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 ctx->start_time = get_timestamp();
29 sorter_speed(struct sort_context *ctx, u64 size, uns *account_to)
31 timestamp_t stop_time = get_timestamp();
34 if (stop_time <= ctx->start_time)
36 *account_to += stop_time - ctx->start_time;
37 return (uns)((double)size / (1<<20) * 1000 / (stop_time-ctx->start_time));
41 sorter_presort(struct sort_context *ctx, struct sort_bucket *in, struct sort_bucket *out, struct sort_bucket *out_only)
43 sorter_alloc_buf(ctx);
44 if (in->flags & SBF_CUSTOM_PRESORT)
46 struct fastbuf *f = sbuck_write(out);
48 return ctx->custom_presort(f, ctx->big_buf, ctx->big_buf_size); // FIXME: out_only optimization?
50 return ctx->internal_sort(ctx, in, out, out_only);
53 static inline struct sort_bucket *
54 sbuck_join_to(struct sort_bucket *b)
56 if (sorter_debug & SORT_DEBUG_NO_JOIN)
59 struct sort_bucket *out = (struct sort_bucket *) b->n.prev; // Such bucket is guaranteed to exist
60 return (out->flags & SBF_FINAL) ? out : NULL;
64 sorter_join(struct sort_bucket *b)
66 struct sort_bucket *join = (struct sort_bucket *) b->n.prev;
67 ASSERT(join->flags & SBF_FINAL);
70 if (!sbuck_has_file(join))
72 // The final bucket doesn't have any file associated yet, so replace
73 // it with the new bucket.
74 SORT_XTRACE(2, "Replaced final bucket");
75 b->flags |= SBF_FINAL;
80 SORT_TRACE("Copying to output file: %s", F_BSIZE(b));
81 struct fastbuf *src = sbuck_read(b);
82 struct fastbuf *dest = sbuck_write(join);
83 bbcopy(src, dest, ~0U);
89 sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
91 struct sort_bucket *ins[3] = { NULL }, *outs[3] = { NULL };
92 cnode *list_pos = b->n.prev;
93 struct sort_bucket *join = sbuck_join_to(b);
95 if (!(sorter_debug & SORT_DEBUG_NO_PRESORT) || (b->flags & SBF_CUSTOM_PRESORT))
97 SORT_XTRACE(3, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
98 sorter_start_timer(ctx);
99 ins[0] = sbuck_new(ctx);
100 if (!sorter_presort(ctx, b, ins[0], join ? : ins[0]))
102 SORT_XTRACE(((b->flags & SBF_SOURCE) ? 1 : 2), "Sorted in memory");
106 clist_insert_after(&ins[0]->n, list_pos);
111 ins[1] = sbuck_new(ctx);
113 while (sorter_presort(ctx, b, ins[i], ins[i]))
116 SORT_TRACE("Presorting pass (%d+%d runs, %s+%s, %dMB/s)",
117 ins[0]->runs, ins[1]->runs,
118 F_BSIZE(ins[0]), F_BSIZE(ins[1]),
119 sorter_speed(ctx, sbuck_size(ins[0]) + sbuck_size(ins[1]), &ctx->total_pre_time));
123 SORT_XTRACE(2, "Presorting disabled");
127 SORT_XTRACE(3, "Main sorting");
131 sorter_start_timer(ctx);
132 if (ins[0]->runs == 1 && ins[1]->runs == 1 && join)
134 // This is guaranteed to produce a single run, so join if possible
135 sh_off_t join_size = sbuck_size(join);
138 ctx->twoway_merge(ctx, ins, outs);
139 ASSERT(join->runs == 2);
141 join_size = sbuck_size(join) - join_size;
142 SORT_TRACE("Mergesort pass %d (final run, %s, %dMB/s)", pass, stk_fsize(join_size), sorter_speed(ctx, join_size, &ctx->total_ext_time));
147 outs[0] = sbuck_new(ctx);
148 outs[1] = sbuck_new(ctx);
150 ctx->twoway_merge(ctx, ins, outs);
151 SORT_TRACE("Mergesort pass %d (%d+%d runs, %s+%s, %dMB/s)", pass,
152 outs[0]->runs, outs[1]->runs,
153 F_BSIZE(outs[0]), F_BSIZE(outs[1]),
154 sorter_speed(ctx, sbuck_size(outs[0]) + sbuck_size(outs[1]), &ctx->total_ext_time));
157 memcpy(ins, outs, 3*sizeof(struct sort_bucket *));
158 } while (sbuck_have(ins[1]));
161 clist_insert_after(&ins[0]->n, list_pos);
165 sorter_radix_bits(struct sort_context *ctx, struct sort_bucket *b)
167 if (!b->hash_bits || !ctx->radix_split ||
168 (b->flags & SBF_CUSTOM_PRESORT) ||
169 (sorter_debug & SORT_DEBUG_NO_RADIX))
172 u64 in = sbuck_size(b);
173 u64 mem = ctx->internal_estimate(ctx, b) * 0.8; // FIXME: Magical factor for hash non-uniformity
177 uns n = sorter_min_radix_bits;
178 while (n < sorter_max_radix_bits && n < b->hash_bits && (in >> n) > mem)
184 sorter_radix(struct sort_context *ctx, struct sort_bucket *b, uns bits)
186 uns nbuck = 1 << bits;
187 SORT_XTRACE(2, "Running radix split on %s with hash %d bits of %d (expecting %s buckets)",
188 F_BSIZE(b), bits, b->hash_bits, stk_fsize(sbuck_size(b) / nbuck));
189 sorter_free_buf(ctx);
190 sorter_start_timer(ctx);
192 struct sort_bucket **outs = alloca(nbuck * sizeof(struct sort_bucket *));
193 for (uns i=nbuck; i--; )
195 outs[i] = sbuck_new(ctx);
196 outs[i]->hash_bits = b->hash_bits - bits;
197 clist_insert_after(&outs[i]->n, &b->n);
200 ctx->radix_split(ctx, b, outs, b->hash_bits - bits, bits);
202 u64 min = ~(u64)0, max = 0, sum = 0;
203 for (uns i=0; i<nbuck; i++)
205 u64 s = sbuck_size(outs[i]);
210 sbuck_swap_out(outs[i]);
213 SORT_TRACE("Radix split (%d buckets, %s min, %s max, %s avg, %dMB/s)", nbuck,
214 stk_fsize(min), stk_fsize(max), stk_fsize(sum / nbuck), sorter_speed(ctx, sum, &ctx->total_ext_time));
219 sorter_run(struct sort_context *ctx)
221 ctx->pool = mp_new(4096);
222 clist_init(&ctx->bucket_list);
223 sorter_prepare_buf(ctx);
225 // Create bucket containing the source
226 struct sort_bucket *bin = sbuck_new(ctx);
227 bin->flags = SBF_SOURCE | SBF_OPEN_READ;
228 if (ctx->custom_presort)
229 bin->flags |= SBF_CUSTOM_PRESORT;
231 bin->fb = ctx->in_fb;
233 bin->size = ctx->in_size;
234 bin->hash_bits = ctx->hash_bits;
235 clist_add_tail(&ctx->bucket_list, &bin->n);
236 SORT_XTRACE(2, "Input size: %s", F_BSIZE(bin));
238 // Create bucket for the output
239 struct sort_bucket *bout = sbuck_new(ctx);
240 bout->flags = SBF_FINAL;
241 if (bout->fb = ctx->out_fb)
242 bout->flags |= SBF_OPEN_WRITE;
245 clist_add_head(&ctx->bucket_list, &bout->n);
247 struct sort_bucket *b;
249 while (bout = clist_head(&ctx->bucket_list), b = clist_next(&ctx->bucket_list, &bout->n))
251 SORT_XTRACE(2, "Next block: %s, %d hash bits", F_BSIZE(b), b->hash_bits);
254 else if (b->runs == 1)
256 else if (bits = sorter_radix_bits(ctx, b))
257 sorter_radix(ctx, b, bits);
259 sorter_twoway(ctx, b);
262 sorter_free_buf(ctx);
263 sbuck_write(bout); // Force empty bucket to a file
264 SORT_XTRACE(2, "Final size: %s", F_BSIZE(bout));
265 SORT_XTRACE(2, "Final timings: %.3fs external sorting, %.3fs presorting, %.3fs internal sorting",
266 ctx->total_ext_time/1000., ctx->total_pre_time/1000., ctx->total_int_time/1000.);
267 ctx->out_fb = sbuck_read(bout);