Performance Optimizations #31
Conversation
mjohnson541
added this to In progress
in Reaction Mechanism Simulator 1.0.0
via automation
|
Thanks for inviting me to review this PR! I may need some time to understand some of the changes (Luckily, you have discussed many of them in subgroups), and I will try to finish reviewing ASAP. |
| Alow[i] = falloff.arr.A | ||
| nlow[i] = falloff.arr.n | ||
| Elow[i] = falloff.arr.Ea | ||
| Ahigh[i] = 1e20 #very high limited by energy transfer process |
There was a problem hiding this comment.
Do you think this may introduce a small difference when comparing the rate we calculate in RMS with the rate from other packages? Do you think it is okay to set it as Inf to make sure the result is consistent with other packages? I just worried that there may be some case that Pr = klow * C/kinf is not far less than 1 as expected. So that rate_third_body != klow * C as defined in other packages.
There was a problem hiding this comment.
So we take the log10 of Pr so it's not numerically feasible to use Inf. k0 is collision limited and C for which ideal gas assumptions are valid is limited as well so I think this should be fine. I found no discrepencies in tests between running the individual reactions and running them vectorized.
|
|
||
| @inline function getkineticstype(kin::Chebyshev) | ||
| return (string(typeof(kin).name),kin.Tmin,kin.Tmax,kin.Pmin,kin.Pmax,size(kin.coefs)) #different opt functions, but always can do | ||
| end |
There was a problem hiding this comment.
Not sure, should we have exports for the above functions?
There was a problem hiding this comment.
It probably won't hurt, added.
| typeassert.(thermo,NASA) | ||
| return NASAvec([sp.thermo for sp in spcs]) | ||
| end | ||
|
|
| function PdepArrheniusvec(pdeparrs::T) where {T<:AbstractArray} | ||
| arrs = Array{Arrhenius,1}() | ||
| Ps = pdeparrs[1].Ps | ||
| arrs = [Array{Arrhenius,1}() for i = 1:length(Ps)] |
There was a problem hiding this comment.
redundant arrs = Array{Arrhenius,1}()?
src/Phase.jl
Outdated
| construct the stochiometric matrix for the reactions and the reaction molecule # change | ||
| """ | ||
| function getstoichmatrix(spcs,rxns) | ||
| M = zeros(length(rxns),length(spcs)) |
There was a problem hiding this comment.
For a model generated by RMG with ~50000 reactions and ~1000 reactants, I guess the RAM usage can be around 400 MB. Although it is not quite big since we are using sparse matrix anyway, can we not use zeros(length(rxns), length(spcs)) but simply create three lists and feed them to sparse()?
There was a problem hiding this comment.
fixed, used spzeros instead of zeros
|
I think I've resolved all the issues with this. |
Vectorization of kinetics and thermo calculations and general optimizations of rate constant and equilibrium constant computations. In tests this resulted in about a 5.2x speed increase on large ~400 species jobs where the rate coefficients need recomputed each evaluation.