Some first performance improvements (#26)

Some very simple (but helpful) performance improvements: speed up signal calculation by roughly 3x

Author: philippwindischhofer

src/CylindricalWeightingFieldCalculator.cxx

@@ -384,20 +384,14 @@ void CylindricalWeightingFieldCalculator::Calculate(std::filesystem::path outdir
     }
 
     cld.ind_t = stepcnt++;
-    std::cout << "entering saving chunkloop" << std::endl;
     m_f -> loop_in_chunks(meep::eisvogel_saving_chunkloop, static_cast<void*>(&cld), m_f -> total_volume());
-    std::cout << "exit saving chunkloop" << std::endl;
 
     if((stepcnt % 400) == 0) {
-      std::cout << "start chunk merging" << std::endl;
       fstor -> MergeChunks(0, 400);
-      std::cout << "end chunk merging" << std::endl;
     }
   }
 
-  std::cout << "start chunk merging" << std::endl;
   fstor -> MergeChunks(0, 400);
-  std::cout << "end chunk merging" << std::endl;
 
   // TODO: again, will get better once the three separate arrays are gone
   // TODO: for large weighting fields, will have to move chunks to the permanent location continuously throughout the calculation so as not to fill up local storage
@@ -418,13 +412,7 @@ void CylindricalWeightingFieldCalculator::Calculate(std::filesystem::path outdir
 
   if(meep::am_master()) {
     std::shared_ptr<CylindricalWeightingField> cwf = std::make_shared<CylindricalWeightingField>(outdir, *m_start_coords, *m_end_coords);
-    cwf -> MakeMetadataPersistent();
-
-    // Sometimes need to wait for all files to show up?
-    // std::this_thread::sleep_for(std::chrono::milliseconds(1000));
-    
-    std::cout << "start defragmentation" << std::endl;
+    cwf -> MakeMetadataPersistent();    
     cwf -> RebuildChunks(requested_chunk_size);
-    std::cout << "end defragmentation" << std::endl;
   }
 }

src/DefaultSerializationTraits.hxx

@@ -111,10 +111,10 @@ namespace stor {
     }
   };
 
-  template<std::size_t n>
-  struct Traits<std::array<float, n>> {
+  // template<std::size_t n>
+  // struct Traits<std::array<float, n>> {
 
-  };
+  // };
 
   // For general vectors
   template <typename T>
@@ -192,7 +192,7 @@ namespace stor {
 
       type retval;
       for(std::size_t ind = 0; ind < keys.size(); ind++) {
-	retval[keys[ind]] = values[ind];
+	retval.insert(retval.end(), std::pair{keys[ind], values[ind]});
       }
 
       return retval;

src/DenseNDArray.hh

@@ -57,6 +57,21 @@ public:
 
     return retval;
   }
+
+  static DenseNDArray<T, dims> FromSparseFile(std::iostream& stream) {
+    T default_value = stor::Traits<T>::deserialize(stream);
+    shape_t shape = stor::Traits<shape_t>::deserialize(stream);
+    DenseNDArray<T, dims> retval(shape, default_value);
+
+    std::vector<std::array<std::size_t, dims>> keys = stor::Traits<std::vector<std::array<std::size_t, dims>>>::deserialize(stream);
+    std::vector<T> values = stor::Traits<std::vector<T>>::deserialize(stream);
+
+    for(std::size_t el_ind = 0; el_ind < keys.size(); el_ind++) {
+      retval(keys[el_ind]) = values[el_ind];
+    }
+    
+    return retval;
+  }
 
   DenseNDArray(const shape_t& shape, std::vector<T>&& data) : NDArray<T, dims>(shape) {
     m_strides[0] = 1;

src/DistributedNDArray.hh

@@ -108,6 +108,8 @@ private:
   // can think about turning this into a class
   index_t m_chunk_index;
 
+  std::size_t m_chunk_last_accessed;
+  
   // The index may not start at {0, 0, 0}
   IndexVector m_global_start_ind;
 

src/DistributedNDArray.hxx

@@ -30,7 +30,7 @@ namespace stor {
 template <class T, std::size_t dims, template<class, std::size_t> class DenseT, template<class, std::size_t> class SparseT, class SerializerT>
 DistributedNDArray<T, dims, DenseT, SparseT, SerializerT>::DistributedNDArray(std::string dirpath, std::size_t max_cache_size, SerializerT& ser) :
   NDArray<T, dims>(), m_dirpath(dirpath), m_indexpath(dirpath + "/index.bin"), m_max_cache_size(max_cache_size),
-  m_global_start_ind(dims, 0), m_ser(ser) {
+  m_chunk_last_accessed(0), m_global_start_ind(dims, 0), m_ser(ser) {
 
   // Create directory if it does not already exist
   if(!std::filesystem::exists(m_dirpath)) {
@@ -104,7 +104,8 @@ void DistributedNDArray<T, dims, DenseT, SparseT, SerializerT>::WriteChunk(const
   std::size_t num_elems = chunk.volume();
 
   // pays off to store as sparse chunk
-  std::size_t sparse_vs_dense_expense_ratio = 3; // sparse storage is approximately 3x as expensive as dense storage per nonzero element
+  // std::size_t sparse_vs_dense_expense_ratio = 3; // when only counting storage space: sparse storage is approximately 3x as expensive as dense storage per nonzero element
+  std::size_t sparse_vs_dense_expense_ratio = 20; // when also counting complexity of deserializing + rebuilding a dense chunk
   if(sparse_vs_dense_expense_ratio * num_nonzero_elems < num_elems) {
 
     std::cout << "going to sparsify" << std::endl;
@@ -232,15 +233,15 @@ void DistributedNDArray<T, dims, DenseT, SparseT, SerializerT>::RebuildChunks(co
     dense_t current_chunk = retrieveChunk(chunk_index);
 
     if(actual_chunk_shape == current_chunk.shape()) {
-      std::cout << "chunk already has the correct size, keep it" << std::endl;
+      // std::cout << "chunk already has the correct size, keep it" << std::endl;
       chunks_to_keep.push_back(m_chunk_index[chunk_index]);
       continue;
     }
 
-    std::cout << "now working on rebuild chunk with inds" << std::endl;
-    std::cout << "chunk_inds_start = " << std::endl;
+    // std::cout << "now working on rebuild chunk with inds" << std::endl;
+    // std::cout << "chunk_inds_start = " << std::endl;
     chunk_inds_start.print();
-    std::cout << "chunk_inds_end = " << std::endl;
+    // std::cout << "chunk_inds_end = " << std::endl;
     chunk_inds_end.print();
 
     dense_t chunk = range(chunk_inds_start, chunk_inds_end);
@@ -436,11 +437,26 @@ bool DistributedNDArray<T, dims, DenseT, SparseT, SerializerT>::chunkContainsInd
 
 template <class T, std::size_t dims, template<class, std::size_t> class DenseT, template<class, std::size_t> class SparseT, class SerializerT>
 std::size_t DistributedNDArray<T, dims, DenseT, SparseT, SerializerT>::getChunkIndex(const IndexVector& inds) {
-  std::size_t chunk_ind = 0;
-  for(chunk_ind = 0; chunk_ind < m_chunk_index.size(); chunk_ind++) {
-    if(chunkContainsInds(m_chunk_index[chunk_ind], inds)) {
-      return chunk_ind;
-    }
+
+  if(m_chunk_index.size() == 0) {
+    [[unlikely]];
+    throw ChunkNotFoundError();
+  }
+  
+  if(chunkContainsInds(m_chunk_index[m_chunk_last_accessed], inds)) {
+    [[likely]];
+    return m_chunk_last_accessed;
+  }
+  else {
+    // Trigger a full chunk lookup
+    // TODO: have a search tree here with logarithmic instead of linear complexity
+    std::size_t chunk_ind = 0;
+    for(chunk_ind = 0; chunk_ind < m_chunk_index.size(); chunk_ind++) {
+      if(chunkContainsInds(m_chunk_index[chunk_ind], inds)) {
+	m_chunk_last_accessed = chunk_ind;
+	return chunk_ind;
+      }
+    }   
   }
 
   std::cout << "HHHHHHH" << std::endl;
@@ -476,7 +492,8 @@ DistributedNDArray<T, dims, DenseT, SparseT, SerializerT>::dense_t& DistributedN
       m_chunk_cache.insert({chunk_ind, m_ser.template deserialize<dense_t>(ifs)});
     }
     else if(meta.chunk_type == ChunkType::sparse) {
-      m_chunk_cache.insert({chunk_ind, dense_t::From(m_ser.template deserialize<sparse_t>(ifs))});
+      m_chunk_cache.insert({chunk_ind, dense_t::FromSparseFile(ifs)});
+      // m_chunk_cache.insert({chunk_ind, dense_t::From(m_ser.template deserialize<sparse_t>(ifs))});
     }
     else {
       throw std::runtime_error("Error: unknown chunk type encountered!");

src/NDArrayOperations.hh

@@ -10,7 +10,7 @@ namespace NDArrayOps {
   template <class T, std::size_t dims>
   DenseNDArray<T, dims> concatenate(const DenseNDArray<T, dims>& arr_1, const DenseNDArray<T, dims>& arr_2, std::size_t axis) {
 
-    std::cout << " ---> START CONCATENATE <---" << std::endl;
+    // std::cout << " ---> START CONCATENATE <---" << std::endl;
 
     if(axis >= dims) {
       throw std::runtime_error("Error: 'axis' out of bounds");
@@ -39,7 +39,7 @@ namespace NDArrayOps {
 
     DenseNDArray<T, dims> retval(final_shape_crutch, 0.0);
 
-    std::cout << " ---> MIGRATE ARR_1 <---" << std::endl;
+    // std::cout << " ---> MIGRATE ARR_1 <---" << std::endl;
 
     // Migrate contents of arr_1
     {
@@ -51,7 +51,7 @@ namespace NDArrayOps {
       }
     }
 
-    std::cout << " ---> MIGRATE ARR_2 <---" << std::endl;
+    // std::cout << " ---> MIGRATE ARR_2 <---" << std::endl;
 
     // Migrate contents of arr_2
     {
@@ -68,7 +68,7 @@ namespace NDArrayOps {
       }
     }
 
-    std::cout << " ---> FINISH CONCATENATE <---" << std::endl;
+    // std::cout << " ---> FINISH CONCATENATE <---" << std::endl;
 
     return retval;
   }

src/SparseNDArray.hh

@@ -3,6 +3,7 @@
 
 #include <array>
 #include <map>
+#include <chrono>
 #include "NDArray.hh"
 #include "DenseNDArray.hh"
 #include "Eisvogel/IteratorUtils.hh"
@@ -108,6 +109,74 @@ namespace stor {
   };
 }
 
+// namespace stor {
+
+//   template<typename T, std::size_t dims>
+//   struct Traits<SparseNDArray<T, dims>> {
+//     using type = SparseNDArray<T, dims>;
+//     using shape_t = typename type::shape_t;
+//     using data_t = typename type::data_t;
+
+//     static void serialize(std::iostream& stream, const type& val) {      
+//       Traits<T>::serialize(stream, val.m_default_value);
+//       Traits<shape_t>::serialize(stream, val.m_shape);
+
+//       // Convert array data from std::map<index, value> into two 1-dimensional vectors:
+//       // (index_1, index_2, ...) with total length of `index_len`, and
+//       // (value_1, value_2, ...) with total length of `number_entries`
+//       std::size_t number_entries = val.m_data.size();
+//       std::size_t index_len = dims * number_entries;
+      
+//       std::vector<std::size_t> index_vec(index_len);
+//       std::vector<T> data_vec(number_entries);
+
+//       // Fill the two vectors ...
+//       auto it_index_vec = index_vec.begin();
+//       auto it_data_vec = data_vec.begin();
+//       for (auto const& [key, val] : val.m_data) {
+// 	std::copy(key.cbegin(), key.cend(), it_index_vec);
+// 	*it_data_vec = val;
+
+// 	std::advance(it_data_vec, 1);
+// 	std::advance(it_index_vec, dims);
+//       }
+      
+//       // ... and serialize them
+//       Traits<std::vector<std::size_t>>::serialize(stream, index_vec);
+//       Traits<std::vector<T>>::serialize(stream, data_vec);
+//     }
+
+//     static type deserialize(std::iostream& stream) {
+
+//       std::chrono::high_resolution_clock::time_point t_start = std::chrono::high_resolution_clock::now();
+      
+//       T default_value = Traits<T>::deserialize(stream);
+//       shape_t shape = Traits<shape_t>::deserialize(stream);
+
+//       std::vector<std::size_t> index_vec = Traits<std::vector<std::size_t>>::deserialize(stream);
+//       std::vector<T> data_vec = Traits<std::vector<T>>::deserialize(stream);
+      
+//       // Fill `data` map from `index_vec` and `data_vec` ...
+//       data_t data;
+//       auto it_index_vec = index_vec.begin();
+//       auto it_data_vec = data_vec.begin();
+//       std::array<std::size_t, dims> cur_ind;
+//       while(it_data_vec != data_vec.end()) {
+// 	std::copy_n(it_index_vec, dims, cur_ind.begin());
+// 	data.insert(data.end(), std::pair{cur_ind, *it_data_vec});
+	
+// 	std::advance(it_data_vec, 1);
+// 	std::advance(it_index_vec, dims);
+//       }
+
+//       // ... and build the sparse array      
+//       SparseNDArray<T, dims> retval(shape, default_value);
+//       retval.m_data = data;
+//       return retval;
+//     }
+//   };
+// }
+
 // Some type shortcuts
 template <class T>
 using SparseScalarField3D = SparseNDArray<T, 3>;

tests/io/testSerialization.cxx

@@ -1,5 +1,7 @@
 #include "Serialization.hh"
 #include "Eisvogel/Common.hh"
+#include "DenseNDArray.hh"
+#include "SparseNDArray.hh"
 
 #include <fstream>
 #include <iostream>
@@ -37,7 +39,7 @@ int test_serialization_vector(std::string ser_path, std::size_t length) {
   std::chrono::high_resolution_clock::time_point t_end = std::chrono::high_resolution_clock::now();
 
   std::chrono::duration<double> time_span = duration_cast<std::chrono::duration<double>>(t_end - t_start);
-  std::cout << "Completed in " << time_span.count() << " seconds." << std::endl;
+  std::cout << "std::vector --> Completed in " << time_span.count() << " seconds." << std::endl;
 
   for(std::size_t ind = 0; ind < length; ind++) {
     if(vec[ind] != res[ind]) {
@@ -48,8 +50,65 @@ int test_serialization_vector(std::string ser_path, std::size_t length) {
   return 0;
 }
 
+template <std::size_t dims>
+int test_serialization_sparse_array(std::string ser_path, std::size_t size) {
+  
+  // Fill random sparse array
+  std::array<std::size_t, dims> shape;
+  for(std::size_t dim = 0; dim < dims; dim++) {
+    shape[dim] = size;
+  }
+  DenseNDArray<scalar_t, dims> darr(shape, 1.0);
+
+  auto to_keep = [](float value) -> bool {
+    return value != 0.0;
+  };
+  SparseNDArray<scalar_t, dims> sparr = SparseNDArray<scalar_t, dims>::From(darr, to_keep, 0.0);
+
+  std::chrono::high_resolution_clock::time_point t_start = std::chrono::high_resolution_clock::now();
+  
+  std::fstream ofs;
+  ofs.open(ser_path, std::ios::out | std::ios::binary);  
+  stor::DefaultSerializer oser;
+  oser.serialize(ofs, sparr);
+  ofs.close();
+  
+  std::chrono::high_resolution_clock::time_point t_end = std::chrono::high_resolution_clock::now();
+  std::chrono::duration<double> time_span = duration_cast<std::chrono::duration<double>>(t_end - t_start);
+  std::cout << "SparseNDArray --> Serialization completed in " << time_span.count() << " seconds." << std::endl;
+
+  t_start = std::chrono::high_resolution_clock::now();
+  
+  std::fstream ifs;
+  ifs.open(ser_path, std::ios::in | std::ios::binary);
+  stor::DefaultSerializer iser; 
+  SparseNDArray<scalar_t, dims> sparr_read = iser.deserialize<SparseNDArray<scalar_t, dims>>(ifs);
+  ifs.close();  
+
+  t_end = std::chrono::high_resolution_clock::now();
+  time_span = duration_cast<std::chrono::duration<double>>(t_end - t_start);
+  std::cout << "SparseNDArray --> Deserialization completed in " << time_span.count() << " seconds." << std::endl;
+
+  DenseNDArray<scalar_t, dims> darr_read = DenseNDArray<scalar_t, dims>::From(sparr_read);
+
+  IndexVector start_inds(dims, 0);
+  IndexVector end_inds = darr_read.shape();
+  for(IndexCounter cnt(start_inds, end_inds); cnt.running(); ++cnt) {
+    IndexVector cur_ind = cnt.index();
+    if(darr(cur_ind) != darr_read(cur_ind)) {
+      throw std::runtime_error("Error: mistake");
+    }
+  }
+
+  std::cout << "Test passed" << std::endl;
+  
+  return 0;
+}
+
 int main(int argc, char* argv[]) {
 
-  std::string ser_path = "ser_test.bin";
-  test_serialization_vector(ser_path, 1e6);
+  std::string ser_path = "/tmp/ser_test.bin";
+  
+  // test_serialization_vector(ser_path, 1e6);
+  test_serialization_sparse_array<3>(ser_path, 200);
 }