SOLVCON v0.1 and Beyond

Create date: 2011/8/11

SOLVCON v0.1 marks a milestone. Before version 0.1, the development of SOLVCON focused on the proof of concept of using Python to construct a software framework for a multi-physics, highly-parallelized system that solves for conservation laws based on unstructured meshes. SOLVCON includes two distinct physical models: One for compressible flows and the other for stress waves in anisotropic elastic solids. The code can deliver more than 0.3 million elements/second on a Xeon or Opteron by using OpenMP, which is a reasonable speed for time-accurate CFD solvers in production use. The calculations can be scaled up to use more than 1,000 CPU cores. In addition to these results, SOLVCON showed tremendous potentials by incorporating in situ visualization, hybrid parallelism, and GPGPU computing (on clusters).

After summarizing the proof of concept by releasing version 0.1, I can now plan for further development of SOLVCON.

The CESE Method

The CESE method will still be employed as the default numerical method of SOLVCON. The current development of the CESE method implementation in SOLVCON is directed to high-order schemes. The high-order CESE method are being implemented by David Bilyeu and me. A prototype of the fourth-order Euler solver has been worked out and is awaiting validation and consolidation.

For complex physical processes that involve more than 10 or 20 conservation variables, the implementation of the CESE method should use compress sparse row (CSR) or a similar format for the buffers of the Jacobian matrices, in order to same memory and potentially to speed up.

Scalability and MPI

Scalability of SOLVCON will be further improved by making the distributed-memory parallel computing comply to the single-process (thread-less), massively-parallelized, message-passing architecture.

The change will also facilitate simplification of the executional model. Parallel calculations will be launched through standard MPI utilities only. Currently, SOLVCON can be launched by MPI utilities or its own remote procedure call (RPC) sub-system. In the future the custom sub-system will be discontinued.

The way for SOLVCON to invoke MPI will also change. Currently, MPI libraries are invoked through ctypes. This approach is cumbersome for deploying SOLVCON onto various supercomputers/clusters. Changing to use mpi4py will solve the issue. mpi4py will introduce an abstraction layer that deals with the wrapping of MPI.

These changes will make the code base more compact and easier to be maintained.

Code for domain decomposition will be further optimized for speed and memory, in order to efficiently use more than 100 million elements in routine simulations along with in situ visualization.

Physical Models

The immediate development task after the release of version 0.1 is to add viscous terms in the compressible flow solver. Based on the 2nd-order CESE method, the treatment of the new diffusive fluxes will be added to the implementation to complement the existing treatment of the convective fluxes.

Two more physical models will be incorporated into SOLVCON into the name space solvcon.kerpak: (i) waves in viscoelastic media and (ii) liquid flows.

Other physical processes under planning include electromagnetic waves and stress waves in piezoelectric solids.