Transition diffeqr to JuliaConnectoR? about diffeqr

[stefan-m-lenz]

Thanks for asking! I'll take a closer look ...

[stefan-m-lenz]

If you "translate" the example from the documentation of DifferentialEquations to R with the JuliaConnectoR, it looks like this:

Original Julia code:

using DifferentialEquations
f(u,p,t) = 1.01*u
u0 = 1/2
tspan = (0.0,1.0)
prob = ODEProblem(f,u0,tspan)
sol = solve(prob,reltol=1e-6,save_everystep=false)
sol(0.45)

"Translated" R code:

Loading the stuff:

library(JuliaConnectoR)

# if it's not installed already
Pkg <- juliaImport("Pkg")
Pkg$add("DifferentialEquations") 

de <- juliaImport("DifferentialEquations")

Then the real action:

f <- function(u,p,t) {1.01*u}
u0 <- 1/2
tspan <- c(0.0, 1.0)
prob <- de$ODEProblem(f,u0,tspan)
sol <- de$solve(prob, de$Tsit5(), reltol=1e-6, save_everystep=FALSE)
sol[[2]]

The JuliaConnectoR doesn't get (yet) that sol is also usable as function, so we need to use map to evaluate it:

juliaCall("map", sol, 0.45)

It is also possible to specify fas a Julia function. This is much more efficient because no communication between Julia and R is necessary during solving the equation.

f <- juliaEval("(u,p,t) -> 1.01u")
prob <- de$ODEProblem(f,u0,tspan)
sol <- de$solve(prob, de$Tsit5(), reltol=1e-6, save_everystep=FALSE)

There is one caveat: The JuliaConnectoR was designed with convenience and stability for interactive use. That's why we chose TCP as communication mechanism between Julia and R. But, of course, TCP has a higher cost for the communication when making calls between Julia and R. In every step of the iteration some binary numbers must be sent between Julia and R if the function that is to be optimized is an R function. (If it's a Julia function, this doesn't matter.) There would probably be the need for doing some further benchmarking against the current implementation of diffeqr to decide.

[stefan-m-lenz]

I tried to do some benchmarking, but I got an error in diffeqr. That is the code that I tried:

de <- juliaImport("DifferentialEquations")
f <- function(u,p,t) {1.01*u}
u0 <- 1/2
tspan <- c(0.0, 1.0)
system.time(for (i in 1:100) {
    prob <- de$ODEProblem(f,u0,tspan)
    sol <- de$solve(prob, reltol=1e-6)
})


library(diffeqr)
diffeqr::diffeq_setup() # edit/update (see below)
f <- function(u,p,t) {1.01*u}
u0 <- 1/2
tspan <- list(0.0, 1.0)
system.time(for (i in 1:100) {
    ode.solve(f, u0 = u0, tspan = tspan, reltol=1e-6)
})

I get the following error with diffeqr when running ode.solve:

> diffeqr::ode.solve(f, u0 = u0, tspan = tspan, reltol=1e-6)
Error: Error happens in Julia.
UndefVarError: ODEProblem not defined
Stacktrace:
 [1] top-level scope at none:0
 [2] eval(::Module, ::Any) at ./boot.jl:331
 [3] eval_string(::String) at /opt/R/R-4.0.2/lib/R/library/JuliaCall/julia/setup.jl:195
 [4] docall(::Ptr{Nothing}) at /opt/R/R-4.0.2/lib/R/library/JuliaCall/julia/setup.jl:168

I don't know what is wrong. I use R version 4.0.2, Julia version 1.4.2 and DifferentialEquations version 6.15.0.

[ChrisRackauckas]

Did you run diffeqr::diffeq_setup()?

[stefan-m-lenz]

No, after running diffeqr::diffeq_setup() it worked. (I added it to the code above.)
The above benchmark shows the following results on my computer (measured after first run):
When executing the code without the loop, the speed difference is not noticable for the human perception, I would say. With the loop, there is a significant difference:

Elapsed time for calling the diffeqr code in the loop above a 100 times: 0.4s

Elapsed time for calling the JuliaConnectoR code in the loop above a 100 times, using an R function: 29s

Elapsed time for calling the JuliaConnectoR code in the loop above a 100 times, using a Julia function: 1.4s

In this case, we have the additional overhead of two calls and one additional assignment for the JuliaConnectoR vs one call in the diffeqr implementation, which explains that there is still a difference when using the Julia function.

I also report that during one time of exeuting the diffeqr code, I had a segfault/system error, which crashed R.

[ChrisRackauckas]

diffeqr is probably more suitable for this specific project, though JuliaConnectoR is great!