As some of you might already know, there is an ongoing effort/project to parallelize parts Stan on the GPU for faster inference ( stanmathcl ). We are currently in the process of empirically evaluating speedups we can achieve with what has already been parallelized (Cholesky, multiplication…) and t…

I have a complex economic model of auto insurance consumer choices (i.e. which plans people choose over time) and their accident/loss distributions (estimated jointly). I can’t share the real data since it is very much proprietary but can share a simulated DGP + model code. Currently the model takes…

I should say that part of what makes the model slow is that we have to compute the “menu of options” (i.e. what consumers would expect to pay) for every possible insurance plan, which boils down to lots of matrix multiplication and simple (but numerous) integration - i.e. things that could benefit g…

Have you considered MPI? If you split things into independent computations then MPI is great. Of course, you can also combine the techniques… which has never been done in stan and would be interesting to see.

Currently we only have the Cholesky Decomposition on the GPU, but multiplication is probably one of the next places we will work on. I agree with @wds15 , if you can split the problem up it sounds like you could get a quicker win with MPI.

You can use MPI with Stan? (I didn’t know what that was and needed to Google it so forgive me if the question is silly)

MPI is coming, so the question is not silly at all. See here: [image] MPI Design Discussion Developers Ok, I have shown the good performance already a few times of MPI, but to this end mostly with synthetic examples. I just compiled a realistic example Hierarchical ODE …

Thanks for all the comments and suggestions! @fabio : We wanted to add a spatial covariance model and I was already looking at that case study. It is close to the top of our list and we will look at it very soon! @mike-lawrence : We are already looking at a GP and we’ll add that hierarchical model …

Awesome! What sort of timeline are you guys thinking of? I can probably share a version of the model + simulated data reasonably soon if you guys or @wds15 would be willing to see if MPI and/or GPU-parallelization could help + how 0-:)

Really exciting - thank you for the response!

MPI has the bottleneck of costly inter process communication. So a hierarchical model where the computation cost per unit is big are a natural candidate to speed things up seriously. The link I posted was the best benchmark I ever saw! But it is a real model for one of our projects. The thing is tha…

Stan on GPU: looking for model+dataset examples for empirical evaluation of speedups

General

wds15 January 25, 2018, 9:12pm 17

MPI is coming, so the question is not silly at all. See here:

However, the first implementation can speed up things dramatically, but it is not truly user friendly since you have to cram your data into the format we expect it for easy distribution via MPI.

Topic		Replies	Views
Integrating GPU support Developers	23	5683	January 31, 2017
Stan on the GPU Project Proposals	16	8635	August 10, 2018
Speeding up hierarchical models Developers	26	2705	June 5, 2020
Making big kernels work General	34	2179	August 10, 2018
Question Regarding Possible Speed Gains ( GPU or CPU ) for a Four Level Hierarchical Logistic Regression Model Modeling fitting-issues , specification , gpu , cmdstanr	7	1147	July 22, 2021

Stan on GPU: looking for model+dataset examples for empirical evaluation of speedups

Related topics