diff --git a/similarity_search/CMakeLists.txt b/similarity_search/CMakeLists.txt
index fae4e53..1451b30 100644
--- a/similarity_search/CMakeLists.txt
+++ b/similarity_search/CMakeLists.txt
@@ -42,28 +42,28 @@ if(${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU")
         message(FATAL_ERROR "GCC version must be at least 4.7!")
     endif()
     # Uncomment the following lines to see how the code compiles without AVX,SSE4.2 and/or SSE2
-    #set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Ofast -lm -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -fopenmp -fpic -march=x86-64")
-    #set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Ofast -lm -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -fopenmp -fpic -march=core2")
-    #set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Ofast -lm -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -fopenmp -fpic -msse4.2")
-    set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Wcast-align -Ofast -lm -lrt -DNDEBUG -std=c++11 -fopenmp  -DHAVE_CXX0X -march=native -Wl,--no-as-needed -fpic")
-    set (CMAKE_CXX_FLAGS_DEBUG   "-Wall -Wunreachable-code -Wcast-align -ggdb  -lm -lrt -DNDEBUG -std=c++11 -fopenmp  -DHAVE_CXX0X -march=native -Wl,--no-as-needed -fpic")
+    #set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Ofast -lm -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -pthread -fpic -march=x86-64")
+    #set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Ofast -lm -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -pthread -fpic -march=core2")
+    #set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Ofast -lm -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -pthread -fpic -msse4.2")
+    set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Wcast-align -Ofast -lm -lrt -DNDEBUG -std=c++11 -pthread  -DHAVE_CXX0X -march=native -Wl,--no-as-needed -fpic")
+    set (CMAKE_CXX_FLAGS_DEBUG   "-Wall -Wunreachable-code -Wcast-align -ggdb  -lm -lrt -DNDEBUG -std=c++11 -pthread  -DHAVE_CXX0X -march=native -Wl,--no-as-needed -fpic")
 elseif(${CMAKE_CXX_COMPILER_ID} STREQUAL "Intel")
     if (CXX_COMPILER_VERSION VERSION_LESS 14.0.1)
         message(FATAL_ERROR "Intel version must be at least 14.0.1!")
     endif()
-    set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Ofast -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -openmp   -march=native -fpic")
-    set (CMAKE_CXX_FLAGS_DEBUG   "-Wall -Wunreachable-code -ggdb  -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -openmp   -march=native -fpic")
+    set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Ofast -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -pthread -march=native -fpic")
+    set (CMAKE_CXX_FLAGS_DEBUG   "-Wall -Wunreachable-code -ggdb  -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -pthread -march=native -fpic")
 elseif(${CMAKE_CXX_COMPILER_ID} STREQUAL "Clang")
     if (CXX_COMPILER_VERSION VERSION_LESS 4.2.1)
         message(FATAL_ERROR "Clang version must be at least 3.4 (GCC >= 4.2.1 equivalent)!")
     endif()
     if (CMAKE_SYSTEM_NAME MATCHES Darwin)
         # MACOSX
-        set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Wcast-align -O3 -DNDEBUG -std=c++11 -DHAVE_CXX0X -fpic -march=native")
-        set (CMAKE_CXX_FLAGS_DEBUG   "-Wall -Wunreachable-code -Wcast-align -ggdb  -DNDEBUG -std=c++11 -DHAVE_CXX0X -fpic -march=native")
+        set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Wcast-align -O3 -DNDEBUG -std=c++11 -DHAVE_CXX0X -pthread -fpic -march=native")
+        set (CMAKE_CXX_FLAGS_DEBUG   "-Wall -Wunreachable-code -Wcast-align -ggdb  -DNDEBUG -std=c++11 -DHAVE_CXX0X -pthread -fpic -march=native")
     else()
-        set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Wcast-align -O3 -DNDEBUG -std=c++11 -DHAVE_CXX0X -fopenmp -march=native -fpic")
-        set (CMAKE_CXX_FLAGS_DEBUG   "-Wall -Wunreachable-code -Wcast-align -ggdb  -DNDEBUG -std=c++11 -DHAVE_CXX0X -fopenmp -march=native -fpic")
+        set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wunreachable-code -Wcast-align -O3 -DNDEBUG -std=c++11 -DHAVE_CXX0X -pthread -march=native -fpic")
+        set (CMAKE_CXX_FLAGS_DEBUG   "-Wall -Wunreachable-code -Wcast-align -ggdb  -DNDEBUG -std=c++11 -DHAVE_CXX0X -pthread -march=native -fpic")
     endif()
     #message(FATAL_ERROR "CLANG ${CMAKE_SYSTEM_NAME}")
 elseif(WIN32)
diff --git a/similarity_search/include/thread_pool.h b/similarity_search/include/thread_pool.h
index 21c0b34..0a98141 100644
--- a/similarity_search/include/thread_pool.h
+++ b/similarity_search/include/thread_pool.h
@@ -54,33 +54,47 @@ namespace similarity {
 
 */
 
-  // replacement for the openmp '#pragma omp parallel for' directive
-  // only handles a subset of functionality (no reductions etc)
+  /* 
+   * replacement for the openmp '#pragma omp parallel for' directive
+   * only handles a subset of functionality (no reductions etc)
+   * Process ids from start (inclusive) to end (EXCLUSIVE)
+   */
   template <class Function>
-  inline void ParallelFor(int initial, int final, int numThreads, Function fn) {
+  inline void ParallelFor(size_t start, size_t end, size_t numThreads, Function fn) {
     if (numThreads <= 0) {
       numThreads = std::thread::hardware_concurrency();
     }
 
-    std::vector<std::thread> threads;
-    std::atomic<int> current(initial);
+    std::vector<std::thread>  threads;
+    std::atomic<size_t>       current(start);
 
     // keep track of exceptions in threads
     // https://stackoverflow.com/a/32428427/1713196
     std::exception_ptr lastException = nullptr;
+    std::mutex         lastExceptMutex;
 
-    for (int i = 0; i < numThreads; ++i) {
+    for (size_t i = 0; i < numThreads; ++i) {
       threads.push_back(std::thread([&] {
         while (true) {
-          int id = current.fetch_add(1);
-          if ((id >= final) || lastException) {
+          size_t id = current.fetch_add(1);
+
+          if ((id >= end)) {
             break;
           }
 
           try {
             fn(id);
           } catch (...) {
+            std::unique_lock<std::mutex> lastExcepLock(lastExceptMutex);
             lastException = std::current_exception();
+            /* 
+             * This will work even when current is the largest value that
+             * size_t can fit, because fetch_add returns the previous value
+             * before the increment (what will result in overflow 
+             * and produce 0 instead of current + 1).
+             */
+            current = end;
+            break;
           }
         }
       }));
diff --git a/similarity_search/src/method/hnsw.cc b/similarity_search/src/method/hnsw.cc
index 9bd9229..bb8009a 100644
--- a/similarity_search/src/method/hnsw.cc
+++ b/similarity_search/src/method/hnsw.cc
@@ -1012,7 +1012,7 @@ namespace similarity {
                 ++currElem;
         }
 
-        for (int_fast32_t i = 0; i < query->GetK() && i < sortedArr.size(); ++i) {
+        for (uint_fast32_t i = 0; i < query->GetK() && i < sortedArr.size(); ++i) {
             query->CheckAndAddToResult(queueData[i].key, queueData[i].data->getData());
         }