benchmarks?

[floswald]

did you guys ever think about benchmarking this against https://cran.r-project.org/web/packages/quantreg/?

if i remember correclty this wraps fortran routines, could think about using those as well (if performance here is far below acceptable)

[pkofod]

Havn't done that. Quantreg's default (if I'm not mistaken) is the interior point method in Portnoy and Koenker, and that is exactly the method I've coded up here and chosen as the default. For large problems it is far faster than the IRLS algorithm we also have here.

[floswald]

that's good news! I just looked at the quantreg package. that's a lot of fortran. anyway, would be interesting to see how good or bad we are doing here.

[pkofod]

Sure, that would be interesting... I originally had the idea to go all in, but then I realized how much work Koenker had already put into quantreg and fainted :)

[pkofod]

Are you an R user? What is the standard way of benchmarking in R?

[floswald]

well, I used to be :-)

you would just time execution of a command with system.time( command ). for example, from the demo folder in the quantreg source:

> library(quantreg)
> data(engel)
> elapsed_time <- system.time(z <- rq(foodexp ~ income, tau= .50, data = engel))
> elapsed_time
   user  system elapsed 
  0.004   0.000   0.008 
> # we want elapsed
> elapsed_time[3]
elapsed 
  0.008 

[floswald]

i just submitted a pull request to Query.jl that does an R vs julia benchmark. maybe useful?
queryverse/Query.jl#154

[yanyu2015]

Yes, a benchmarks is important for people to choose jl over R.