How to improve performance

[efaulhaber]

This issue is to discuss options to improve single-processor performance (multithreaded, but no MPI yet).

The plan so far:

• Solve wrap_array allocates #29 and resolve all type instabilities (solved in Remove type instabilities and allocations #36).
• Investigate if thread=OrdinaryDiffEq.True() can improve performance of the time integration method once StackOverflowError in broadcasting JuliaSIMD/StrideArrays.jl#62 is resolved. Edit: We are now using ThreadedBroadcastArrays for this.
• Further optimize PK1 computation for structure dynamics (solved in Compute deformation gradient in matrix form #38).
• Use symmetry of interactions? We could skip half of the fluid-fluid interaction by applying the same force with a flipped sign to the neighbor particle. However, some SPH codes don't do this because it makes the computations and memory accesses less optimal, especially on GPUs. We could also skip the solid-fluid interaction by using symmetry in the fluid-solid interaction.
• A single @threaded loop over all particles, which then includes all fluid-* interactions could potentially improve performance. Edit: Benchmarked in Combine interact! kernels to reduce GPU latency #1082. Didn't make a difference for large enough problems.
• Improve neighborhood search update. The NHS update is a bottleneck on multiple threads because the implementation does not use multithreading (How to improve neighborhood search #65).
• Rework the neighborhood search. We might be able to get a significant speedup for large simulations by using a contiguous memory layout. Edit: Benchmarked in Compact Vector of Vectors PointNeighbors.jl#120