Drastic difference in the mean running times of Julia and Python

[dinskid]

The benchmarking (if it is correct) shows results where Julia's version of solveSystem is really slow when compared to Python's version of the same

Python results

Julia results

[ChrisRackauckas]

Do you want me to take a look? What are you running?

[dinskid]

Sure, the code has been merged in the last commit. I am running Tsit5() on it (if that is what you asked).

[ChrisRackauckas]

what integrator are you using in SciPy?

[dinskid]

I don't have a great understanding about these, but I think it is RK23.
This is the corresponding python code.

out = integrate.solve_ivp(
            fun = lambda t,y: dydt_Complete(t,y, **kwargs),
            t_span=(0.,total_days),
            y0 = cur_stateTensor,
            method='RK23',
            t_eval=range(total_days),
            rtol = 1e-3, #default 1e-3
            atol = 1e-3, # default 1e-6
        )

Sorry about the previous comment, the code is in the PR ( it hasn't been merged yet ).

[ChrisRackauckas]

Alright. I'll follow the benchmark and when you merge it I'll take a look.

[vollmersj]

I have just reviewed the PR and merged it. Maybe the issue is #6 the einsum hack is allocating arrays at every step - this might explain the 637mb allocation for the solving the ODE (the transition tensor should be less < 1MB).

[ChrisRackauckas]

Alright, I'll take a look and see what can be done. Indeed, 637mb is far too much.

[ChrisRackauckas]

Accelerated by 36,000x in #24