diff --git a/extern/benchmark_runner/benchmark_runner.h b/extern/benchmark_runner/benchmark_runner.h
index 37a2ac6..2a399b3 100644
--- a/extern/benchmark_runner/benchmark_runner.h
+++ b/extern/benchmark_runner/benchmark_runner.h
@@ -10,6 +10,14 @@
 #include <iostream>
 #include <fstream>
 #include <bits/stdc++.h>
+#include <thread>
+#include <map>
+
+#include <tuple>
+#include <unistd.h>
+#include <string>
+#include <fstream>
+#include <sstream>
 
 using namespace std;
 
@@ -21,9 +29,16 @@ class benchmark_runner {
   chrono::steady_clock::time_point last_start_time_;
   vector<long> times_;
 
+  map<string, vector<long>> custom_stats_;
+
   void print_statistics() {
     long time_sum = std::accumulate(times_.begin(), times_.end(), 0l);
     cout << "Average Runtime (us): " << (time_sum / times_.size()) << endl;
+
+    for (auto &iter : custom_stats_) {
+      long custom_stat_sum = std::accumulate(iter.second.begin(), iter.second.end(), 0l);
+      cout << "Average " << iter.first << ": " << (custom_stat_sum / times_.size()) << endl;
+    }
   }
 
   inline bool file_exists(const std::string &name) {
@@ -32,17 +47,53 @@ class benchmark_runner {
   }
 
  public:
-  benchmark_runner(string csv_path, string csv_name, int num_measurements = 10000) : csv_path_{std::move(csv_path)},
-                                                                                     csv_name_{std::move(csv_name)},
-                                                                                     times_{} {
+  benchmark_runner(string csv_path, string csv_name) : csv_path_{std::move(csv_path)},
+                                                       csv_name_{std::move(csv_name)},
+                                                       times_{} {
     string command = "mkdir -p " + csv_path_;
     int res = system(command.c_str());
     if (res) {
       cout << "Error while creating directory!" << endl;
       exit(1);
     }
+  }
+
+  /**
+   * Queries the pages used by the current process.
+   * Splits between private and shared pages.
+   *
+   * @return Tuple(private_pages, shared_pages) used by the process.
+   */
+  static std::pair<unsigned long, unsigned long> query_process_memory_pages() {
+    pid_t my_pid = getpid();
+    std::ostringstream proc_stats_path_builder;
+    proc_stats_path_builder << "/proc/" << my_pid << "/smaps_rollup";
+    std::string proc_stats_path = proc_stats_path_builder.str();
+
+    std::ifstream proc_stats_file{proc_stats_path};
+
+    std::string line;
+    unsigned long total_shared = 0;
+    unsigned long total_private = 0;
+    while (std::getline(proc_stats_file, line)) {
+      std::stringstream line_input(line);
+      std::string type;
+      unsigned long size;
+      line_input >> type >> size;
+
+      if (type.find("Shared") != std::string::npos) {
+        total_shared += size;
+      }
+      if (type.find("Private") != std::string::npos) {
+        total_private += size;
+      }
+    }
+
+    return std::make_pair(total_private, total_shared);
+  }
 
-    times_.reserve(num_measurements);
+  void add_custom_stats_field(const string name) {
+    custom_stats_.insert({name, {}});
   }
 
   static void read_args(int argc, char **argv, int &num_threads, string &path) {
@@ -65,22 +116,35 @@ class benchmark_runner {
     auto end_time = chrono::steady_clock::now();
     long time = chrono::duration_cast<chrono::microseconds>(end_time - last_start_time_).count();
     times_.emplace_back(time);
+
+    for (auto &iter : custom_stats_) {
+      iter.second.emplace_back(0);
+    }
+  }
+
+  void store_custom_stat(const string &name, long value) {
+    auto &stat_vector = custom_stats_[name];
+    stat_vector[stat_vector.size() - 1] = value;
   }
 
   void run_iterations(int count,
                       const function<void(void)> measure,
                       int warmup_count,
-                      const function<void(void)> prepare = []() {}) {
+                      const function<void(void)> prepare = []() {},
+                      const function<void(void)> finish = []() {}) {
     for (int i = 0; i < warmup_count; i++) {
       prepare();
       measure();
     }
 
     for (int i = 0; i < count; i++) {
+      using namespace std::literals;
+      this_thread::sleep_for(100us);
       prepare();
       start_iteration();
       measure();
       end_iteration();
+      finish();
     }
   }
 
@@ -94,15 +158,29 @@ class benchmark_runner {
 
     { // Scope for output file
       ofstream o(full_filename, std::fstream::out | std::fstream::app);
+
       if (write_header) {
-        o << "runtime_us" << endl;
+        o << "runtime_us";
+        for (auto &iter : custom_stats_) {
+          o << ";" << iter.first;
+        }
+        o << endl;
       }
-      for (auto time : times_) {
-        o << time << endl;
+
+      // TODO: make this more efficient
+      for (size_t i = 0; i < times_.size(); i++) {
+        o << times_[i];
+        for (auto &iter : custom_stats_) {
+          o << ";" << iter.second[i];
+        }
+        o << endl;
       }
     } // End Scope for output file
 
     times_.clear();
+    for (auto &iter : custom_stats_) {
+      iter.second.clear();
+    }
   }
 };