Optimized MPI simulator

qiskit/providers/aer/backends/aer_simulator.py

@@ -238,6 +238,13 @@ class AerSimulator(AerBackend):
       This option should be set when using option ``blocking_enable=True``
       (Default: 0).
 
+    * ``chunk_swap_buffer_qubits`` (int): Sets the number of qubits of
+      maximum buffer size (=2^chunk_swap_buffer_qubits) used for multiple
+      chunk-swaps over MPI processes. This parameter should be smaller than
+      ``blocking_qubits`` otherwise multiple chunk-swaps is disabled.
+      ``blocking_qubits`` - ``chunk_swap_buffer_qubits`` swaps are applied
+      at single all-to-all communication. (Default: 15).
+
     * ``batched_shots_gpu`` (bool): This option enables batched execution
       of multiple shot simulations on GPU devices for GPU enabled simulation
       methods. This optimization is intended for statevector simulations with
@@ -553,6 +560,7 @@ def _default_options(cls):
             # cache blocking for multi-GPUs/MPI options
             blocking_qubits=None,
             blocking_enable=False,
+            chunk_swap_buffer_qubits=None,
             # multi-shots optimization options (GPU only)
             batched_shots_gpu=True,
             batched_shots_gpu_max_qubits=16,

releasenotes/notes/MPI-chunk-swap-optimization-8e693483ed271583.yaml

@@ -0,0 +1,7 @@
+---
+upgrade:
+  - |
+    MPI parallelization for large number of qubits is optimized to apply
+    multiple chunk-swaps as all-to-all communication that can decrease
+    data size exchanged over MPI processes. This upgrade improve scalability
+    of parallelization.

src/simulators/density_matrix/densitymatrix_state.hpp

@@ -313,6 +313,9 @@ class State : public Base::StateChunk<densmat_t> {
   //-----------------------------------------------------------------------
   //swap between chunks
   void apply_chunk_swap(const reg_t &qubits) override;
+
+  //apply multiple swaps between chunks
+  void apply_multi_chunk_swap(const reg_t &qubits) override;
 };
 
 //=========================================================================
@@ -1966,6 +1969,34 @@ void State<densmat_t>::apply_chunk_swap(const reg_t &qubits)
   BaseState::apply_chunk_swap(qs1);
 }
 
+template <class densmat_t>
+void State<densmat_t>::apply_multi_chunk_swap(const reg_t &qubits)
+{
+  reg_t qubits_density;
+
+  for(int_t i=0;i<qubits.size();i+=2){
+    uint_t q0,q1;
+    q0 = qubits[i*2];
+    q1 = qubits[i*2+1];
+
+    std::swap(BaseState::qubit_map_[q0],BaseState::qubit_map_[q1]);
+
+    if(q1 >= BaseState::chunk_bits_){
+      q1 += BaseState::chunk_bits_;
+    }
+    qubits_density.push_back(q0);
+    qubits_density.push_back(q1);
+
+    q0 += BaseState::chunk_bits_;
+    if(q1 >= BaseState::chunk_bits_){
+      q1 += (BaseState::num_qubits_ - BaseState::chunk_bits_*2);
+    }
+  }
+
+  BaseState::apply_multi_chunk_swap(qubits_density);
+}
+
+
 //-------------------------------------------------------------------------
 } // end namespace DensityMatrix
 //-------------------------------------------------------------------------