This blog marks end of a really wonderful experience with amazing people! Special shoutout to Chris Rackauckas, Yingbo Ma and David Widmann who immensely helped me in getting my work done and helping me throughout this whole summer. I would always be greatful to them for accepting me as a student.
Regarding my work, my project dealt with more of general fixes for many of the JuliaDiffEq repositories. My work was focussed the most on Callbacks, Non Linear solvers and other Derivative utilities. It involved a lot of moving code back and forth from one repository to another as we were making the common tools in all the repositories shifted to DiffEqBase. Also I made a small benchmarking service called DiffEqBot for the organization which also got featured on JuliaLang Blog! Below is the list of my contribution grouped by the type of work -
We used to have seperate copies of the NLSolver methods in each of the repositories. All of them were basically same, but with small changes. We made that common for all of them and shifted it to DiffEqBase. We are further refactoring it right now to an even more extendible interface.
| Repository | PR | Title |
|---|---|---|
| DiffEqBase | #209 | Shift NLSolve code to DiffEqBase |
| OrdinaryDiffEq | #721 | Moves NLSolve code to DiffEqBase |
| StochasticDiffEq | #141 | Jacobian Reuse in StochasticDiffEq |
| DiffEqBase | #250 | Split NLSolve macros |
| OrdinaryDiffEq | #790 | Changes for Split of nlsolve macros |
| DiffEqDiffTools | #63 | Adds resize! for JacobianCache |
| DiffEqBase | #265 | Resizing nlsolver things |
| StochasticDiffEq | #194 | Changes for NLSolver change |
| OrdinaryDiffEq | #830 | [WIP] Remove redundant NLSolve fields from alg cache |
| DiffEqBase | #297 | Resize changes for new nlsolve |
| StochasticDiffEq | #205 | Remove redundant NLSolve Fields |
| DiffEqDocs | #255 | NLSolver Documentation |
I developed a basic bot which is very similar to Nanosoldier, we call it on GitHub, it runs benchmarking scripts on a PC and posts the results comparing the PR with master. All of the work and documentaion is contained in the repository DiffEqBenchmarkServer.jl
| Repository | PR | Title |
|---|---|---|
| OrdinaryDiffEq | #743 | Adds Benchmark Suite |
| OrdinaryDiffEq | #772,#773,#774,#775 | Benchmark Time Adjustments |
| DiffEqBenchmarkServer | #1 | Complete the documentation and instructions |
| DiffEqBenchmarkServer | #2 | Remove benchmarked package after job |
Our packages have a feature called Callbacks, which helps users plug their own code when a certain event happes - like a particular variable reaches a certain value. My work was to implement VectorContinuousCallback, which is basically equivalent to CallbackSet, but it's very lightweight for the case of large number of callbacks.
| Repository | PR | Title |
|---|---|---|
| OrdinaryDiffEq | #754 | Changes for VectorContinuousCallback |
| DiffEqBase | #229 | Adds CallbackCache |
| DiffEqBase | #231 | Fixes type broken inference |
| StochasticDiffEq | #157 | Changes for VectorContinuousCallback |
| DiffEqDocs | #226 | Adds VectorContinuousCallback documentation |
| DiffEqBase | #235 | Fix callback tests failures |
| DiffEqBase | #238 | Fix OrdinaryDiffEq failure |
| Sundials | #214 | Changes for VectorContinuousCallback |
| ODEInterfaceDiffEq | #27 | Changes for VectorContinuousCallback |
| DiffEqDocs | #227 | Adds VectorContinuousCallback example |
| DiffEqBase | #253 | CallbackSet fix |
| DiffEqBase | #257 | Better epsilon handling in callbacks |
| DiffEqBase | #271 | Fix missed callback |
Originally, the toolkit developed was not ready for parallel algorithms, especially the derivative utilities and nlsolver. I made the required changes and implemented first Parallel DIRK (PDIRK44). Now that one algorithm is implemented, it opens new ways for future algorithms to be developed using the same toolkit we have.
| Repository | PR | Title |
|---|---|---|
| OrdinaryDiffEq | #815 | [WIP] Parallel 4th order DIRK |
| DiffEqBase | #288 | [WIP] Make nlsolve thread safe |
We are almost there when we have OrdinaryDiffEq natively support Fully Implicit ODEs including DAEs. I aim to implement two algorithms - DImplicitEuler and DBDF. DImplicitEuler is ready, but we need changes in NLSolver, so for now, we are using NLsolver.jl package
| Repository | PR | Title |
|---|---|---|
| OrdinaryDiffEq | #863 | DAEProblem Support |
On the way down, I also fixed some small fixes -
| Repository | PR | Title |
|---|---|---|
| OrdinaryDiffEq | #751 | Qsteady Handling in JVODE, QNDF and Standard Controller |
| DiffEqBase | #255 | Fix Typo and Tests |
| OrdinaryDiffEq | #757 | Fix DifferentialEquations warning |
| StochasticDiffEq | #152 | Fix DifferentialEquations warning |
| DelayDiffEq | #105 | Fixes unwrap_cache error |
| DelayDiffEq | #108 | Test Failure Fixes |
| OrdinaryDiffEq | #796 | Fix analytic functions in tests |
| OrdinaryDiffEq | #812 | [WIP] Resize remaining fields |
| DiffEqDocs | #251 | SKenCarp warning |
| DiffEqProblemLibrary | #50 | fix Brusselator Problem |
| StochasticDiffEq | #223 | [WIP] SKenCarp convergence fix |
| OrdinaryDiffEq | #881 | Fix ImplicitEulerExtrapolation |
I have decided that I will stay with JuliaDiffEq and get involved in the community more. People here are amazing and I have never seen a community that works towards efficiency in code more than this. GSoC was a great way of getting involved in JuliaDiffEq and Julia organizations. I want to complete DAE Support and then move on to ModelingToolkit.jl, a toolkit for modeling DSLs. Much more things I need to explore, so much more to learn.