runs <= 1 && ins[1]->runs <= 1 && join)
168 // This is guaranteed to produce a single run, so join if possible
171 ctx->twoway_merge(ctx, ins, outs);
172 sh_off_t size = sbuck_ins_or_join(NULL, NULL, join, join_size);
173 sorter_stop_timer(ctx, &ctx->total_ext_time);
174 SORT_TRACE("Mergesort pass %d (final run, %s, %dMB/s)", pass, stk_fsize(size), sorter_speed(ctx, size));
179 outs[0] = sbuck_new(ctx);
180 outs[1] = sbuck_new(ctx);
182 ctx->twoway_merge(ctx, ins, outs);
183 sorter_stop_timer(ctx, &ctx->total_ext_time);
184 SORT_TRACE("Mergesort pass %d (%d+%d runs, %s+%s, %dMB/s)", pass,
185 outs[0]->runs, outs[1]->runs,
186 F_BSIZE(outs[0]), F_BSIZE(outs[1]),
187 sorter_speed(ctx, sbuck_size(outs[0]) + sbuck_size(outs[1])));
190 memcpy(ins, outs, 3*sizeof(struct sort_bucket *));
191 } while (sbuck_have(ins[1]));
194 clist_insert_after(&ins[0]->n, list_pos);
198 sorter_multiway(struct sort_context *ctx, struct sort_bucket *b)
201 cnode *list_pos = b->n.prev;
203 struct sort_bucket *join = sbuck_join_to(b, &join_size);
204 uns trace_level = (b->flags & SBF_SOURCE) ? 1 : 3;
207 ASSERT(!(sorter_debug & SORT_DEBUG_NO_PRESORT));
208 SORT_XTRACE(3, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
212 sorter_start_timer(ctx);
215 struct sort_bucket *p = sbuck_new(ctx);
216 cont = sorter_presort(ctx, b, p, (!part_cnt && join) ? join : p);
220 clist_add_tail(&parts, &p->n);
221 total_size += sbuck_size(p);
228 sorter_stop_timer(ctx, &ctx->total_pre_time);
229 sorter_free_buf(ctx);
234 sh_off_t size = sbuck_ins_or_join(clist_head(&parts), list_pos, (part_cnt ? NULL : join), join_size);
235 SORT_XTRACE(trace_level, "Sorted in memory (%s, %dMB/s)", stk_fsize(size), sorter_speed(ctx, size));
239 SORT_TRACE("Multi-way presorting pass (%d parts, %s, %dMB/s)", part_cnt, stk_fsize(total_size), sorter_speed(ctx, total_size));
241 uns max_ways = 1 << sorter_max_multiway_bits;
242 struct sort_bucket *ways[max_ways+1];
243 SORT_XTRACE(3, "Starting up to %d-way merge", max_ways);
247 struct sort_bucket *p;
248 while (n < max_ways && (p = clist_head(&parts)))
256 struct sort_bucket *out;
257 if (clist_empty(&parts) && join)
260 out = sbuck_new(ctx);
261 sorter_start_timer(ctx);
262 ctx->multiway_merge(ctx, ways, out);
263 sorter_stop_timer(ctx, &ctx->total_ext_time);
265 for (uns i=0; i<n; i++)
268 if (clist_empty(&parts))
270 sh_off_t size = sbuck_ins_or_join((join ? NULL : out), list_pos, join, join_size);
271 SORT_TRACE("Multi-way merge completed (%d ways, %s, %dMB/s)", n, stk_fsize(size), sorter_speed(ctx, size));
277 clist_add_tail(&parts, &out->n);
278 SORT_TRACE("Multi-way merge pass (%d ways, %s, %dMB/s)", n, F_BSIZE(out), sorter_speed(ctx, sbuck_size(out)));
284 sorter_radix(struct sort_context *ctx, struct sort_bucket *b, uns bits)
286 // Add more bits if requested and allowed.
287 bits = MIN(bits + sorter_add_radix_bits, sorter_max_radix_bits);
289 uns nbuck = 1 << bits;
290 SORT_XTRACE(3, "Running radix split on %s with hash %d bits of %d (expecting %s buckets)",
291 F_BSIZE(b), bits, b->hash_bits, stk_fsize(sbuck_size(b) / nbuck));
292 sorter_free_buf(ctx);
293 sorter_start_timer(ctx);
295 struct sort_bucket **outs = alloca(nbuck * sizeof(struct sort_bucket *));
296 for (uns i=nbuck; i--; )
298 outs[i] = sbuck_new(ctx);
299 outs[i]->hash_bits = b->hash_bits - bits;
300 clist_insert_after(&outs[i]->n, &b->n);
303 ctx->radix_split(ctx, b, outs, b->hash_bits - bits, bits);
305 u64 min = ~(u64)0, max = 0, sum = 0;
306 for (uns i=0; i<nbuck; i++)
308 u64 s = sbuck_size(outs[i]);
313 sbuck_swap_out(outs[i]);
316 sorter_stop_timer(ctx, &ctx->total_ext_time);
317 SORT_TRACE("Radix split (%d buckets, %s min, %s max, %s avg, %dMB/s)", nbuck,
318 stk_fsize(min), stk_fsize(max), stk_fsize(sum / nbuck), sorter_speed(ctx, sum));
323 sorter_decide(struct sort_context *ctx, struct sort_bucket *b)
325 // Drop empty buckets
328 SORT_XTRACE(4, "Dropping empty bucket");
333 // How many bits of bucket size we have to reduce before it fits in the RAM?
334 // (this is insanely large if the input size is unknown, but it serves our purpose)
335 u64 insize = sbuck_size(b);
336 u64 mem = ctx->internal_estimate(ctx, b) * 0.8; // Magical factor accounting for various non-uniformities
338 while ((insize >> bits) > mem)
341 // Calculate the possibilities of radix splits
343 if (!ctx->radix_split ||
344 (b->flags & SBF_CUSTOM_PRESORT) ||
345 (sorter_debug & SORT_DEBUG_NO_RADIX))
349 radix_bits = MIN(bits, b->hash_bits);
350 radix_bits = MIN(radix_bits, sorter_max_radix_bits);
351 if (radix_bits < sorter_min_radix_bits)
355 // The same for multi-way merges
357 if (!ctx->multiway_merge ||
358 (sorter_debug & SORT_DEBUG_NO_MULTIWAY) ||
359 (sorter_debug & SORT_DEBUG_NO_PRESORT))
363 multiway_bits = MIN(bits, sorter_max_multiway_bits);
364 if (multiway_bits < sorter_min_multiway_bits)
368 SORT_XTRACE(3, "Decisions: size=%s max=%s runs=%d bits=%d hash=%d -> radix=%d multi=%d",
369 stk_fsize(insize), stk_fsize(mem), b->runs, bits, b->hash_bits,
370 radix_bits, multiway_bits);
372 // If the input already consists of a single run, just join it
374 return sorter_join(b);
376 // If everything fits in memory, the 2-way strategy will sort it in memory
378 return sorter_twoway(ctx, b);
380 // If we can reduce everything in one pass, do so and prefer radix splits
381 if (radix_bits == bits)
382 return sorter_radix(ctx, b, radix_bits);
383 if (multiway_bits == bits)
384 return sorter_multiway(ctx, b);
386 // Otherwise, reduce as much as possible and again prefer radix splits
388 return sorter_radix(ctx, b, radix_bits);
390 return sorter_multiway(ctx, b);
392 // Fall back to 2-way strategy if nothing else applies
393 return sorter_twoway(ctx, b);
397 sorter_run(struct sort_context *ctx)
399 ctx->pool = mp_new(4096);
400 clist_init(&ctx->bucket_list);
401 sorter_prepare_buf(ctx);
402 asort_start_threads(0);
404 // Create bucket containing the source
405 struct sort_bucket *bin = sbuck_new(ctx);
406 bin->flags = SBF_SOURCE | SBF_OPEN_READ;
407 if (ctx->custom_presort)
408 bin->flags |= SBF_CUSTOM_PRESORT;
410 bin->fb = ctx->in_fb;
412 bin->size = ctx->in_size;
413 bin->hash_bits = ctx->hash_bits;
414 clist_add_tail(&ctx->bucket_list, &bin->n);
415 SORT_XTRACE(2, "Input size: %s, %d hash bits", F_BSIZE(bin), bin->hash_bits);
416 ctx->fb_params = (bin->size < sorter_small_input) ? &sorter_small_fb_params : &sorter_fb_params;
418 // Create bucket for the output
419 struct sort_bucket *bout = sbuck_new(ctx);
420 bout->flags = SBF_FINAL;
421 if (bout->fb = ctx->out_fb)
422 bout->flags |= SBF_OPEN_WRITE;
425 clist_add_head(&ctx->bucket_list, &bout->n);
427 // Repeatedly sort buckets
428 struct sort_bucket *b;
429 while (bout = clist_head(&ctx->bucket_list), b = clist_next(&ctx->bucket_list, &bout->n))
430 sorter_decide(ctx, b);
432 asort_stop_threads();
433 sorter_free_buf(ctx);
434 sbuck_write(bout); // Force empty bucket to a file
435 SORT_XTRACE(2, "Final size: %s", F_BSIZE(bout));
436 SORT_XTRACE(2, "Final timings: %.3fs external sorting, %.3fs presorting, %.3fs internal sorting",
437 ctx->total_ext_time/1000., ctx->total_pre_time/1000., ctx->total_int_time/1000.);
438 ctx->out_fb = sbuck_read(bout);
439 mp_delete(ctx->pool);