szilard

sqlite vs R's data.table

TLDR;   sqlite (:memory:) 250 sec    data.table 7 sec


data: 100 million rows, 1 million groups
generated by: https://github.com/szilard/benchm-databases/blob/master/0-gendata.txt

szilard

########### R

library(h2o)

h2oServer <- h2o.init(max_mem_size = "50g", nthreads = -1)

d <- h2o.importFile(h2oServer, path = "d.csv")

system.time({

szilard

library(nnet)
library(h2o)
h2o.init()

set.seed(123)

n <- 1000

x1 <- runif(n)

szilard

## count

benchmark(
  nrow(d[dm, nomatch=0L, on="x"]),
  d[dm, .N, nomatch=0L, on="x"],
replications = 5, columns = c("test", "replications", "elapsed", "relative"))

#                                test replications elapsed relative
#2   d[dm, .N, nomatch = 0, on = "x"]            5  28.535    1.000
#1 nrow(d[dm, nomatch = 0, on = "x"])            5  38.562    1.351

szilard

system.time(z <- 1:1e9)

system.time(d <- data.frame(x = 1:1e9))
system.time(d$y <- 1:1e9)
system.time(d$z <- z)

system.time(d$x[1] <- 0L)

szilard

library(h2o)
h2o.init(max_mem_size = "20g", nthreads = -1)

# R is connected to the H2O cluster: 
#   H2O cluster uptime:         1 seconds 704 milliseconds 
# H2O cluster version:        3.8.2.8 
# H2O cluster name:           H2O_started_from_R_szilard_lcr105 
# H2O cluster total nodes:    1 
# H2O cluster total memory:   17.78 GB 

szilard

> system.time(x <- 1:1e11)
   user  system elapsed
221.491 210.466 432.030

> object.size(x)/1e9
800.00000004 bytes

> system.time(sum(x))
   user  system elapsed
145.913  78.183 230.063

szilard

## allocate <1TB first, stuff works
> x <- 1:1.2e11
> gc()
               used     (Mb)  gc trigger      (Mb)     max used     (Mb)
Ncells       214403     11.5 4.60000e+05      24.6       350000     18.7
Vcells 120000397293 915530.4 1.72801e+11 1318366.9 120000398080 915530.4
> system("echo 1")
1

szilard

library(h2o)
h2o.init(nthreads = -1)

dx_train <- h2o.importFile("https://s3.amazonaws.com/benchm-ml--main/train-1m.csv")


system.time({
  md_10 <- h2o.gbm(x = 1:(ncol(dx_train)-1), y = ncol(dx_train), training_frame = dx_train, 
                   model_id = "airline_depth10",

szilard

## Stochastic Variational Deep Kernel Learning
## paper: https://arxiv.org/abs/1611.00336
## code+data from the authors (thanks!!!): https://people.orie.cornell.edu/andrew/code/#SVDKL


## get data + prepare sample authors used for evaluation

wget https://people.orie.cornell.edu/andrew/code/svdklcode.zip
unzip  svdklcode.zip