Our code is on GitHub here: https://github.com/ImperialCollegeLondon/covid19model/ I’m running base.r from Rscript (i.e. Rscript base.r; this runs 4 chains for 200 iterations, 100 warmup; takes < 4 minutes). Everything runs out of the box on my Macbook Pro, with R version 3.5.1 and RStan version 2…

Maybe this model should be rewritten to take advantage of within-chain parallelism with the newest reduce_sum feature about to land in 2.23? Do you have sufficient cores to drive this? I don’t know the data sizes, but if your data-likelihood evaluation involves lots of terms, then this should speed…

Please give us a hand, we have many many cores (thanks to the generosity of Microsoft and Amazon) and this could be a huge speedup! How do we get started?

Is there a way for me to run the model with some fake data (ideally a subset)? Then I can edit the stan file directly. You need to install the current release candidate to get started; next week should be the final release out (some slight changes will happen to the facility, but nothing big). EDI…

Here’s a branch where we can collaborate ( https://github.com/flaxter/covid19model/tree/parallel )–it’s setup to run with a subset of real data, look in parallel.r. (Happy to switch to synthetic data if that’s important.) Are you wds15 on github? I’ll add you.

Yes, I am wds15 on GitHub. Thanks for the branch.

Here we go. This was an interesting case and I had to write two flavours of this. Since I decided that I want to parallelise over countries, I created my own data-set using the base.r script with the defaults. This gave me 14 countries and 50ish observations per country. My first attempt was to do …

Very cool, looking forward to trying this on longer runs and more cores. Just to check–I need to install cmdstan 2.23 release candidate, right? There’s no RStan 2.23 release candidate?

You need cpus with a lot of fast cpu cache to improve the situation. Right now it looks as if speed ups are somewhat limited beyond 3 cores with the data set I used. Possibly more data will allow you to speed up more with increasing cpu count (because then a larger fraction of cpu time is spend in t…

[image] wds15: now I am getting these timings on my 4-core laptop for 25 warmup / 25 iterations on the base data set: I’m getting really big variance in the timings from chain to chain when chains are this short, so I’d be wary of these timings. Unless you’ve done something extra to eliminate…

Convergence / latest version of Stan issues with our covid-19 model?

Modeling

mike-lawrence April 16, 2020, 2:48pm 47

I think I also see opportunities to use the performance trick posted here. I’ll start working on that now and let you know if I have any success with a pull request.

Topic		Replies	Views
Speed issues since upgrading to RStan v2.21.2 rstanarm	39	1317	August 30, 2020
Stuck at Warmup iteration with no error : CmdStanR CmdStan techniques , fitting-issues	48	3627	April 21, 2020
How to speed up sampling in rstan? Modeling rstan , performance , hierarchical-model	6	6269	July 24, 2020
The sampling has no response, when rstan 2.21 run in R4.0.5 General rstan , fitting-issues	15	1226	September 15, 2021
Stan on computing cluster: strange results CmdStan	11	1713	June 8, 2018

Convergence / latest version of Stan issues with our covid-19 model?

Related topics