jass.c

// #define TIME_EACH_POSTINGS_SEGMENT 1
#define PRINT_PER_QUERY_STATS 1
// #define TIME_EACH_QUERY_PHASE 1
#define ANYTIME 1

#include <chrono>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#ifdef _MSC_VER
#include <intrin.h>
#include <windows.h>
#else
#ifdef __APPLE__
#include <mach/mach.h>
#include <mach/mach_time.h>
#endif
#include <unistd.h>
#include <glob.h>
#include <sys/times.h>
#include <dlfcn.h>
#endif
#ifdef __linux__
#include <sys/time.h>
#endif

#include <sys/types.h>
#include <sys/stat.h>
#include <emmintrin.h>
#include <smmintrin.h>

#include "CI.h"
#include "compress_qmx.h"
#include "compress_qmx_d4.h"
#include "compress_simple8b.h"
#include "process_postings.h"

using namespace std;

uint16_t *CI_accumulators;              // the accumulators
uint16_t **CI_accumulator_pointers;     // an array of pointers into the accumulators (used to avoid computing docIDs)
uint32_t CI_top_k;                      // the number of results to find (top-k)
uint32_t CI_results_list_length;        // the number of results we found (at most top-k)

uint8_t *CI_accumulator_clean_flags;    // is the "row" of the accumulator table
uint32_t CI_accumulators_shift;         // number of bits to shift (right) the docid by to get the CI_accumulator_clean_flags
uint32_t CI_accumulators_width;         // the "width" of the accumulator table
uint32_t CI_accumulators_height;        // the "height" of the accumulator table
ANT_heap<uint16_t *, add_rsv_compare> *CI_heap;

char **CI_documentlist;                 // the list of document IDs (TREC document IDs)

CI_vocab_heap *CI_dictionary;           // the vocab array
uint32_t CI_unique_terms;
uint32_t CI_unique_documents;

#define MAX_TERMS_PER_QUERY 10
#define MAX_QUERIES 1000
#define MAX_QUANTUM 0xFF

#ifdef _MSC_VER
#define atoll(x) _atoi64(x)
double log2(double n) {return log(n) / log(2.0);}
#endif

void print_os_time(void) {
#ifdef __APPLE__
  struct tms tmsbuf;
  long clock_speed = sysconf(_SC_CLK_TCK);

  if (times(&tmsbuf) > 0) {
    printf("OS reports kernel time: %.3f seconds\n", (double)tmsbuf.tms_stime / clock_speed);
    printf("OS reports user time  : %.3f seconds\n", (double)tmsbuf.tms_utime / clock_speed);
  }
#endif
}

void trec_dump_results(uint32_t topic_id, FILE *out, uint32_t output_length) {
  uint32_t current, id;

  for (current = 0; current < (output_length < CI_results_list_length ? output_length : CI_results_list_length); current++) {
    id = CI_accumulator_pointers[current] - CI_accumulators;
    fprintf(out, "%d Q0 %s %d %d COMPILED (ID:%u)\n", topic_id, CI_documentlist[id], current + 1, CI_accumulators[id], id);
  }
}

char *read_entire_file(const char *filename, uint64_t *length = NULL) {
  char *block = NULL;
  FILE *fp;
  struct stat details;

  if (filename == NULL) {
    return NULL;
  }

  if ((fp = fopen(filename, "rb")) == NULL) {
    return NULL;
  }

  if (fstat(fileno(fp), &details) == 0) {
    if (details.st_size != 0) {
      if ((block = new char [(size_t)(details.st_size + 1)]) != NULL) { // +1 for the '\0' on the end
        if (length != NULL) {
          *length = details.st_size;
        }

        if (fread(block, details.st_size, 1, fp) == 1) {
          block[details.st_size] = '\0';
        } else {
          delete [] block;
          block = NULL;
        }
      }
    }
  }
  fclose(fp);

  return block;
}

class CI_quantum_header {
public:
  uint16_t impact;
  uint64_t offset;
  uint64_t end;
  uint32_t quantum_frequency;
}__attribute__((packed));

int quantum_compare(const void *a, const void *b) {
  CI_quantum_header *lhs = (CI_quantum_header *) ((*(uint64_t *) a) + postings);
  CI_quantum_header *rhs = (CI_quantum_header *) ((*(uint64_t *) b) + postings);

  /*
   sort from highest to lowest impact, but break ties by placing the lowest quantum-frequency first and the highest quantum-drequency last
   */
  return lhs->impact > rhs->impact ? -1 : lhs->impact < rhs->impact ? 1 :
         lhs->quantum_frequency > rhs->quantum_frequency ? 1 :
         lhs->quantum_frequency == rhs->quantum_frequency ? 0 : -1;
}

void print_postings_list(CI_vocab_heap *postings_list) {
  CI_quantum_header *header;

  uint32_t current;
  uint64_t *data;

  printf("\n\noffset:0x%llX\n", postings_list->offset);
  fflush(stdout);
  printf("impacts:%llu\n", postings_list->impacts);
  for (uint32_t x = 0; x < 64; x++) {
    printf("%02X ", *(postings + postings_list->offset + x));
  }
  puts("");

  data = (uint64_t *) (postings + postings_list->offset);
  for (current = 0; current < postings_list->impacts; current++) {
    header = (CI_quantum_header *) (postings + (data[current]));
    printf("OFFSET:%llx Impact:%hx Offset:%llx End:%llx docids:%u\n", data[current], header->impact,
        header->offset, header->end, header->quantum_frequency);

    for (uint8_t *byte = postings + header->offset; byte < postings + header->end; byte++) {
      printf("0x%02X, ", *byte);
      if (*byte & 0x80) {
        printf("\n");
      }
    }
    puts("");
  }
}

void read_doclist(void) {
  char *doclist;
  uint64_t total_docs;
  uint64_t length;
  uint64_t *offset_base;

  printf("Loading doclist... ");
  fflush(stdout);
  if ((doclist = read_entire_file("CIdoclist.bin", &length)) == 0) {
    exit(printf("Can't read CIdoclist.bin"));
  }

  CI_unique_documents = total_docs = *((uint64_t *) (doclist + length - sizeof(uint64_t)));
  CI_documentlist = new char * [total_docs];

  offset_base = (uint64_t *) (doclist + length - (total_docs * sizeof(uint64_t) + sizeof(uint64_t)));
  for (uint64_t id = 0; id < total_docs; id++) {
    CI_documentlist[id] = doclist + offset_base[id];
  }

  puts("done");
  fflush(stdout);
}

void read_vocab(void) {
  char *vocab, *vocab_terms;
  uint64_t total_terms;
  uint64_t length;
  uint64_t *base;
  uint64_t term;

  printf("Loading vocab... ");
  fflush(stdout);
  if ((vocab = read_entire_file("CIvocab.bin", &length)) == NULL) {
    exit(printf("Can't read CIvocab.bin"));
  }
  if ((vocab_terms = read_entire_file("CIvocab_terms.bin")) == NULL) {
    exit(printf("Can't read CIvocab_terms.bin"));
  }

  CI_unique_terms = total_terms = length / (sizeof(uint64_t) + sizeof(uint64_t) + sizeof(uint64_t));

  CI_dictionary = new CI_vocab_heap[total_terms];
  for (term = 0; term < total_terms; term++) {
    base = (uint64_t *) (vocab + (3 * sizeof(uint64_t)) * term);

    CI_dictionary[term].term = vocab_terms + base[0];
    CI_dictionary[term].offset = base[1];
    CI_dictionary[term].impacts = base[2];
  }

  puts("done");
  fflush(stdout);
}

int main(int argc, char *argv[]) {
  chrono::time_point<chrono::steady_clock, chrono::nanoseconds> full_query_timer = chrono::steady_clock::now();
  chrono::time_point<chrono::steady_clock, chrono::nanoseconds> start_timer, end_timer, full_query_without_io_timer;
  chrono::nanoseconds stats_total_time_to_search;
  chrono::nanoseconds stats_total_time_to_search_without_io;
  chrono::nanoseconds stats_tmp;
#ifdef TIME_EACH_QUERY_PHASE
  chrono::nanoseconds stats_accumulator_time;
  chrono::nanoseconds stats_vocab_time;
  chrono::nanoseconds stats_postings_time;
  chrono::nanoseconds stats_sort_time;
  chrono::nanoseconds stats_quantum_prep_time;
  chrono::nanoseconds stats_early_terminate_check_time;
#endif

  static char buffer[1024];
  const char *SEPERATORS = " \t\r\n";
  int i;
  FILE *fp, *out;
  char *term, *id;
  uint64_t query_id;
  CI_vocab_heap *postings_list;

  uint32_t postings_to_process;
  uint32_t postings_processed = 0;
  uint64_t total_number_of_topics;
  uint32_t accumulators_needed;
  uint64_t stats_quantum_check_count, stats_quantum_count, stats_early_terminations;
  uint64_t experimental_repeat = 0, times_to_repeat_experiment = 1;
  uint64_t *quantum_order, *current_quantum;
  uint64_t max_remaining_impact;
  uint16_t **quantum_check_pointers;
  uint64_t early_terminate;
  uint16_t **partial_rsv;
  CI_quantum_header *current_header;
  void (*process_postings_list)(uint8_t *doclist, uint8_t *end, uint16_t impact, uint32_t integers);
  uint32_t parameter, decompress_then_process;

  uint64_t queries_id[MAX_QUERIES];
  uint32_t queries_num_postings[MAX_QUERIES];
  uint64_t queries_latency[MAX_QUERIES];

  /*
   Parameter parsing
   */
  CI_top_k = CI_unique_documents + 1;
  decompress_then_process = false;

#ifdef ANYTIME
  if (argc < 1 || argc > 5) {
    exit(printf("Usage:%s <queryfile> [<top-k-number>] [<num-postings-to-proces>] [-d<ecompress then process>]\n", argv[0]));
  }

  for (parameter = 2; parameter < argc; parameter++) {
    if (strcmp(argv[parameter], "-d") == 0) {
      decompress_then_process = true;
    } else if (parameter == 2) {
      CI_top_k = atoll(argv[parameter]);
    } else if (parameter == 3) {
      postings_to_process = atoi(argv[parameter]);
    }
  }

  printf("Ranking mode: anytime\n");
  printf("Settings: top k = %d, postings to process = %d", CI_top_k, postings_to_process);
  if (decompress_then_process) {
    printf(", decompress then process (non-interleaved)\n");
  } else {
    printf(", interleaving decompression and postings processing\n");
  }

#else
  if (argc < 1 || argc > 4) {
    exit(printf("Usage:%s <queryfile> [<top-k-number>] [-d<ecompress then process>]\n", argv[0]));
  }

  for (parameter = 2; parameter < argc; parameter++) {
    if (strcmp(argv[parameter], "-d") == 0) {
      decompress_then_process = true;
    } else {
      CI_top_k = atoll(argv[parameter]);
    }
  }

  printf("Ranking mode: early termination\n");
  printf("Settings: top k = %d", CI_top_k);
  if (decompress_then_process) {
    printf(", decompress then process (non-interleaved)\n");
  } else {
    printf(", interleaving decompression and postings processing\n");
  }
#endif

  printf("Loading postings... ");
  fflush(stdout);
  if ((postings = (uint8_t *) read_entire_file("CIpostings.bin")) == NULL) {
    exit(printf("Cannot open postings file 'CIpostings.bin'\n"));
  }
  puts("done");
  fflush(stdout);

  if ((fp = fopen(argv[1], "r")) == NULL) {
    exit(printf("Can't open query file:%s\n", argv[1]));
  }

  if ((out = fopen("ranking.txt", "w")) == NULL) {
    exit(printf("Can't open output file.\n"));
  }

  /*
   Sort out how to decode the postings (either compressed or not)
   */
  if (*postings == 's') {
    puts("Uncompressed Index");
    process_postings_list = CIt_process_list_not_compressed;
  } else if (*postings == 'c') {
    puts("Variable Byte Compressed Index");
    if (decompress_then_process) {
      process_postings_list = CIt_process_list_decompress_then_process;
    } else {
      process_postings_list = CIt_process_list_compressed_vbyte;
    }
  } else if (*postings == '8') {
    puts("Simple-8b Compressed Index");
    if (decompress_then_process) {
      process_postings_list = CIt_process_list_compressed_simple8b_ATIRE;
    } else {
      process_postings_list = CIt_process_list_compressed_simple8b;
    }
  } else if (*postings == 'q') {
    puts("QMX Compressed Index");
    if (!decompress_then_process) {
      exit(printf("Cannot interleave QMX\n"));
    }
    process_postings_list = CIt_process_list_compressed_qmx;
  } else if (*postings == 'Q') {
    puts("QMX-D4 Compressed Index");
    if (!decompress_then_process) {
      exit(printf("Cannot interleave QMX\n"));
    }
    process_postings_list = CIt_process_list_compressed_qmx_d4;
  } else if (*postings == 'R') {
    puts("QMX-D0 Compressed Index");
    if (!decompress_then_process) {
      exit(printf("Cannot interleave QMX\n"));
    }
    process_postings_list = CIt_process_list_compressed_qmx_d0;
  } else {
    exit(printf("This index appears to be invalid as it is neither compressed nor not compressed!\n"));
  }

  read_doclist();
  read_vocab();

  /*
   Compute the details of the accumulator table
   */
  CI_accumulators_shift = log2(sqrt((double) CI_unique_documents));
  CI_accumulators_width = 1 << CI_accumulators_shift;
  CI_accumulators_height = (CI_unique_documents + CI_accumulators_width) / CI_accumulators_width;
  accumulators_needed = CI_accumulators_width * CI_accumulators_height;	// guaranteed to be larger than the highest accumulagtor that can be initialised
  CI_accumulator_clean_flags = new uint8_t[CI_accumulators_height];

  /*
   Create a buffer to store the decompressed postings lists
   */
  CI_decompressed_postings = new uint32_t[CI_unique_documents + 1024];	// we add because some decompressors are allowed to overflow this buffer

  /*
   Now prime the search engine
   */
  CI_accumulators = new uint16_t[accumulators_needed];
  CI_accumulator_pointers = new uint16_t * [accumulators_needed];

  /*
   For QaaT early termination we need K+1 elements in the heap so that we can check that nothing else can get into the top-k.
   */
  CI_top_k++;
  CI_heap = new ANT_heap<uint16_t *, add_rsv_compare>(*CI_accumulator_pointers, CI_top_k);

  /*
   Allocate the quantum at a time table
   */
  quantum_order = new uint64_t[MAX_TERMS_PER_QUERY * MAX_QUANTUM];
  quantum_check_pointers = new uint16_t * [accumulators_needed];

#ifdef TIME_EACH_POSTINGS_SEGMENT
  times_to_repeat_experiment = 1;
#endif

  /*
   Now start searching
   */
  while (experimental_repeat < times_to_repeat_experiment) {
    experimental_repeat++;

    stats_total_time_to_search = chrono::nanoseconds(0);
    stats_total_time_to_search_without_io = chrono::nanoseconds(0);
#ifdef TIME_EACH_QUERY_PHASE
    stats_accumulator_time = chrono::nanoseconds(0);
    stats_vocab_time = chrono::nanoseconds(0);
    stats_quantum_prep_time = chrono::nanoseconds(0);
    stats_early_terminate_check_time = chrono::nanoseconds(0);
    stats_postings_time = chrono::nanoseconds(0);
    stats_sort_time = chrono::nanoseconds(0);
#endif

    total_number_of_topics = 0;
    stats_quantum_check_count = 0;
    stats_quantum_count = 0;
    stats_early_terminations = 0;

    rewind(fp);					// the query file
    rewind(out);				// the TREC run file

    while (fgets(buffer, sizeof(buffer), fp) != NULL) {
      full_query_without_io_timer = chrono::steady_clock::now();
      if ((id = strtok(buffer, SEPERATORS)) == NULL) {
        continue;
      }

      total_number_of_topics++;
      CI_results_list_length = 0;

      query_id = atoll(id);

      // Initialize the accumulators
#ifdef TIME_EACH_QUERY_PHASE
      start_timer = chrono::steady_clock::now();
#endif
      memset(CI_accumulator_clean_flags, 0, CI_accumulators_height);
#ifdef TIME_EACH_QUERY_PHASE
      end_timer = chrono::steady_clock::now();
      stats_accumulator_time += chrono::duration_cast<chrono::nanoseconds>(end_timer - start_timer);
#endif

      max_remaining_impact = 0;
      current_quantum = quantum_order;
      early_terminate = false;

      while ((term = strtok(NULL, SEPERATORS)) != NULL) {
#ifdef TIME_EACH_QUERY_PHASE
        start_timer = chrono::steady_clock::now();
#endif
        postings_list = (CI_vocab_heap *) bsearch(term, CI_dictionary, CI_unique_terms, sizeof(*CI_dictionary), CI_vocab_heap::compare_string);
#ifdef TIME_EACH_QUERY_PHASE
        end_timer = chrono::steady_clock::now();
        stats_vocab_time += chrono::duration_cast<chrono::nanoseconds>(end_timer - start_timer);
#endif

        // Initializee the SaaT structures
#ifdef TIME_EACH_QUERY_PHASE
        start_timer = chrono::steady_clock::now();
#endif
        if (postings_list != NULL) {
          // Copy this term's pointers to the segments list
          memcpy(current_quantum, postings + postings_list->offset, postings_list->impacts * sizeof(*quantum_order));

          // Compute the maximum possibe impact score (that is, assume one document has the maximum impact for each term)
          max_remaining_impact += ((CI_quantum_header *) (postings + *current_quantum))->impact;

          // Advance to the place we want to place the next set of segments
          current_quantum += postings_list->impacts;
        }
      }
      /*
       NULL termainate the list of quantums
       */
      *current_quantum = 0;

      /*
       Sort the quantum list from highest to lowest
       */
      qsort(quantum_order, current_quantum - quantum_order, sizeof(*quantum_order), quantum_compare);

#ifdef TIME_EACH_QUERY_PHASE
      end_timer = chrono::steady_clock::now();
      stats_quantum_prep_time += chrono::duration_cast<chrono::nanoseconds>(end_timer - start_timer);
#endif
      /*
       Now process each quantum, one at a time
       */

#ifdef ANYTIME
      postings_processed = 0;
      for (current_quantum = quantum_order; *current_quantum != 0; current_quantum++) {
        current_header = (CI_quantum_header *) (postings + *current_quantum);
        if (postings_processed + current_header->quantum_frequency > postings_to_process) {
          break;
        }

        stats_quantum_count++;
#ifdef TIME_EACH_QUERY_PHASE
        start_timer = chrono::steady_clock::now();
#endif
        (*process_postings_list)(postings + current_header->offset, postings + current_header->end,
            current_header->impact, current_header->quantum_frequency);
#ifdef TIME_EACH_QUERY_PHASE
        end_timer = chrono::steady_clock::now();
        stats_tmp = chrono::duration_cast<chrono::nanoseconds>(end_timer - start_timer);
        stats_postings_time += stats_tmp;
#endif
        postings_processed += current_header->quantum_frequency;

#ifdef TIME_EACH_POSTINGS_SEGMENT
        printf("I:%lld L:%lld T:%lld\n", (long long)current_header->impact, (long long)current_header->quantum_frequency, stats_tmp);
#endif

      }

#else

      for (current_quantum = quantum_order; *current_quantum != 0; current_quantum++) {
        stats_quantum_count++;

        current_header = (CI_quantum_header *)(postings + *current_quantum);
#ifdef TIME_EACH_QUERY_PHASE
        start_timer = chrono::steady_clock::now();
#endif
        (*process_postings_list)(postings + current_header->offset, postings + current_header->end, current_header->impact, current_header->quantum_frequency);
#ifdef TIME_EACH_QUERY_PHASE
        end_timer = chrono::steady_clock::now();
        stats_tmp = chrono::duration_cast<chrono::nanoseconds>(end_timer - start_timer);
        stats_postings_time += stats_tmp;
#endif

#ifdef TIME_EACH_POSTINGS_SEGMENT
        printf("I:%lld L:%lld T:%lld\n", (long long)current_header->impact, (long long)current_header->quantum_frequency, stats_tmp);
#endif

        // Check to see if it's posible for the remaining impacts to affect the order of the top-k.
#ifdef TIME_EACH_QUERY_PHASE
        start_timer = chrono::steady_clock::now();
#endif
        // Subtract the current impact score and then add the next impact score for the current term.
        max_remaining_impact -= current_header->impact;
        max_remaining_impact += (current_header + 1)->impact;

        if (CI_results_list_length > CI_top_k - 1) {
          stats_quantum_check_count++;
          /*
           We need to run through the top-(k+1) to see if its possible for any re-ordering to occur
           1. copy the accumulators;
           2. sort the top k + 1
           3. go through consequative rsvs checking to see if reordering is possible (check rsv[k] - rsv[k + 1])
           */
          memcpy(quantum_check_pointers, CI_accumulator_pointers, CI_top_k * sizeof(*quantum_check_pointers));
          top_k_qsort(quantum_check_pointers, CI_top_k, CI_top_k);

          early_terminate = true;

          for (partial_rsv = quantum_check_pointers; partial_rsv < quantum_check_pointers + CI_top_k - 1; partial_rsv++) {
            if (**partial_rsv - **(partial_rsv + 1) < max_remaining_impact) { // We're sorted from largest to smallest so a[x] - a[x+1] >= 0
              early_terminate = false;
              break;
            }
          }
        }
#ifdef TIME_EACH_QUERY_PHASE
        end_timer = chrono::steady_clock::now();
        stats_early_terminate_check_time += chrono::duration_cast<chrono::nanoseconds>(end_timer - start_timer);
#endif
        if (early_terminate) {
          stats_early_terminations++;
          break;
        }
      }

#endif

      // Sort the accumulator pointers to put the highest RSV document at the top of the list.
#ifdef TIME_EACH_QUERY_PHASE
      start_timer = chrono::steady_clock::now();
#endif
      top_k_qsort(CI_accumulator_pointers, CI_results_list_length, CI_top_k - 1);
#ifdef TIME_EACH_QUERY_PHASE
      end_timer = chrono::steady_clock::now();
      stats_sort_time += chrono::duration_cast<chrono::nanoseconds>(end_timer - start_timer);
#endif

      /*
       At this point we know the number of hits (CI_results_list_length) and they can be decode out of the CI_accumulator_pointers array
       where CI_accumulator_pointers[0] points into CI_accumulators[] and therefore CI_accumulator_pointers[0] - CI_accumulators is the docid
       and *CI_accumulator_pointers[0] is the rsv.
       */
      end_timer = chrono::steady_clock::now();
      stats_tmp = chrono::duration_cast<chrono::nanoseconds>(end_timer - full_query_without_io_timer);
      stats_total_time_to_search_without_io += stats_tmp;

#ifdef PRINT_PER_QUERY_STATS
      queries_id[total_number_of_topics-1] = query_id;
      queries_num_postings[total_number_of_topics-1] = postings_processed;
      queries_latency[total_number_of_topics-1] = stats_tmp.count();
#endif

      // Dump TREC output
      trec_dump_results(query_id, out, CI_top_k - 1);	// Subtract 1 from top_k because we added 1 for the early termination checks.
    }
  }

  fclose(out);
  fclose(fp);

  end_timer = chrono::steady_clock::now();
  stats_total_time_to_search += chrono::duration_cast<chrono::nanoseconds>(end_timer - full_query_timer);
  print_os_time();

  printf("\nAverages over %llu queries\n", total_number_of_topics);
  printf("--------------------------------------------------------\n");
#ifdef TIME_EACH_QUERY_PHASE
  printf("Accumulator initialization (per query) : %12llu ns\n", stats_accumulator_time.count() / total_number_of_topics);
  printf("Vocabulary lookup          (per query) : %12llu ns\n", stats_vocab_time.count() / total_number_of_topics);
  printf("SaaT prep time             (per query) : %12llu ns\n", stats_quantum_prep_time.count() / total_number_of_topics);
  printf("Processing postings        (per query) : %12llu ns\n", stats_postings_time.count() / total_number_of_topics);
#ifndef ANYTIME
  printf("SaaT early terminate check (per query) : %12llu ns\n", stats_early_terminate_check_time.count() / total_number_of_topics);
#endif
  printf("Extracting top-k           (per query) : %12llu ns\n", stats_sort_time.count() / total_number_of_topics);
#endif
  printf("Total time excluding I/O   (per query) : %12llu ns\n", stats_total_time_to_search_without_io.count() / total_number_of_topics);
  printf("--------------------------------------------------------\n");
  printf("Total running time                     : %12llu ns\n", stats_total_time_to_search.count());
  printf("--------------------------------------------------------\n");
#ifndef ANYTIME
  printf("Total number of early terminate checks : %12llu\n", stats_quantum_check_count);
  printf("Total number of early terminations     : %12llu\n", stats_early_terminations);
#endif
  printf("Total number of segments processed     : %12llu\n", stats_quantum_count);
  printf("--------------------------------------------------------\n");

#ifdef PRINT_PER_QUERY_STATS
  for (i=0; i<total_number_of_topics; i++) {
    printf("query id,postings processed,ns:%llu,%d,%llu\n", queries_id[i], queries_num_postings[i], queries_latency[i]);
  }
#endif

  return 0;
}