distributed/merge.c

   1 /* The master keeps stats received from slaves in a queue of received
   2  * buffers that are merged together with the functions implemented
   3  * here. It also has one hash table per slave to maintain cumulative
   4  * stats that have not yet been sent to the slave machine. The queue
   5  * and the hash tables are cleared at each new move. */
   6
   7 #include <assert.h>
   8 #include <stdio.h>
   9 #include <limits.h>
  10
  11 #include "debug.h"
  12 #include "timeinfo.h"
  13 #include "distributed/distributed.h"
  14 #include "distributed/merge.h"
  15
  16 /* We merge together debug stats for all hash tables. */
  17 static struct hash_counts h_counts;
  18
  19 /* Display and reset hash statistics. For debugging only. */
  20 void
  21 merge_print_stats(int total_hnodes)
  22 {
  23         if (DEBUGL(3)) {
  24                 char buf[BSIZE];
  25                 snprintf(buf, sizeof(buf),
  26                          "stats occupied %ld %.1f%% inserts %ld collisions %ld/%ld %.1f%%\n",
  27                          h_counts.occupied, h_counts.occupied * 100.0 / total_hnodes,
  28                          h_counts.inserts, h_counts.collisions, h_counts.lookups,
  29                          h_counts.collisions * 100.0 / (h_counts.lookups + 1));
  30                 logline(NULL, "* ", buf);
  31         }
  32         if (DEBUG_MODE) h_counts.occupied = 0;
  33 }
  34
  35 /* We maintain counts per bucket to avoid sorting large arrays.
  36  * All nodes with n updates since last send go to bucket n.
  37  * We have at most max_merged_nodes = (max_slaves-1) * shared_nodes
  38  * nodes to merge, 230K nodes for 24 slaves. If we put all nodes above
  39  * 1K updates in the top bucket, we get at most 230 nodes in this
  40  * bucket. So we can select exactly the best shared_nodes nodes if
  41  * shared_nodes >= 230. In practice there is overlap between
  42  * nodes sent by different slaves so shared_nodes can be lower. */
  43 #define MAX_BUCKETS 1024
  44
  45 /* Update the hash table for the given increment stats,
  46  * and increment the bucket count. Return the hash index.
  47  * The slave lock is not held on either entry or exit of this function */
  48 static inline int
  49 stats_tally(struct incr_stats *s, struct slave_state *sstate, int *bucket_count)
  50 {
  51         int h;
  52         bool found;
  53         struct incr_stats *stats_htable = sstate->stats_htable;
  54         find_hash(h, stats_htable, sstate->stats_hbits, s->coord_path, found, h_counts);
  55         if (found) {
  56                 assert(stats_htable[h].incr.playouts > 0);
  57                 stats_add_result(&stats_htable[h].incr, s->incr.value, s->incr.playouts);
  58         } else {
  59                 stats_htable[h] = *s;
  60                 if (DEBUG_MODE) h_counts.inserts++, h_counts.occupied++;
  61         }
  62
  63         int incr = stats_htable[h].incr.playouts;
  64         if (incr >= MAX_BUCKETS) incr = MAX_BUCKETS - 1;
  65         bucket_count[incr]++;
  66         return h;
  67 }
  68
  69 static struct incr_stats terminator = { .coord_path = INT64_MAX };
  70
  71 /* Initialize the next pointers (see merge_new_stats()).
  72  * Exclude invalid buffers and my own buffers by setting their next pointer
  73  * to a terminator value. Update min if there are too many nodes to merge,
  74  * so that merge time remains reasonable and the merge buffer doesn't overflow.
  75  * (We skip the oldest buffers if the slave thread is too much behind. It is
  76  * more important to get frequent incomplete updates than late complete updates.)
  77  * Return the total number of nodes to be merged.
  78  * The slave lock is not held on either entry or exit of this function. */
  79 static int
  80 filter_buffers(struct slave_state *sstate, struct incr_stats **next,
  81                int *min, int max)
  82 {
  83         int size = 0;
  84         int max_size = sstate->max_merged_nodes * sizeof(struct incr_stats);
  85
  86         for (int q = max; q >= *min; q--) {
  87                 if (!receive_queue[q] || receive_queue[q]->owner == sstate->thread_id) {
  88                         next[q] = &terminator;
  89                 } else if (size + receive_queue[q]->size > max_size) {
  90                         *min = q + 1;
  91                         assert(*min <= max);
  92                         break;
  93                 } else {
  94                         next[q] = (struct incr_stats *)receive_queue[q]->buf;
  95                         size += receive_queue[q]->size;
  96                 }
  97         }
  98         return size / sizeof(struct incr_stats);
  99 }
 100
 101 /* Return the minimum coord path of next[min..max].
 102  * This implementation is optimized for small values of max - min,
 103  * which is the case if slaves are not too much behind.
 104  * A heap (priority queue) could be used otherwise.
 105  * The returned value might be come from a buffer that has
 106  * been invalidated, the caller must check for this; in this
 107  * case the returned value is < the correct value. */
 108 static inline path_t
 109 min_coord(struct incr_stats **next, int min, int max)
 110 {
 111         path_t min_c = next[min]->coord_path;
 112         for (int q = min + 1; q <= max; q++) {
 113                 if (next[q]->coord_path < min_c)
 114                         min_c = next[q]->coord_path;
 115         }
 116         return min_c;
 117 }
 118
 119 /* Merge all valid incremental stats in receive_queue[min..max],
 120  * update the hash table, set the bucket counts, and save the
 121  * list of updated hash table entries. The input buffers and
 122  * the output buffer are all sorted by increasing coord path.
 123  * The input buffers end with a terminator value INT64_MAX.
 124  * Return the number of updated hash table entries. */
 125
 126 /* The slave lock is not held on either entry or exit of this function,
 127  * so receive_queue entries may be invalidated while we scan them.
 128  * The receive queue might grow while we scan it but we ignore
 129  * entries above max, they will be processed at the next call.
 130  * This function does not modify the receive queue. */
 131 static int
 132 merge_new_stats(struct slave_state *sstate, int min, int max,
 133                 int *bucket_count, int *nodes_read, int last_queue_age)
 134 {
 135         *nodes_read = 0;
 136         if (max < min) return 0;
 137
 138         /* next[q] is the next value to be checked in receive_queue[q]->buf */
 139         struct incr_stats *next_[max - min + 1];
 140         struct incr_stats **next = next_ - min;
 141         *nodes_read = filter_buffers(sstate, next, &min, max);
 142
 143         /* prev_min_c is only used for debugging. */
 144         path_t prev_min_c = 0;
 145
 146         /* Do N-way merge, processing one coord path per iteration.
 147          * If the minimum coord is INT64_MAX, either all buffers are
 148          * invalidated, or at least one is valid and we are at the
 149          * end of all valid buffers. In both cases we're done. */
 150         int merge_count = 0;
 151         path_t min_c;
 152         while ((min_c = min_coord(next, min, max)) != INT64_MAX) {
 153
 154                 struct incr_stats sum = { .coord_path = min_c,
 155                                           .incr = { .playouts = 0, .value = 0.0 }};
 156                 for (int q = min; q <= max; q++) {
 157                         struct incr_stats s = *(next[q]);
 158
 159                         /* If s.coord_path != min_c, we must skip s.coord_path for now.
 160                          * If min_c is invalid, a future iteration will get a stable
 161                          * value since the call of min_coord(), so at some point we will
 162                          * get s.coord_path == min_c and we will not loop forever. */
 163                          if (s.coord_path != min_c) continue;
 164
 165                         /* We check the buffer validity after s.coord has been checked
 166                          * to avoid a race condition, and also to avoid multiple useless
 167                          * checks for the same coord_path. */
 168                         if (unlikely(!receive_queue[q])) {
 169                                 next[q] = &terminator;
 170                                 continue;
 171                         }
 172
 173                         /* Stop if we have a new move. If queue_age is incremented
 174                          * after this check, the merged output will be discarded. */
 175                         if (unlikely(queue_age > last_queue_age)) return 0;
 176
 177                         /* s.coord_path is valid here, so min_c is valid too.
 178                          * (An invalid min_c would be < s.coord_path.) */
 179                         assert(min_c > prev_min_c);
 180
 181                         assert(s.coord_path && s.incr.playouts);
 182                         stats_add_result(&sum.incr, s.incr.value, s.incr.playouts);
 183                         next[q]++;
 184                 }
 185                 /* All the buffers containing min_c may have been invalidated
 186                  * so sum may still be zero. But in this case the next[q] which
 187                  * contained min_c have been reset to &terminator so we will
 188                  * not loop forever. */
 189                 if (!sum.incr.playouts) continue;
 190
 191                 assert(min_c > prev_min_c);
 192                 if (DEBUG_MODE) prev_min_c = min_c;
 193
 194                 /* At this point sum contains only valid increments,
 195                  * so we can add it to the hash table. */
 196                 assert(merge_count < sstate->max_merged_nodes);
 197                 sstate->merged[merge_count++] = stats_tally(&sum, sstate, bucket_count);
 198         }
 199         return merge_count;
 200 }
 201
 202 /* Save in buf the best increments from other slaves merged previously.
 203  * To avoid a costly scan of the entire hash table we only send nodes
 204  * that were previously sent recently by other slaves. It is possible
 205  * but very unlikely that the hash table contains some nodes with
 206  * higher number of playouts.
 207  * Return the number of nodes to be sent.
 208  * The slave lock is not held on either entry or exit of this function. */
 209 static int
 210 output_stats(struct incr_stats *buf, struct slave_state *sstate,
 211              int *bucket_count, int merge_count)
 212 {
 213         /* Find the minimum increment to send. The bucket with minimum
 214          * increment may be sent only partially. */
 215         int out_count = 0;
 216         int min_incr = MAX_BUCKETS;
 217         int shared_nodes = sstate->max_buf_size / sizeof(*buf);
 218         do {
 219                 out_count += bucket_count[--min_incr];
 220         } while (min_incr > 1 && out_count < shared_nodes);
 221
 222         /* Send all all increments > min_incr plus whatever we can at min_incr. */
 223         int min_count = bucket_count[min_incr] - (out_count - shared_nodes);
 224         out_count = 0;
 225         int *merged = sstate->merged;
 226         struct incr_stats *stats_htable = sstate->stats_htable;
 227         while (merge_count--) {
 228                 int h = *merged++;
 229                 int delta = stats_htable[h].incr.playouts - min_incr;
 230                 if (delta < 0 || (delta == 0 && --min_count < 0)) continue;
 231
 232                 assert (out_count < shared_nodes);
 233                 buf[out_count++] = stats_htable[h];
 234
 235                 /* Clear the hash table entry. (We could instead
 236                  * just clear the playouts but clearing the entry
 237                  * leads to fewer collisions later.) */
 238                 stats_htable[h].coord_path = 0;
 239                 if (DEBUG_MODE) h_counts.occupied--;
 240         }
 241         /* The slave expects increments sorted by coord path
 242          * but they are sorted already. */
 243         return out_count;
 244 }
 245
 246 /* Get all incremental stats received from other slaves since the
 247  * last send. Store in buf the stats with largest playout increments.
 248  * Return the byte size of the resulting buffer. The caller must
 249  * check that the result is still valid.
 250  * The slave lock is held on both entry and exit of this function. */
 251 static int
 252 get_new_stats(struct incr_stats *buf, struct slave_state *sstate, int cmd_id)
 253 {
 254         /* Process all valid buffers in receive_queue[min..max] */
 255         int min = sstate->last_processed + 1;
 256         int max = queue_length - 1;
 257         if (max < min && cmd_id == sstate->stats_id) return 0;
 258
 259         sstate->last_processed = max;
 260         int last_queue_age = queue_age;
 261
 262         /* It takes time to clear the hash table and merge the stats
 263          * so do this unlocked. */
 264         protocol_unlock();
 265
 266         double start = time_now();
 267         double clear_time = 0;
 268
 269         /* Clear the hash table at a new move; the old paths in
 270          * the hash table are now meaningless. */
 271         if (cmd_id != sstate->stats_id) {
 272                 memset(sstate->stats_htable, 0,
 273                        (1 << sstate->stats_hbits) * sizeof(sstate->stats_htable[0]));
 274                 sstate->stats_id = cmd_id;
 275                 clear_time = time_now() - start;
 276         }
 277
 278         /* Set the bucket counts and update the hash table stats. */
 279         int bucket_count[MAX_BUCKETS];
 280         memset(bucket_count, 0, sizeof(bucket_count));
 281         int nodes_read;
 282         int merge_count = merge_new_stats(sstate, min, max, bucket_count,
 283                                           &nodes_read, last_queue_age);
 284
 285         /* Put the best increments in the output buffer. */
 286         int output_nodes = output_stats(buf, sstate, bucket_count, merge_count);
 287
 288         if (DEBUGVV(2)) {
 289                 char b[1024];
 290                 snprintf(b, sizeof(b), "merged %d..%d  %d/%d nodes,"
 291                          " output %d/%d nodes in %.3fms (clear %.3fms)\n",
 292                          min, max, merge_count, nodes_read, output_nodes,
 293                          sstate->max_buf_size / (int)sizeof(*buf),
 294                          (time_now() - start)*1000, clear_time*1000);
 295                 logline(&sstate->client, "= ", b);
 296         }
 297
 298         protocol_lock();
 299
 300         return output_nodes * sizeof(*buf);
 301 }
 302
 303 /* Allocate the buffers in the merge specific part of the slave sate,
 304  * and reserve space for a terminator value (see merge_insert_hook). */
 305 static void
 306 merge_state_alloc(struct slave_state *sstate)
 307 {
 308         sstate->stats_htable = calloc2(1 << sstate->stats_hbits, sizeof(struct incr_stats));
 309         sstate->merged = malloc2(sstate->max_merged_nodes * sizeof(int));
 310         sstate->max_buf_size -= sizeof(struct incr_stats);
 311 }
 312
 313 /* Append a terminator value to make merge_new_stats() more
 314  * efficient. merge_state_alloc() has reserved enough space. */
 315 static void
 316 merge_insert_hook(struct incr_stats *buf, int size)
 317 {
 318         int nodes = size / sizeof(*buf);
 319         buf[nodes].coord_path = INT64_MAX;
 320 }
 321
 322 /* Initiliaze merge-related fields of the default slave state. */
 323 void
 324 merge_init(struct slave_state *sstate, int shared_nodes, int stats_hbits, int max_slaves)
 325 {
 326         /* See merge_state_alloc() for shared_nodes + 1 */
 327         sstate->max_buf_size = (shared_nodes + 1) * sizeof(struct incr_stats);
 328         sstate->stats_hbits = stats_hbits;
 329
 330         sstate->insert_hook = (buffer_hook)merge_insert_hook;
 331         sstate->alloc_hook = merge_state_alloc;
 332         sstate->args_hook = (getargs_hook)get_new_stats;
 333
 334         /* At worst one late slave thread may have to merge up to
 335          *   shared_nodes * BUFFERS_PER_SLAVE * (max_slaves - 1)
 336          * nodes but on average it should not have to merge more than
 337          *   dist->shared_nodes * (max_slaves - 1)
 338          * Restricting the maximum number of merged nodes to the latter avoids
 339          * spending excessive time on the merge. */
 340         sstate->max_merged_nodes = shared_nodes * (max_slaves - 1);
 341 }