Skip to content

Latest commit

 

History

History
44 lines (27 loc) · 1.55 KB

README.md

File metadata and controls

44 lines (27 loc) · 1.55 KB

Stan Rocker images

These images contain RStudio, tidyverse, and Stan/brms. rocker_stan_base builds off of Rocker's tidyverse image, whereas rocker_stan_vice builds from a CyVerse VICE image.

Benchmarking Stan speed

Quick semi-realistic test of speed running brms models in RStudio on various platforms.

Platforms

gpu06: a speedy 256-core Ubuntu node

local: non-Docker on 16" MacBook Pro 2.4 GHz 8-Core Intel Core i9 32GB RAM

VICE: CyVerse VICE cloud, set with 8 cores, 16GB RAM, 64GB scratch

NOTE I ran local while doing other stuff on my laptop (writing some code, watching a video), which is, for my money, the most honest comparison to cloud-based options, since I can also do other stuff while I'm running models remotely.

Code

library(brms)
library(tidyverse)

install.packages("microbenchmark")


microbenchmark::microbenchmark({
  m1 <- brm(wind ~ year + (1 | status), data = storms, chains = 4, cores = 4)
},
times = 10)

Run times (seconds)

method min lq mean median uq max
gpu06 376.9 390.5 427.7 417.2 459.9 503.1
local 482.9 590.4 646.1 629.4 703.6 846.0
VICE 591.0 647.5 669.3 650.8 706.7 737.5

Conclusions

gpu06 wins decisively, but my fairly beefy MacBook Pro was really not that much faster than VICE, and actually had significantly higher variance in speed (probably due to uneven loads on my computer while I was doing other tasks). I think this is a pretty compelling case for using VICE.