initial code repo
[stor4nfv.git] / src / ceph / src / crush / CrushTester.cc
diff --git a/src/ceph/src/crush/CrushTester.cc b/src/ceph/src/crush/CrushTester.cc
new file mode 100644 (file)
index 0000000..b6b4b2f
--- /dev/null
@@ -0,0 +1,725 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#include "include/stringify.h"
+#include "CrushTester.h"
+#include "CrushTreeDumper.h"
+#include "include/ceph_features.h"
+
+#include <algorithm>
+#include <stdlib.h>
+#include <boost/lexical_cast.hpp>
+// to workaround https://svn.boost.org/trac/boost/ticket/9501
+#ifdef _LIBCPP_VERSION
+#include <boost/version.hpp>
+#if BOOST_VERSION < 105600
+#define ICL_USE_BOOST_MOVE_IMPLEMENTATION
+#endif
+#endif
+#include <boost/icl/interval_map.hpp>
+#include <boost/algorithm/string/join.hpp>
+#include "common/SubProcess.h"
+#include "common/fork_function.h"
+
+void CrushTester::set_device_weight(int dev, float f)
+{
+  int w = (int)(f * 0x10000);
+  if (w < 0)
+    w = 0;
+  if (w > 0x10000)
+    w = 0x10000;
+  device_weight[dev] = w;
+}
+
+int CrushTester::get_maximum_affected_by_rule(int ruleno)
+{
+  // get the number of steps in RULENO
+  int rule_size = crush.get_rule_len(ruleno);
+  vector<int> affected_types;
+  map<int,int> replications_by_type;
+
+  for (int i = 0; i < rule_size; i++){
+    // get what operation is done by the current step
+    int rule_operation = crush.get_rule_op(ruleno, i);
+
+    // if the operation specifies choosing a device type, store it
+    if (rule_operation >= 2 && rule_operation != 4){
+      int desired_replication = crush.get_rule_arg1(ruleno,i);
+      int affected_type = crush.get_rule_arg2(ruleno,i);
+      affected_types.push_back(affected_type);
+      replications_by_type[affected_type] = desired_replication;
+    }
+  }
+
+  /*
+   * now for each of the affected bucket types, see what is the
+   * maximum we are (a) requesting or (b) have
+   */
+
+  map<int,int> max_devices_of_type;
+
+  // loop through the vector of affected types
+  for (vector<int>::iterator it = affected_types.begin(); it != affected_types.end(); ++it){
+    // loop through the number of buckets looking for affected types
+    for (map<int,string>::iterator p = crush.name_map.begin(); p != crush.name_map.end(); ++p){
+      int bucket_type = crush.get_bucket_type(p->first);
+      if ( bucket_type == *it)
+        max_devices_of_type[*it]++;
+    }
+  }
+
+  for(std::vector<int>::iterator it = affected_types.begin(); it != affected_types.end(); ++it){
+    if ( replications_by_type[*it] > 0 && replications_by_type[*it] < max_devices_of_type[*it] )
+      max_devices_of_type[*it] = replications_by_type[*it];
+  }
+
+  /*
+   * get the smallest number of buckets available of any type as this is our upper bound on
+   * the number of replicas we can place
+  */
+  int max_affected = max( crush.get_max_buckets(), crush.get_max_devices() );
+
+  for(std::vector<int>::iterator it = affected_types.begin(); it != affected_types.end(); ++it){
+    if (max_devices_of_type[*it] > 0 && max_devices_of_type[*it] < max_affected )
+      max_affected = max_devices_of_type[*it];
+  }
+
+  return max_affected;
+}
+
+
+map<int,int> CrushTester::get_collapsed_mapping()
+{
+  int num_to_check = crush.get_max_devices();
+  int next_id = 0;
+  map<int, int> collapse_mask;
+
+  for (int i = 0; i < num_to_check; i++){
+    if (crush.check_item_present(i)){
+      collapse_mask[i] = next_id;
+      next_id++;
+    }
+  }
+  
+  return collapse_mask;
+}
+
+void CrushTester::adjust_weights(vector<__u32>& weight)
+{
+
+  if (mark_down_device_ratio > 0) {
+    // active buckets
+    vector<int> bucket_ids;
+    for (int i = 0; i < crush.get_max_buckets(); i++) {
+      int id = -1 - i;
+      if (crush.get_bucket_weight(id) > 0) {
+        bucket_ids.push_back(id);
+      }
+    }
+
+    // get buckets that are one level above a device
+    vector<int> buckets_above_devices;
+    for (unsigned i = 0; i < bucket_ids.size(); i++) {
+      // grab the first child object of a bucket and check if it's ID is less than 0
+      int id = bucket_ids[i];
+      if (crush.get_bucket_size(id) == 0)
+        continue;
+      int first_child = crush.get_bucket_item(id, 0); // returns the ID of the bucket or device
+      if (first_child >= 0) {
+        buckets_above_devices.push_back(id);
+      }
+    }
+
+    // permute bucket list
+    for (unsigned i = 0; i < buckets_above_devices.size(); i++) {
+      unsigned j = lrand48() % (buckets_above_devices.size() - 1);
+      std::swap(buckets_above_devices[i], buckets_above_devices[j]);
+    }
+
+    // calculate how many buckets and devices we need to reap...
+    int num_buckets_to_visit = (int) (mark_down_bucket_ratio * buckets_above_devices.size());
+
+    for (int i = 0; i < num_buckets_to_visit; i++) {
+      int id = buckets_above_devices[i];
+      int size = crush.get_bucket_size(id);
+      vector<int> items;
+      for (int o = 0; o < size; o++)
+        items.push_back(crush.get_bucket_item(id, o));
+
+      // permute items
+      for (int o = 0; o < size; o++) {
+        int j = lrand48() % (crush.get_bucket_size(id) - 1);
+        std::swap(items[o], items[j]);
+      }
+
+      int local_devices_to_visit = (int) (mark_down_device_ratio*size);
+      for (int o = 0; o < local_devices_to_visit; o++){
+        int item = crush.get_bucket_item(id, o);
+        weight[item] = 0;
+      }
+    }
+  }
+}
+
+bool CrushTester::check_valid_placement(int ruleno, vector<int> in, const vector<__u32>& weight)
+{
+
+  bool valid_placement = true;
+  vector<int> included_devices;
+  map<string,string> seen_devices;
+
+  // first do the easy check that all devices are "up"
+  for (vector<int>::iterator it = in.begin(); it != in.end(); ++it) {
+    if (weight[(*it)] == 0) {
+      valid_placement = false;
+      break;
+    } else if (weight[(*it)] > 0) {
+      included_devices.push_back( (*it) );
+    }
+  }
+
+  /*
+   * now do the harder test of checking that the CRUSH rule r is not violated
+   * we could test that none of the devices mentioned in out are unique,
+   * but this is a special case of this test
+   */
+
+  // get the number of steps in RULENO
+  int rule_size = crush.get_rule_len(ruleno);
+  vector<string> affected_types;
+
+  // get the smallest type id, and name
+  int min_map_type = crush.get_num_type_names();
+  for (map<int,string>::iterator it = crush.type_map.begin(); it != crush.type_map.end(); ++it ) {
+    if ( (*it).first < min_map_type ) {
+      min_map_type = (*it).first;
+    }
+  }
+
+  string min_map_type_name = crush.type_map[min_map_type];
+
+  // get the types of devices affected by RULENO
+  for (int i = 0; i < rule_size; i++) {
+    // get what operation is done by the current step
+    int rule_operation = crush.get_rule_op(ruleno, i);
+
+    // if the operation specifies choosing a device type, store it
+    if (rule_operation >= 2 && rule_operation != 4) {
+      int affected_type = crush.get_rule_arg2(ruleno,i);
+      affected_types.push_back( crush.get_type_name(affected_type));
+    }
+  }
+
+  // find in if we are only dealing with osd's
+  bool only_osd_affected = false;
+  if (affected_types.size() == 1) {
+    if ((affected_types.back() == min_map_type_name) && (min_map_type_name == "osd")) {
+      only_osd_affected = true;
+    }
+  }
+
+  // check that we don't have any duplicate id's
+  for (vector<int>::iterator it = included_devices.begin(); it != included_devices.end(); ++it) {
+    int num_copies = std::count(included_devices.begin(), included_devices.end(), (*it) );
+    if (num_copies > 1) {
+      valid_placement = false;
+    }
+  }
+
+  // if we have more than just osd's affected we need to do a lot more work
+  if (!only_osd_affected) {
+    // loop through the devices that are "in/up"
+    for (vector<int>::iterator it = included_devices.begin(); it != included_devices.end(); ++it) {
+      if (valid_placement == false)
+        break;
+
+      // create a temporary map of the form (device type, device name in map)
+      map<string,string> device_location_hierarchy = crush.get_full_location(*it);
+
+      // loop over the types affected by RULENO looking for duplicate bucket assignments
+      for (vector<string>::iterator t = affected_types.begin(); t != affected_types.end(); ++t) {
+        if (seen_devices.count( device_location_hierarchy[*t])) {
+          valid_placement = false;
+          break;
+        } else {
+          // store the devices we have seen in the form of (device name, device type)
+          seen_devices[ device_location_hierarchy[*t] ] = *t;
+        }
+      }
+    }
+  }
+
+  return valid_placement;
+}
+
+int CrushTester::random_placement(int ruleno, vector<int>& out, int maxout, vector<__u32>& weight)
+{
+  // get the total weight of the system
+  int total_weight = 0;
+  for (unsigned i = 0; i < weight.size(); i++)
+    total_weight += weight[i];
+
+  if (total_weight == 0 ||
+      crush.get_max_devices() == 0)
+    return -EINVAL;
+
+  // determine the real maximum number of devices to return
+  int devices_requested = min(maxout, get_maximum_affected_by_rule(ruleno));
+  bool accept_placement = false;
+
+  vector<int> trial_placement(devices_requested);
+  int attempted_tries = 0;
+  int max_tries = 100;
+  do {
+    // create a vector to hold our trial mappings
+    int temp_array[devices_requested];
+    for (int i = 0; i < devices_requested; i++){
+      temp_array[i] = lrand48() % (crush.get_max_devices());
+    }
+
+    trial_placement.assign(temp_array, temp_array + devices_requested);
+    accept_placement = check_valid_placement(ruleno, trial_placement, weight);
+    attempted_tries++;
+  } while (accept_placement == false && attempted_tries < max_tries);
+
+  // save our random placement to the out vector
+  if (accept_placement)
+    out.assign(trial_placement.begin(), trial_placement.end());
+
+  // or don't....
+  else if (attempted_tries == max_tries)
+    return -EINVAL;
+
+  return 0;
+}
+
+void CrushTester::write_integer_indexed_vector_data_string(vector<string> &dst, int index, vector<int> vector_data)
+{
+  stringstream data_buffer (stringstream::in | stringstream::out);
+  unsigned input_size = vector_data.size();
+
+  // pass the indexing variable to the data buffer
+  data_buffer << index;
+
+  // pass the rest of the input data to the buffer
+  for (unsigned i = 0; i < input_size; i++) {
+    data_buffer << ',' << vector_data[i];
+  }
+
+  data_buffer << std::endl;
+
+  // write the data buffer to the destination
+  dst.push_back( data_buffer.str() );
+}
+
+void CrushTester::write_integer_indexed_vector_data_string(vector<string> &dst, int index, vector<float> vector_data)
+{
+  stringstream data_buffer (stringstream::in | stringstream::out);
+  unsigned input_size = vector_data.size();
+
+  // pass the indexing variable to the data buffer
+  data_buffer << index;
+
+  // pass the rest of the input data to the buffer
+  for (unsigned i = 0; i < input_size; i++) {
+    data_buffer << ',' << vector_data[i];
+  }
+
+  data_buffer << std::endl;
+
+  // write the data buffer to the destination
+  dst.push_back( data_buffer.str() );
+}
+
+void CrushTester::write_integer_indexed_scalar_data_string(vector<string> &dst, int index, int scalar_data)
+{
+  stringstream data_buffer (stringstream::in | stringstream::out);
+
+  // pass the indexing variable to the data buffer
+  data_buffer << index;
+
+  // pass the input data to the buffer
+  data_buffer << ',' << scalar_data;
+  data_buffer << std::endl;
+
+  // write the data buffer to the destination
+  dst.push_back( data_buffer.str() );
+}
+void CrushTester::write_integer_indexed_scalar_data_string(vector<string> &dst, int index, float scalar_data)
+{
+  stringstream data_buffer (stringstream::in | stringstream::out);
+
+  // pass the indexing variable to the data buffer
+  data_buffer << index;
+
+  // pass the input data to the buffer
+  data_buffer << ',' << scalar_data;
+  data_buffer << std::endl;
+
+  // write the data buffer to the destination
+  dst.push_back( data_buffer.str() );
+}
+
+int CrushTester::test_with_fork(int timeout)
+{
+  ostringstream sink;
+  int r = fork_function(timeout, sink, [&]() {
+      return test();
+    });
+  if (r == -ETIMEDOUT) {
+    err << "timed out during smoke test (" << timeout << " seconds)";
+  }
+  return r;
+}
+
+namespace {
+  class BadCrushMap : public std::runtime_error {
+  public:
+    int item;
+    BadCrushMap(const char* msg, int id)
+      : std::runtime_error(msg), item(id) {}
+  };
+  // throws if any node in the crush fail to print
+  class CrushWalker : public CrushTreeDumper::Dumper<void> {
+    typedef void DumbFormatter;
+    typedef CrushTreeDumper::Dumper<DumbFormatter> Parent;
+    int max_id;
+  public:
+    CrushWalker(const CrushWrapper *crush, unsigned max_id)
+      : Parent(crush, CrushTreeDumper::name_map_t()), max_id(max_id) {}
+    void dump_item(const CrushTreeDumper::Item &qi, DumbFormatter *) override {
+      int type = -1;
+      if (qi.is_bucket()) {
+       if (!crush->get_item_name(qi.id)) {
+         throw BadCrushMap("unknown item name", qi.id);
+       }
+       type = crush->get_bucket_type(qi.id);
+      } else {
+       if (max_id > 0 && qi.id >= max_id) {
+         throw BadCrushMap("item id too large", qi.id);
+       }
+       type = 0;
+      }
+      if (!crush->get_type_name(type)) {
+       throw BadCrushMap("unknown type name", qi.id);
+      }
+    }
+  };
+}
+
+bool CrushTester::check_name_maps(unsigned max_id) const
+{
+  CrushWalker crush_walker(&crush, max_id);
+  try {
+    // walk through the crush, to see if its self-contained
+    crush_walker.dump(NULL);
+    // and see if the maps is also able to handle straying OSDs, whose id >= 0.
+    // "ceph osd tree" will try to print them, even they are not listed in the
+    // crush map.
+    crush_walker.dump_item(CrushTreeDumper::Item(0, 0, 0, 0), NULL);
+  } catch (const BadCrushMap& e) {
+    err << e.what() << ": item#" << e.item << std::endl;
+    return false;
+  }
+  return true;
+}
+
+static string get_rule_name(CrushWrapper& crush, int rule)
+{
+  if (crush.get_rule_name(rule))
+    return crush.get_rule_name(rule);
+  else
+    return string("rule") + std::to_string(rule);
+}
+
+void CrushTester::check_overlapped_rules() const
+{
+  namespace icl = boost::icl;
+  typedef std::set<string> RuleNames;
+  typedef icl::interval_map<int, RuleNames> Rules;
+  // <ruleset, type> => interval_map<size, {names}>
+  typedef std::map<std::pair<int, int>, Rules> RuleSets;
+  using interval = icl::interval<int>;
+
+  // mimic the logic of crush_find_rule(), but it only return the first matched
+  // one, but I am collecting all of them by the overlapped sizes.
+  RuleSets rulesets;
+  for (int rule = 0; rule < crush.get_max_rules(); rule++) {
+    if (!crush.rule_exists(rule)) {
+      continue;
+    }
+    Rules& rules = rulesets[{crush.get_rule_mask_ruleset(rule),
+                            crush.get_rule_mask_type(rule)}];
+    rules += make_pair(interval::closed(crush.get_rule_mask_min_size(rule),
+                                       crush.get_rule_mask_max_size(rule)),
+                      RuleNames{get_rule_name(crush, rule)});
+  }
+  for (auto i : rulesets) {
+    auto ruleset_type = i.first;
+    const Rules& rules = i.second;
+    for (auto r : rules) {
+      const RuleNames& names = r.second;
+      // if there are more than one rules covering the same size range,
+      // print them out.
+      if (names.size() > 1) {
+       err << "overlapped rules in ruleset " << ruleset_type.first << ": "
+           << boost::join(names, ", ") << "\n";
+      }
+    }
+  }
+}
+
+int CrushTester::test()
+{
+  if (min_rule < 0 || max_rule < 0) {
+    min_rule = 0;
+    max_rule = crush.get_max_rules() - 1;
+  }
+  if (min_x < 0 || max_x < 0) {
+    min_x = 0;
+    max_x = 1023;
+  }
+
+  // initial osd weights
+  vector<__u32> weight;
+
+  /*
+   * note device weight is set by crushtool
+   * (likely due to a given a command line option)
+   */
+  for (int o = 0; o < crush.get_max_devices(); o++) {
+    if (device_weight.count(o)) {
+      weight.push_back(device_weight[o]);
+    } else if (crush.check_item_present(o)) {
+      weight.push_back(0x10000);
+    } else {
+      weight.push_back(0);
+    }
+  }
+
+  if (output_utilization_all)
+    err << "devices weights (hex): " << hex << weight << dec << std::endl;
+
+  // make adjustments
+  adjust_weights(weight);
+
+
+  int num_devices_active = 0;
+  for (vector<__u32>::iterator p = weight.begin(); p != weight.end(); ++p)
+    if (*p > 0)
+      num_devices_active++;
+
+  if (output_choose_tries)
+    crush.start_choose_profile();
+  
+  for (int r = min_rule; r < crush.get_max_rules() && r <= max_rule; r++) {
+    if (!crush.rule_exists(r)) {
+      if (output_statistics)
+        err << "rule " << r << " dne" << std::endl;
+      continue;
+    }
+    if (ruleset >= 0 &&
+       crush.get_rule_mask_ruleset(r) != ruleset) {
+      continue;
+    }
+    int minr = min_rep, maxr = max_rep;
+    if (min_rep < 0 || max_rep < 0) {
+      minr = crush.get_rule_mask_min_size(r);
+      maxr = crush.get_rule_mask_max_size(r);
+    }
+    
+    if (output_statistics)
+      err << "rule " << r << " (" << crush.get_rule_name(r)
+      << "), x = " << min_x << ".." << max_x
+      << ", numrep = " << minr << ".." << maxr
+      << std::endl;
+
+    for (int nr = minr; nr <= maxr; nr++) {
+      vector<int> per(crush.get_max_devices());
+      map<int,int> sizes;
+
+      int num_objects = ((max_x - min_x) + 1);
+      float num_devices = (float) per.size(); // get the total number of devices, better to cast as a float here 
+
+      // create a structure to hold data for post-processing
+      tester_data_set tester_data;
+      vector<float> vector_data_buffer_f;
+
+      // create a map to hold batch-level placement information
+      map<int, vector<int> > batch_per;
+      int objects_per_batch = num_objects / num_batches;
+      int batch_min = min_x;
+      int batch_max = min_x + objects_per_batch - 1;
+
+      // get the total weight of the system
+      int total_weight = 0;
+      for (unsigned i = 0; i < per.size(); i++)
+        total_weight += weight[i];
+
+      if (total_weight == 0)
+       continue;
+
+      // compute the expected number of objects stored per device in the absence of weighting
+      float expected_objects = min(nr, get_maximum_affected_by_rule(r)) * num_objects;
+
+      // compute each device's proportional weight
+      vector<float> proportional_weights( per.size() );
+
+      for (unsigned i = 0; i < per.size(); i++)
+        proportional_weights[i] = (float) weight[i] / (float) total_weight;
+
+      if (output_data_file) {
+        // stage the absolute weight information for post-processing
+        for (unsigned i = 0; i < per.size(); i++) {
+          tester_data.absolute_weights[i] = (float) weight[i] / (float)0x10000;
+        }
+
+        // stage the proportional weight information for post-processing
+        for (unsigned i = 0; i < per.size(); i++) {
+          if (proportional_weights[i] > 0 )
+            tester_data.proportional_weights[i] = proportional_weights[i];
+
+          tester_data.proportional_weights_all[i] = proportional_weights[i];
+        }
+
+      }
+      // compute the expected number of objects stored per device when a device's weight is considered
+      vector<float> num_objects_expected(num_devices);
+
+      for (unsigned i = 0; i < num_devices; i++)
+        num_objects_expected[i] = (proportional_weights[i]*expected_objects);
+
+      for (int current_batch = 0; current_batch < num_batches; current_batch++) {
+        if (current_batch == (num_batches - 1)) {
+          batch_max = max_x;
+          objects_per_batch = (batch_max - batch_min + 1);
+        }
+
+        float batch_expected_objects = min(nr, get_maximum_affected_by_rule(r)) * objects_per_batch;
+        vector<float> batch_num_objects_expected( per.size() );
+
+        for (unsigned i = 0; i < per.size() ; i++)
+          batch_num_objects_expected[i] = (proportional_weights[i]*batch_expected_objects);
+
+        // create a vector to hold placement results temporarily 
+        vector<int> temporary_per ( per.size() );
+
+        for (int x = batch_min; x <= batch_max; x++) {
+          // create a vector to hold the results of a CRUSH placement or RNG simulation
+          vector<int> out;
+
+          if (use_crush) {
+            if (output_mappings)
+             err << "CRUSH"; // prepend CRUSH to placement output
+            uint32_t real_x = x;
+            if (pool_id != -1) {
+              real_x = crush_hash32_2(CRUSH_HASH_RJENKINS1, x, (uint32_t)pool_id);
+            }
+            crush.do_rule(r, real_x, out, nr, weight, 0);
+          } else {
+            if (output_mappings)
+             err << "RNG"; // prepend RNG to placement output to denote simulation
+            // test our new monte carlo placement generator
+            random_placement(r, out, nr, weight);
+          }
+
+         if (output_mappings)
+           err << " rule " << r << " x " << x << " " << out << std::endl;
+
+          if (output_data_file)
+            write_integer_indexed_vector_data_string(tester_data.placement_information, x, out);
+
+          bool has_item_none = false;
+          for (unsigned i = 0; i < out.size(); i++) {
+            if (out[i] != CRUSH_ITEM_NONE) {
+              per[out[i]]++;
+              temporary_per[out[i]]++;
+            } else {
+              has_item_none = true;
+            }
+          }
+
+          batch_per[current_batch] = temporary_per;
+          sizes[out.size()]++;
+          if (output_bad_mappings && 
+              (out.size() != (unsigned)nr ||
+               has_item_none)) {
+            err << "bad mapping rule " << r << " x " << x << " num_rep " << nr << " result " << out << std::endl;
+          }
+        }
+
+        batch_min = batch_max + 1;
+        batch_max = batch_min + objects_per_batch - 1;
+      }
+
+      for (unsigned i = 0; i < per.size(); i++)
+        if (output_utilization && !output_statistics)
+          err << "  device " << i
+          << ":\t" << per[i] << std::endl;
+
+      for (map<int,int>::iterator p = sizes.begin(); p != sizes.end(); ++p)
+        if (output_statistics)
+          err << "rule " << r << " (" << crush.get_rule_name(r) << ") num_rep " << nr
+          << " result size == " << p->first << ":\t"
+          << p->second << "/" << (max_x-min_x+1) << std::endl;
+
+      if (output_statistics)
+        for (unsigned i = 0; i < per.size(); i++) {
+          if (output_utilization) {
+            if (num_objects_expected[i] > 0 && per[i] > 0) {
+              err << "  device " << i << ":\t"
+                  << "\t" << " stored " << ": " << per[i]
+                  << "\t" << " expected " << ": " << num_objects_expected[i]
+                  << std::endl;
+            }
+          } else if (output_utilization_all) {
+            err << "  device " << i << ":\t"
+                << "\t" << " stored " << ": " << per[i]
+                << "\t" << " expected " << ": " << num_objects_expected[i]
+                << std::endl;
+          }
+        }
+
+      if (output_data_file)
+        for (unsigned i = 0; i < per.size(); i++) {
+          vector_data_buffer_f.clear();
+          vector_data_buffer_f.push_back( (float) per[i]);
+          vector_data_buffer_f.push_back( (float) num_objects_expected[i]);
+
+          write_integer_indexed_vector_data_string(tester_data.device_utilization_all, i, vector_data_buffer_f);
+
+          if (num_objects_expected[i] > 0 && per[i] > 0)
+            write_integer_indexed_vector_data_string(tester_data.device_utilization, i, vector_data_buffer_f);
+        }
+
+      if (output_data_file && num_batches > 1) {
+        // stage batch utilization information for post-processing
+        for (int i = 0; i < num_batches; i++) {
+          write_integer_indexed_vector_data_string(tester_data.batch_device_utilization_all, i, batch_per[i]);
+          write_integer_indexed_vector_data_string(tester_data.batch_device_expected_utilization_all, i, batch_per[i]);
+        }
+      }
+
+      string rule_tag = crush.get_rule_name(r);
+
+      if (output_csv)
+        write_data_set_to_csv(output_data_file_name+rule_tag,tester_data);
+    }
+  }
+
+  if (output_choose_tries) {
+    __u32 *v = 0;
+    int n = crush.get_choose_profile(&v);
+    for (int i=0; i<n; i++) {
+      cout.setf(std::ios::right);
+      cout << std::setw(2)
+      << i << ": " << std::setw(9) << v[i];
+      cout.unsetf(std::ios::right);
+      cout << std::endl;
+    }
+
+    crush.stop_choose_profile();
+  }
+
+  return 0;
+}