A. Ghaffarizadeh, S.H. Friedman, and P. Macklin. BioFVM: an efficient, parallelized diffusive transport solver for 3-D biological simulations. Bioinformatics, 2015.
DOI: 10.1093/bioinformatics/btv730 (free; open access)
BioFVM (stands for "Finite Volume Method for biological problems) is an open source package to solve for 3-D diffusion of several substrates with desktop workstations, single supercomputer nodes, or even laptops (for smaller problems). We built it from the ground up for biological problems, with optimizations in C++ and OpenMP to take advantage of all those cores on your CPU. The code is available at SourceForge and BioFVM.MathCancer.org.
The main idea here is to make it easier to simulate big, cool problems in 3-D multicellular biology. We'll take care of secretion, diffusion, and uptake of things like oxygen, glucose, metabolic waste products, signaling factors, and drugs, so you can focus on the rest of your model.
Design philosophy and main capabilities
Solving diffusion equations efficiently and accurately is hard, especially in 3D. Almost all biological simulations deal with this, many by using explicit finite differences (easy to code and accurate, but very slow!) or implicit methods like ADI (accurate and relatively fast, but difficult to code with complex linking to libraries). While real biological systems often depend upon many diffusing things (lots of signaling factors for cell-cell communication, growth substrates, drugs, etc.), most solvers only scale well to simulating two or three. We solve a system of PDEs of the following form:
Above, all vector-vector products are term-by-term.
Solving for many diffusing substrates
We set out to write a package that could simulate many diffusing substrates using algorithms that were fast but simple enough to optimize. To do this, we wrote the entire solver to work on vectors of substrates, rather than on individual PDEs. In performance testing, we found that simulating 10 diffusing things only takes about 2.6 times longer than simulating one. (In traditional codes, simulating ten things takes ten times as long as simulating one.) We tried our hardest to break the code in our testing, but we failed. We simulated all the way from 1 diffusing substrate up to 128 without any problems. Adding new substrates increases the computational cost linearly.
Combining simple but tailored solvers
We used an approach called operator splitting: breaking a complicated PDE into a series of simpler PDEs and ODEs, which can be solved one at a time with implicit methods. This allowed us to write a very fast diffusion/decay solver, a bulk supply/uptake solver, and a cell-based secretion/uptake solver. Each of these individual solvers was individually optimized. Theory tells us that if each individual solver is first-order accurate in time and stable, then the overall approach is first-order accurate in time and stable.
The beauty of the approach is that each solver can individually be improved over time. For example, in BioFVM 1.0.2, we doubled the performance of the cell-based secretion/uptake solver. The operator splitting approach also lets us add new terms to the "main" PDE by writing new solvers, rather than rewriting a large, monolithic solver. We will take advantage of this to add advective terms (critical for interstitial flow) in future releases.
Optimizing the diffusion solver for large 3-D domains
For the first main release of BioFVM, we restricted ourselves to Cartesian meshes, which allowed us to write very tailored mesh data structures and diffusion solvers. (Note: the finite volume method reduces to finite differences on Cartesian meshes with trivial Neumann boundary conditions.) We intend to work on more general Voronoi meshes in a future release. (This will be particularly helpful for sources/sinks along blood vessels.)
By using constant diffusion and decay coefficients, we were able to write very fast solvers for Cartesian meshes. We use the locally one-dimensional (LOD) method--a specialized form of operator splitting--to break the 3-D diffusion problem into a series of 1-D diffusion problems. For each (y,z) in our mesh, we have a 1-D diffusion problem along x. This yields a tridiagonal linear system which we can solve efficiently with the Thomas algorithm. Moreover, because the forward-sweep steps only depend upon the coefficient matrix (which is unchanging over time), we can pre-compute and store the results in memory for all the x-diffusion problems. In fact, the structure of the matrix allows us to pre-compute part of the back-substitution steps as well. Same for y- and z-diffusion. This gives a big speedup.
Next, we can use all those CPU cores to speed up our work. While the back-substitution steps of the Thomas algorithm can't be easily parallelized (it's a serial operation), we can solve many x-diffusion problems at the same time, using independent copies (instances) of the Thomas solver. So, we break up all the x-diffusion problems up across a big OpenMP loop, and repeat for y- and z-diffusion.
Lastly, we used overloaded +=, axpy and similar operations on the vector of substrates, to avoid unnecessary (and very expensive) memory allocation and copy operations wherever we could. This was a really fun code to write!
The work seems to have payed off: we have found that solving on 1 million voxel meshes (about 8 mm3 at 20 μm resolution) is easy even for laptops.
By using constant diffusion and decay coefficients, we were able to write very fast solvers for Cartesian meshes. We use the locally one-dimensional (LOD) method--a specialized form of operator splitting--to break the 3-D diffusion problem into a series of 1-D diffusion problems. For each (y,z) in our mesh, we have a 1-D diffusion problem along x. This yields a tridiagonal linear system which we can solve efficiently with the Thomas algorithm. Moreover, because the forward-sweep steps only depend upon the coefficient matrix (which is unchanging over time), we can pre-compute and store the results in memory for all the x-diffusion problems. In fact, the structure of the matrix allows us to pre-compute part of the back-substitution steps as well. Same for y- and z-diffusion. This gives a big speedup.
Next, we can use all those CPU cores to speed up our work. While the back-substitution steps of the Thomas algorithm can't be easily parallelized (it's a serial operation), we can solve many x-diffusion problems at the same time, using independent copies (instances) of the Thomas solver. So, we break up all the x-diffusion problems up across a big OpenMP loop, and repeat for y- and z-diffusion.
Lastly, we used overloaded +=, axpy and similar operations on the vector of substrates, to avoid unnecessary (and very expensive) memory allocation and copy operations wherever we could. This was a really fun code to write!
The work seems to have payed off: we have found that solving on 1 million voxel meshes (about 8 mm3 at 20 μm resolution) is easy even for laptops.
Simulating many cells
We tailored the solver to allow both lattice- and off-lattice cell sources and sinks. Desktop workstations should have no trouble with 1,000,000 cells secreting and uptaking a few substrates.Simplifying the non-science
We worked to minimize external dependencies, because few things are more frustrating than tracking down a bunch of libraries that may not work together on your platform. The first release BioFVM only has one external dependency: pugixml (an XML parser). We didn't link an entire linear algebra library just to get axpy and a Thomas solver--it wouldn't have been optimized for our system anyway. We implemented what we needed of the freely available .mat file specification, rather than requiring a separate library for that. (We have used these matlab read/write routines in house for several years.)
Similarly, we stuck to a very simple mesh data structure so we wouldn't have to maintain compatibility with general mesh libraries (which can tend to favor feature sets and generality over performance and simplicity). Rather than use general-purpose ODE solvers (with yet more library dependencies, and more work for maintaining compatibility), we wrote simple solvers tailored specifically to our equations.
The upshot of this is that you don't have to do anything fancy to replicate results with BioFVM. Just grab a copy of the source, drop it into your project directory, include it in your project (e.g., your makefile), and you're good to go.
Similarly, we stuck to a very simple mesh data structure so we wouldn't have to maintain compatibility with general mesh libraries (which can tend to favor feature sets and generality over performance and simplicity). Rather than use general-purpose ODE solvers (with yet more library dependencies, and more work for maintaining compatibility), we wrote simple solvers tailored specifically to our equations.
The upshot of this is that you don't have to do anything fancy to replicate results with BioFVM. Just grab a copy of the source, drop it into your project directory, include it in your project (e.g., your makefile), and you're good to go.
All the juicy details
The Bioinformatics paper is just 2 pages long, using the standard "Applications Note" format. It's a fantastic format for announcing and disseminating a piece of code, and we're grateful to be published there. But you should pop open the supplementary materials, because all the fun mathematics are there:
- The full details of the numerical algorithm, including information on our optimizations.
- Convergence tests: For several examples, we showed:
- First-order convergence in time (with respect to Δt), and stability
- Second-order convergence in space (with respect to Δx)
- Accuracy tests: For each convergence test, we looked at how small Δt has to be to ensure 5% relative accuracy at Δx = 20 μm resolution. For oxygen-like problems with cell-based sources and sinks, Δt = 0.01 min will do the trick. This is about 15 times larger than the stability-restricted time step for explicit methods.
- Performance tests:
- Computational cost (wall time to simulate a fixed problem on a fixed domain size with fixed time/spatial resolution) increases linearly with the number of substrates. 5-10 substrates are very feasible on desktop workstations.
- Computational cost increases linearly with the number of voxels
- Computational cost increases linearly in the number of cell-based source/sinks
What's next?
- As MultiCellDS (multicellular data standard) matures, we will implement read/write support for <microenvironment> data in digital snapshots.
- We have a few ideas to improve the speed of the cell-based sources and sinks. In particular, switching to a higher-order accurate solver may allow larger time step sizes, so long as the method is still stable. For the specific form of the sources/sinks, the trapezoid rule could work well here.
- I'd like to allow a spatially-varying diffusion coefficient. We could probably do this (at very great memory cost) by writing separate Thomas solvers for each strip in x, y, and z, or by giving up the pre-computation part of the optimization. I'm still mulling this one over.
- I'd also like to implement non-Cartesian meshes. The data structure isn't a big deal, but we lose the LOD optimization and Thomas solvers. In this case, we'd either use explicit methods (very slow!), use an iterative matrix solver (trickier to parallelize nicely, except in matrix-vector multiplication operations), or start with quasi-steady problems that let us use Gauss-Seidel iterative type methods, like this old paper.
- Since advective flow (particularly interstitial flow) is so important for many problems, I'd like to add an advective solver. This will require some sort of upwinding to maintain stability.
- At some point, we'd like to port this to GPUs. However, I don't currently have time / resources to maintain a separate CUDA or OpenCL branch. (Perhaps this will be an excuse to learn Julia on GPUs.)
Well, we hope you find BioFVM useful. If you give it a shot, I'd love to hear back from you!
Very best -- Paul
