Redo mapping entirely; now dumps constant-sized shards for reducing. Much more consis...

diff --git a/binloader.cpp b/binloader.cpp
index 44088856081f05eb05e7ddb464a00ccbd4ddb98a..2ed60cbe78250ff34eba9f7c358968f90c4ddb97 100644 (file)
--- a/binloader.cpp
+++ b/binloader.cpp
@@ -1,6 +1,6 @@
 //#define _GLIBCXX_PARALLEL
 
-// Usage: ./binloader IN1 IN2 IN3 ... OUT NUM_BUCKETS
+// Usage: ./binloader IN1 IN2 IN3 ... OUT NUM_BUCKETS NUM_POS_PER_SUBSHARD
 
 #include <stdio.h>
 #include <vector>
@@ -45,27 +45,107 @@ struct Element {
        }
 };
 
+struct ShardData {
+       vector<Element> elems;
+       unique_ptr<Arena> arena;  // Used to allocate bpfen.
+       int num_written_subshards = 0;
+};
+
+void write_subshard(const char *basename, ShardData* shard, int bucket)
+{
+       string buf;  // Keep allocated.
+       char filename[256];
+       snprintf(filename, sizeof(filename), "%s.part%04d.subshard%04d",
+               basename, bucket, shard->num_written_subshards++);
+       printf("Writing SSTable %s...\n", filename);
+
+       sort(shard->elems.begin(), shard->elems.end());
+
+       mtbl_writer_options* wopt = mtbl_writer_options_init();
+       mtbl_writer_options_set_compression(wopt, MTBL_COMPRESSION_SNAPPY);
+       mtbl_writer* mtbl = mtbl_writer_init(filename, wopt);
+       Count c;
+       unordered_set<string> moves;
+       for (size_t i = 0; i < shard->elems.size(); ++i) {
+               const Element &e = shard->elems[i];
+               if (e.result == WHITE) {
+                       c.set_white(c.white() + 1);
+               } else if (e.result == DRAW) {
+                       c.set_draw(c.draw() + 1);
+               } else if (e.result == BLACK) {
+                       c.set_black(c.black() + 1);
+               }
+               if (e.white_elo >= 100 && e.black_elo >= 100) {
+                       c.set_sum_white_elo(c.sum_white_elo() + e.white_elo);
+                       c.set_sum_black_elo(c.sum_black_elo() + e.black_elo);
+                       c.set_num_elo(c.num_elo() + 1);
+               }
+               if (!c.has_first_timestamp() || e.timestamp < c.first_timestamp()) {
+                       if (e.timestamp != DUMMY_TIMESTAMP) {
+                               c.set_first_timestamp(e.timestamp);
+                       }
+                       c.set_opening_num(e.opening_num);
+                       c.set_pgn_file_num(e.file_num);
+                       c.set_pgn_start_position(e.start_position);
+               }
+               if (!moves.count(e.move)) {
+                       moves.insert(e.move);
+                       c.add_move(e.move);
+               }
+               if (i == shard->elems.size() - 1 ||
+                   e.bpfen_len != shard->elems[i + 1].bpfen_len ||
+                   memcmp(e.bpfen, shard->elems[i + 1].bpfen, e.bpfen_len) != 0) {
+                       c.SerializeToString(&buf);
+                       mtbl_writer_add(mtbl,
+                               (const uint8_t *)e.bpfen, e.bpfen_len,
+                               (const uint8_t *)buf.data(), buf.size());
+                       c = Count();
+                       moves.clear();
+               }
+       }
+       mtbl_writer_destroy(&mtbl);
+
+       shard->elems.clear();
+       shard->arena.reset(new Arena);
+}
+
 int main(int argc, char **argv)
 {
-       int num_buckets = atoi(argv[argc - 1]);
+       int num_buckets = atoi(argv[argc - 2]);
+       size_t num_pos_per_subshard = atoi(argv[argc - 1]);  // 500000 is a reasonable value.
 
-       vector<vector<Element>> elems;
-       elems.resize(num_buckets);
+       vector<ShardData> shards;
+       shards.resize(num_buckets);
+
+       for (int i = 0; i < num_buckets; ++i) {
+               shards[i].elems.reserve(num_pos_per_subshard);
+               shards[i].arena.reset(new Arena);
+       }
 
        size_t num_elems = 0;
-       for (int i = 1; i < argc - 2; ++i) {
-               FILE *fp = fopen(argv[i], "rb");
-               if (fp == NULL) {
-                       perror(argv[i]);
-                       exit(1);
+       for (int i = 1; i < argc - 3; ++i) {
+               FILE *fp;
+               if (strcmp(argv[i], "-") == 0) {
+                       fp = stdin;
+               } else {
+                       fp = fopen(argv[i], "rb");
+                       if (fp == NULL) {
+                               perror(argv[i]);
+                               exit(1);
+                       }
                }
                for ( ;; ) {
+                       char bpfen[256];
+
                        int bpfen_len = getc(fp);
                        if (bpfen_len == -1) {
                                break;
                        }
-               
-                       char *bpfen = arena.alloc(bpfen_len);
+                       if (bpfen_len >= int(sizeof(bpfen))) {
+                               fprintf(stderr, "Overlong BPFEN (%d bytes)\n", bpfen_len);
+               //              exit(1);
+                               break;
+                       }
                        if (fread(bpfen, bpfen_len, 1, fp) != 1) {
                                perror("fread()");
                //              exit(1);
@@ -133,7 +213,8 @@ int main(int argc, char **argv)
 
                        int bucket = hash_key_to_bucket(bpfen, bpfen_len, num_buckets);
                        Element e;
-                       e.bpfen = bpfen;
+                       e.bpfen = shards[bucket].arena->alloc(bpfen_len);
+                       memcpy(e.bpfen, bpfen, bpfen_len);
                        e.bpfen_len = bpfen_len;
                        strcpy(e.move, move);
                        e.result = Result(r);
@@ -143,67 +224,21 @@ int main(int argc, char **argv)
                        e.file_num = file_num;
                        e.timestamp = timestamp;
                        e.start_position = start_position;
-                       elems[bucket].push_back(e);
+                       shards[bucket].elems.push_back(e);
                        ++num_elems;
+
+                       if (shards[bucket].elems.size() >= num_pos_per_subshard) {
+                               write_subshard(argv[argc - 3], &shards[bucket], bucket);
+                               shards[bucket].elems.reserve(num_pos_per_subshard);
+                       }
                }
                fclose(fp);
 
                printf("Read %ld elems\n", num_elems);
        }
 
-       printf("Sorting...\n");
-       for (int i = 0; i < num_buckets; ++i) {
-               sort(elems[i].begin(), elems[i].end());
-       }
-
-       printf("Writing SSTables...\n");
-       string buf;  // Keep allocated.
        for (int i = 0; i < num_buckets; ++i) {
-               char filename[256];
-               snprintf(filename, sizeof(filename), "%s.part%04d", argv[argc - 2], i);
-
-               mtbl_writer_options* wopt = mtbl_writer_options_init();
-               mtbl_writer_options_set_compression(wopt, MTBL_COMPRESSION_SNAPPY);
-               mtbl_writer* mtbl = mtbl_writer_init(filename, wopt);
-               Count c;
-               unordered_set<string> moves;
-               for (size_t j = 0; j < elems[i].size(); ++j) {
-                       const Element &e = elems[i][j];
-                       if (e.result == WHITE) {
-                               c.set_white(c.white() + 1);
-                       } else if (e.result == DRAW) {
-                               c.set_draw(c.draw() + 1);
-                       } else if (e.result == BLACK) {
-                               c.set_black(c.black() + 1);
-                       }
-                       if (e.white_elo >= 100 && e.black_elo >= 100) {
-                               c.set_sum_white_elo(c.sum_white_elo() + e.white_elo);
-                               c.set_sum_black_elo(c.sum_black_elo() + e.black_elo);
-                               c.set_num_elo(c.num_elo() + 1);
-                       }
-                       if (!c.has_first_timestamp() || e.timestamp < c.first_timestamp()) {
-                               if (e.timestamp != DUMMY_TIMESTAMP) {
-                                       c.set_first_timestamp(e.timestamp);
-                               }
-                               c.set_opening_num(e.opening_num);
-                               c.set_pgn_file_num(e.file_num);
-                               c.set_pgn_start_position(e.start_position);
-                       }
-                       if (!moves.count(e.move)) {
-                               moves.insert(e.move);
-                               c.add_move(e.move);
-                       }
-                       if (j == elems[i].size() - 1 ||
-                           e.bpfen_len != elems[i][j + 1].bpfen_len ||
-                           memcmp(e.bpfen, elems[i][j + 1].bpfen, e.bpfen_len) != 0) {
-                               c.SerializeToString(&buf);
-                               mtbl_writer_add(mtbl,
-                                       (const uint8_t *)e.bpfen, e.bpfen_len,
-                                       (const uint8_t *)buf.data(), buf.size());
-                               c = Count();
-                               moves.clear();
-                       }
-               }
-               mtbl_writer_destroy(&mtbl);
+               write_subshard(argv[argc - 3], &shards[i], i);
        }
 }
+

diff --git a/build-book.sh b/build-book.sh
index 60d5710ed481eda49ee4e372bbf778778688e1dd..67181c68167bdf555bb4157d05fc5c2aa829dbc8 100755 (executable)
--- a/build-book.sh
+++ b/build-book.sh
@@ -3,29 +3,32 @@
 set -e
 
 export SHARDS=40
-export PARALLEL_LOADS=15  # Reduce if you have problems with OOM
 export PARALLEL_MERGES=40
 
-rm -f pgnnames.txt part-*.bin part-*.mtbl part-*.mtbl.part???? open.mtbl.new open.mtbl.part???? open.mtbl.part????.new 2>/dev/null
+rm -f pgnnames.txt tmp.mtbl* open.mtbl.new open.mtbl.part???? open.mtbl.part????.new 2>/dev/null
 
-PGNNUM=0
 for FILE in "$@"; do
-       date | tr -d "\n"
-       echo "  $FILE"
-       ./parallel-parse-pgn.sh "$FILE" "$PGNNUM"
        echo "$FILE" >> pgnnames.txt
-       PGNNUM=$(( PGNNUM + 1 ))
 done
-date 
 
-echo "Loading..."
-parallel -i -j ${PARALLEL_LOADS} bash -c 'FILE="{}"; set -x; ./binloader "$FILE" "${FILE/bin/mtbl}" $SHARDS' -- part-*.bin
-
-rm -f part-*.bin
+for X in $( seq 0 39 ); do
+       (
+               (
+                       PGNNUM=0
+                       for FILE in "$@"; do
+                               START=$( ./find-pgn-split-point.sh "$FILE" $X 40 )
+                               END=$( ./find-pgn-split-point.sh "$FILE" $(( X + 1 )) 40 )
+                               ~/nmu/pgn-extract/pgn-extract --startpos $START --endpos $END --startfilenum $PGNNUM -e -Wsessebin "$FILE" 2>/dev/null
+                               PGNNUM=$(( PGNNUM + 1 ))
+                       done
+               ) | ./binloader - tmp.mtbl.mapper$X $SHARDS 500000
+       ) &
+done
+wait
 
-parallel -i -j ${PARALLEL_MERGES} bash -c 'X={}; set -x; ./binmerger part-*.mtbl.part$( printf %04d $X ) open.mtbl.part$( printf %04d $X ).new' -- $( seq 0 $(( SHARDS - 1 )) )
+parallel -i -j ${PARALLEL_MERGES} bash -c 'X={}; set -x; ./binmerger tmp.mtbl.mapper*.part$( printf %04d $X ).subshard* open.mtbl.part$( printf %04d $X ).new' -- $( seq 0 $(( SHARDS - 1 )) )
 
 for X in $( seq 0 $(( SHARDS - 1 )) ); do
        mv open.mtbl.part$( printf %04d $X ).new open.mtbl.part$( printf %04d $X )
 done
-rm -f part-*.mtbl.part????
+rm -f tmp.mtbl*