[R] R 2.0.0 (Windows): slow startup over the network

[Thomas Petzoldt]

Prof Brian Ripley wrote:

 > On Wed, 6 Oct 2004, Thomas Petzoldt wrote:
 >
 >
>> I installed R 2.0.0/Windows on our network server (Win NT 4.0) with
>> all CRAN packages and observed a quick startup on this machine
>> (3s).
 >
 > That's actually very slow: decent machine are well under 1s.

The NT server is optimized for network traffic, not for running local 
applications and it's not brand new.

>> However, when started from different client machines (e.g. Win NT
>> 4.0 or Win XP Pro SP2, Athlon XP 1700, 100 Mbit Network) the R Gui
>> Window appears immediately but then it takes 90 seconds to get the
>> startup message and the command prompt.
>> 
>> The R base configuration (without additional packages) takes only 4
>>  seconds on both, the server and the client.
 >
 > So it's slower without the additional packages!

Yes, but the difference lies within the confidence interval (median of 
two experiments with sample size n=3). The load of the server during the 
test was very low, but not zero and there might be some autocorrelation 
between the trials.

> Read the NEWS file, which says
> 
> As the lazy-loading databases will be consulted often, R will be
> slower if run from a slow network-mounted disc.

I see, but I don't yet understand why additional packages which are 
*not* loaded do require startup time.

> and I have only seen this with Windows network-mounted discs. With
> our 7-year-old Sun Sparc 1 server running Samba it takes about 25s,
> and that is a pretty extreme situation (and that server is about to
> be retired), which were are solving by installing R 2.0.0 locally.

This is not a good solution, but I may try a mixed approach. Small 
version on the network and complete versions on dedicated clients.

> Your network looks very slow, and maybe needs checking out. No one 
> checking alpha or beta versions has reported any problems, including
> you.

During the testing phase with the development version of R 2.0 I 
installed only the necessary packages, not all, but from time to time 
it's nice to look into some more contributed packages.

I can do some additional tests, e.g. on a Samba share and see, if I can 
isolate the problem.

> I do plan to address this in a future version, and did not do for
> 2.0.0 precisely because no end-user reported any problem during
> testing.

Thank you very much!

Thomas P.


-- 
Thomas Petzoldt
Dresden University of Technology
Institute of Hydrobiology          petzoldt at rcs.urz.tu-dresden.de
01062 Dresden                      http://www.tu-dresden.de/fghhihb/