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)
51 struct fastbuf *f = sbuck_write(out);
53 return ctx->custom_presort(f, ctx->big_buf, ctx->big_buf_size); // FIXME: out_only optimization?
55 return ctx->internal_sort(ctx, in, out, out_only);
58 static struct sort_bucket *
59 sbuck_join_to(struct sort_bucket *b, sh_off_t *sizep)
61 if (sorter_debug & SORT_DEBUG_NO_JOIN)
64 struct sort_bucket *out = (struct sort_bucket *) b->n.prev; // Such bucket is guaranteed to exist
65 if (!(out->flags & SBF_FINAL))
67 ASSERT(out->runs == 1);
68 *sizep = sbuck_size(out);
73 sbuck_ins_or_join(struct sort_bucket *b, cnode *list_pos, struct sort_bucket *join, sh_off_t join_size)
79 ASSERT(join->runs == 2);
81 return sbuck_size(join) - join_size;
85 clist_insert_after(&b->n, list_pos);
93 sorter_join(struct sort_bucket *b)
95 struct sort_bucket *join = (struct sort_bucket *) b->n.prev;
96 ASSERT(join->flags & SBF_FINAL);
99 if (!sbuck_has_file(join))
101 // The final bucket doesn't have any file associated yet, so replace
102 // it with the new bucket.
103 SORT_XTRACE(2, "Replaced final bucket");
104 b->flags |= SBF_FINAL;
109 SORT_TRACE("Copying to output file: %s", F_BSIZE(b));
110 struct fastbuf *src = sbuck_read(b);
111 struct fastbuf *dest = sbuck_write(join);
112 bbcopy(src, dest, ~0U);
118 sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
120 struct sort_bucket *ins[3] = { NULL }, *outs[3] = { NULL };
121 cnode *list_pos = b->n.prev;
123 struct sort_bucket *join = sbuck_join_to(b, &join_size);
125 if (!(sorter_debug & SORT_DEBUG_NO_PRESORT) || (b->flags & SBF_CUSTOM_PRESORT))
127 SORT_XTRACE(3, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
128 sorter_start_timer(ctx);
129 ins[0] = sbuck_new(ctx);
130 if (!sorter_presort(ctx, b, ins[0], join ? : ins[0]))
132 sorter_stop_timer(ctx, &ctx->total_pre_time);
133 sh_off_t size = sbuck_ins_or_join(ins[0], list_pos, join, join_size);
134 SORT_XTRACE(((b->flags & SBF_SOURCE) ? 1 : 2), "Sorted in memory (%s, %dMB/s)", stk_fsize(size), sorter_speed(ctx, size));
139 ins[1] = sbuck_new(ctx);
141 while (sorter_presort(ctx, b, ins[i], ins[i]))
144 sorter_stop_timer(ctx, &ctx->total_pre_time);
145 SORT_TRACE("Presorting pass (%d+%d runs, %s+%s, %dMB/s)",
146 ins[0]->runs, ins[1]->runs,
147 F_BSIZE(ins[0]), F_BSIZE(ins[1]),
148 sorter_speed(ctx, sbuck_size(ins[0]) + sbuck_size(ins[1])));
152 SORT_XTRACE(2, "Presorting disabled");
156 SORT_XTRACE(3, "Main sorting");
160 sorter_start_timer(ctx);
161 if (ins[0]->runs == 1 && ins[1]->runs == 1 && join)
163 // This is guaranteed to produce a single run, so join if possible
166 ctx->twoway_merge(ctx, ins, outs);
167 sh_off_t size = sbuck_ins_or_join(NULL, NULL, join, join_size);
168 sorter_stop_timer(ctx, &ctx->total_ext_time);
169 SORT_TRACE("Mergesort pass %d (final run, %s, %dMB/s)", pass, stk_fsize(size), sorter_speed(ctx, size));
174 outs[0] = sbuck_new(ctx);
175 outs[1] = sbuck_new(ctx);
177 ctx->twoway_merge(ctx, ins, outs);
178 sorter_stop_timer(ctx, &ctx->total_ext_time);
179 SORT_TRACE("Mergesort pass %d (%d+%d runs, %s+%s, %dMB/s)", pass,
180 outs[0]->runs, outs[1]->runs,
181 F_BSIZE(outs[0]), F_BSIZE(outs[1]),
182 sorter_speed(ctx, sbuck_size(outs[0]) + sbuck_size(outs[1])));
185 memcpy(ins, outs, 3*sizeof(struct sort_bucket *));
186 } while (sbuck_have(ins[1]));
189 clist_insert_after(&ins[0]->n, list_pos);
193 sorter_multiway(struct sort_context *ctx, struct sort_bucket *b)
196 cnode *list_pos = b->n.prev;
198 struct sort_bucket *join = sbuck_join_to(b, &join_size);
199 uns trace_level = (b->flags & SBF_SOURCE) ? 1 : 2;
202 ASSERT(!(sorter_debug & SORT_DEBUG_NO_PRESORT));
203 SORT_XTRACE(3, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
207 sorter_start_timer(ctx);
210 struct sort_bucket *p = sbuck_new(ctx);
211 cont = sorter_presort(ctx, b, p, (!part_cnt && join) ? join : p);
215 clist_add_tail(&parts, &p->n);
216 total_size += sbuck_size(p);
223 sorter_stop_timer(ctx, &ctx->total_pre_time);
224 sorter_free_buf(ctx);
229 sh_off_t size = sbuck_ins_or_join(clist_head(&parts), list_pos, join, join_size);
230 SORT_XTRACE(trace_level, "Sorted in memory (%s, %dMB/s)", stk_fsize(size), sorter_speed(ctx, size));
234 SORT_TRACE("Multi-way presorting pass (%d parts, %s, %dMB/s)", part_cnt, stk_fsize(total_size), sorter_speed(ctx, total_size));
236 uns max_ways = 1 << sorter_max_multiway_bits;
237 struct sort_bucket *ways[max_ways+1];
238 SORT_XTRACE(2, "Starting up to %d-way merge", max_ways);
242 struct sort_bucket *p;
243 while (n < max_ways && (p = clist_head(&parts)))
251 struct sort_bucket *out;
252 if (clist_empty(&parts) && join)
255 out = sbuck_new(ctx);
256 sorter_start_timer(ctx);
257 ctx->multiway_merge(ctx, ways, out);
258 sorter_stop_timer(ctx, &ctx->total_ext_time);
260 for (uns i=0; i<n; i++)
263 if (clist_empty(&parts))
265 sh_off_t size = sbuck_ins_or_join((join ? NULL : out), list_pos, join, join_size);
266 SORT_TRACE("Multi-way merge completed (%d ways, %s, %dMB/s)", n, stk_fsize(size), sorter_speed(ctx, size));
272 clist_add_tail(&parts, &out->n);
273 SORT_TRACE("Multi-way merge pass (%d ways, %s, %dMB/s)", n, F_BSIZE(out), sorter_speed(ctx, sbuck_size(out)));
279 sorter_radix(struct sort_context *ctx, struct sort_bucket *b, uns bits)
281 uns nbuck = 1 << bits;
282 SORT_XTRACE(2, "Running radix split on %s with hash %d bits of %d (expecting %s buckets)",
283 F_BSIZE(b), bits, b->hash_bits, stk_fsize(sbuck_size(b) / nbuck));
284 sorter_free_buf(ctx);
285 sorter_start_timer(ctx);
287 struct sort_bucket **outs = alloca(nbuck * sizeof(struct sort_bucket *));
288 for (uns i=nbuck; i--; )
290 outs[i] = sbuck_new(ctx);
291 outs[i]->hash_bits = b->hash_bits - bits;
292 clist_insert_after(&outs[i]->n, &b->n);
295 ctx->radix_split(ctx, b, outs, b->hash_bits - bits, bits);
297 u64 min = ~(u64)0, max = 0, sum = 0;
298 for (uns i=0; i<nbuck; i++)
300 u64 s = sbuck_size(outs[i]);
305 sbuck_swap_out(outs[i]);
308 sorter_stop_timer(ctx, &ctx->total_ext_time);
309 SORT_TRACE("Radix split (%d buckets, %s min, %s max, %s avg, %dMB/s)", nbuck,
310 stk_fsize(min), stk_fsize(max), stk_fsize(sum / nbuck), sorter_speed(ctx, sum));
315 sorter_decide(struct sort_context *ctx, struct sort_bucket *b)
317 // Drop empty buckets
320 SORT_XTRACE(3, "Dropping empty bucket");
325 // How many bits of bucket size we have to reduce before it fits in the RAM?
326 // (this is insanely large if the input size is unknown, but it serves our purpose)
327 u64 insize = sbuck_size(b);
328 u64 mem = ctx->internal_estimate(ctx, b) * 0.8; // FIXME: Magical factor for various non-uniformities
330 while ((insize >> bits) > mem)
333 // Calculate the possibilities of radix splits
335 if (!ctx->radix_split ||
336 (b->flags & SBF_CUSTOM_PRESORT) ||
337 (sorter_debug & SORT_DEBUG_NO_RADIX))
341 radix_bits = MIN(bits, b->hash_bits);
342 radix_bits = MIN(radix_bits, sorter_max_radix_bits);
343 if (radix_bits < sorter_min_radix_bits)
347 // The same for multi-way merges
349 if (!ctx->multiway_merge ||
350 (sorter_debug & SORT_DEBUG_NO_MULTIWAY) ||
351 (sorter_debug & SORT_DEBUG_NO_PRESORT))
355 multiway_bits = MIN(bits, sorter_max_multiway_bits);
356 if (multiway_bits < sorter_min_multiway_bits)
360 SORT_XTRACE(2, "Decisions: size=%s max=%s runs=%d bits=%d hash=%d -> radix=%d multi=%d",
361 stk_fsize(insize), stk_fsize(mem), b->runs, bits, b->hash_bits,
362 radix_bits, multiway_bits);
364 // If the input already consists of a single run, just join it
366 return sorter_join(b);
368 // If everything fits in memory, the 2-way strategy will sort it in memory
370 return sorter_twoway(ctx, b);
372 // If we can reduce everything in one pass, do so and prefer radix splits
373 if (radix_bits == bits)
374 return sorter_radix(ctx, b, radix_bits);
375 if (multiway_bits == bits)
376 return sorter_multiway(ctx, b);
378 // Otherwise, reduce as much as possible and again prefer radix splits
380 return sorter_radix(ctx, b, radix_bits);
382 return sorter_multiway(ctx, b);
384 // Fall back to 2-way strategy if nothing else applies
385 return sorter_twoway(ctx, b);
389 sorter_run(struct sort_context *ctx)
391 ctx->pool = mp_new(4096);
392 clist_init(&ctx->bucket_list);
393 sorter_prepare_buf(ctx);
395 // Create bucket containing the source
396 struct sort_bucket *bin = sbuck_new(ctx);
397 bin->flags = SBF_SOURCE | SBF_OPEN_READ;
398 if (ctx->custom_presort)
399 bin->flags |= SBF_CUSTOM_PRESORT;
401 bin->fb = ctx->in_fb;
403 bin->size = ctx->in_size;
404 bin->hash_bits = ctx->hash_bits;
405 clist_add_tail(&ctx->bucket_list, &bin->n);
406 SORT_XTRACE(2, "Input size: %s, %d hash bits", F_BSIZE(bin), bin->hash_bits);
408 // Create bucket for the output
409 struct sort_bucket *bout = sbuck_new(ctx);
410 bout->flags = SBF_FINAL;
411 if (bout->fb = ctx->out_fb)
412 bout->flags |= SBF_OPEN_WRITE;
415 clist_add_head(&ctx->bucket_list, &bout->n);
417 // Repeatedly sort buckets
418 struct sort_bucket *b;
419 while (bout = clist_head(&ctx->bucket_list), b = clist_next(&ctx->bucket_list, &bout->n))
420 sorter_decide(ctx, b);
422 sorter_free_buf(ctx);
423 sbuck_write(bout); // Force empty bucket to a file
424 SORT_XTRACE(2, "Final size: %s", F_BSIZE(bout));
425 SORT_XTRACE(2, "Final timings: %.3fs external sorting, %.3fs presorting, %.3fs internal sorting",
426 ctx->total_ext_time/1000., ctx->total_pre_time/1000., ctx->total_int_time/1000.);
427 ctx->out_fb = sbuck_read(bout);