Slow performance on large data set

[xcompass]

It seems the library has slow performance on larger data set. Here is the example:

use Statistics::R::IO;

my $c = Statistics::R::IO::Rserve->new(server => 'r', _usesocket => 1);

$c->eval('n <- 1500');
$c->eval('x1 <- round(runif(n, min=-1, max=7),3)');
$c->eval('x2 <- round(runif(n, min=-1, max=7),3)');
$c->eval('x3 <- round(rnorm(n, sd=3),3)');
$c->eval('x4 <- round(x3 + abs(rnorm(n, sd=2)),3)');
$c->eval('x5 <- round(runif(n, min=-1, max=7),3)');
$c->eval('x6 <- round(sample(x1),3)');
$c->eval('x7 <- round(sample(x3),3)');
$c->eval('x8 <- round(sample(x2),3)');
$c->eval('x9 <- round(x4 + x1 + rnorm(n, sd=.5),3)');
$c->eval('x10 <- round(x2 + rnorm(n, sd=.5),3)');
my $r = $c->eval('x <- cbind(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10)');
print $r;
$c->close();

root@6b3ba909a6e2:/opt/webwork# time perl test.pl
....
real    0m3.931s
user    0m3.890s
sys     0m0.030s

Same R statements running in python client:

import pyRserve

conn = pyRserve.connect(host='r');

conn.eval('n <- 1500');
conn.eval('x1 <- round(runif(n, min=-1, max=7),3)');
conn.eval('x2 <- round(runif(n, min=-1, max=7),3)');
conn.eval('x3 <- round(rnorm(n, sd=3),3)');
conn.eval('x4 <- round(x3 + abs(rnorm(n, sd=2)),3)');
conn.eval('x5 <- round(runif(n, min=-1, max=7),3)');
conn.eval('x6 <- round(sample(x1),3)');
conn.eval('x7 <- round(sample(x3),3)');
conn.eval('x8 <- round(sample(x2),3)');
conn.eval('x9 <- round(x4 + x1 + rnorm(n, sd=.5),3)');
conn.eval('x10 <- round(x2 + rnorm(n, sd=.5),3)');
result = conn.eval('x <- cbind(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10)');
print result;
conn.close();

root@6b3ba909a6e2:/opt/webwork# time python test.py
...
real    0m0.291s
user    0m0.080s
sys     0m0.110s

Native Rscript:

n <- 1500
x1 <- round(runif(n, min=-1, max=7),3)
x2 <- round(runif(n, min=-1, max=7),3)
x3 <- round(rnorm(n, sd=3),3)
x4 <- round(x3 + abs(rnorm(n, sd=2)),3)
x5 <- round(runif(n, min=-1, max=7),3)
x6 <- round(sample(x1),3)
x7 <- round(sample(x3),3)
x8 <- round(sample(x2),3)
x9 <- round(x4 + x1 + rnorm(n, sd=.5),3)
x10 <- round(x2 + rnorm(n, sd=.5),3)
x <- cbind(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10)
print(x)

root@8f3ee1c31d84:/# time Rscript test.R
...
real    0m0.146s
user    0m0.090s
sys     0m0.030s

Python is slow than native Rscript, but still comparable. But perl client is more than 10x slower. I wonder if there is something inefficient or a bug in the client.

I tested on both remote and local host to Rserve.

R version 3.1.1
Rserve 1.8-5, 1.7-3, 0.6-8 all similar results.

[cubranic]

Can you check the performance of just a single assignment? (E.g., x1) I want to check whether it's any object, even a simple vector (x1), or if it's especially the more complex objects, like matrices (i.e., cbind).

The library was never optimized for performance, it's kind of like a recursive parser to make it easy to work from the protocol spec.

[cubranic]

Also, I suggested to the problem creators to avoid returning those objects to webwork if they don't really need them. In other words, run:

$c->eval('x1 <- round(runif(10000, min=-1, max=7),3); NULL');

unless x1 is actually needed in the PG problem. I know it's just a workaround, but I don't have the time right now to deep dive into performance issues in the library.