Threading and performance

A plurality (if not majority) of the time spent in MPO construction is in performing the sparse QR decomposition. This can often be sped-up significantly by loading MKL.

try
  using MKL
catch
end

ITensorMPOConstruction can take advantage of multiple threads in a variety of ways.

1. ITensorMPS.add!(os::OpIDSum, ...) is thread-safe. This can be used to construct a OpIDSum in parallel.

2. During the MPO construction Threads.@threads are used, primarily to iterate over the connected components (i.e. QN blocks) in parallel.

3. In the sparse QR decomposition that is done for every connected component. The threading here is controlled by BLAS.set_num_threads. Over subscription is possible since this QR decomposition is itself called from within Threads.@threads.

Threading tips

• The MPO construction is memory bound, and the performance gain (or loss!) from multi-threading is system and operator dependent.
• If the MPO has lots of connected components (i.e. is sparse), use Julia threads.
• If the MPO has few connected components (i.e. is dense) use BLAS threads.