Re: st: Stata (gllamm) benchmarks for different platforms?

[Buzz Burhans <wsb2@cornell.edu>]

Over the past few days Fred Wolfe and I have exchanged some information 
relative to differences in running Stata on the newer AMD chips under Linux 
(64-bit processing) or Win XP (32-bit processing), where the chips were 
similar (but not identical) clockspeed.  We thought that perhaps some of 
the results and other info we exchanged might be of interest to others, so 
we are submitting it for general information. I will split our exchanges 
into two posts, one about the results of our platform comparisons, the 
other containing some thoughts about working from multiple computers 
exchanged between us, some of which was extracted from an off-list exchange 
between Michael Ingre and myself. The following is a mix of our comments to 
each other, I have extracted and patched the more pertinent parts together 
from several exchanges, mostly it will not be easy to discern who is making 
which comment, which is not germaine to the information anyway.  This 
exchange got us started:


> I went a somewhat similar route, the 2.2 opteron. A velocity micro box 
> with Raid 0 and 2 GB of memory. However, I stayed with XP Pro.
>
> You know the usual XP versus Linux arguments. I decided that I needed XP 
> for more general compatability, and I figured I would take a little hit in 
> processing power by not going with Linux.
>
> However, I really don't know how much better, if any, is Linux in running 
> Stata. Perhaps you have a file and program that you have run and timed 
> that you could send to me send to me. I would run here and then we could 
> compare speed. I could run GLLAMM (on your data?)
Following this exchange, we both ran a small dataset and the same .do file 
which ran two almost identical calls to -gllamm-. The difference between 
Program1 and Program2 is that the exact same model is run in both cases, 
with the single difference being that in Program1 there are 2 level 2 
random effects, and in program 2 an additional random effect was added so 
there are 3 random effects. The third RE made much much more difference in 
speed than the difference in processors, which in my opinion supports the 
idea expressed recently by Philp Ryan and Ronan Conroy that integrating 
-gllamm- into Stata via either incorporating more into the core code, as 
was previously done, or by using a plugin, are merited and probably the 
more important issue than processor differences.  I heartily second Ronan 
Conroy's sentiments that I repeat here: "It's [-gllamm-] a major piece of 
intellectual work, and is really well supported with a website 
(www.gllamm.org) documentation etc. Sophia, Andrew and Anders are doing a 
remarkable job"

These were our results:


1) Both machines produced exactly the same numeric results for the GLLAMM 
programs.
2) The Linux machine was faster.

Program 1
Linux                         Windows XP                           % Faster
 16.75                            25.51         minutes         29.0%

Program 2
13.18                            18.47          hours           34.3%


The Win XP machine  was manufactured by Velocity Micro 1 month age, and as 
far as I could determine at the time was the fastest Windows XP Pro machine 
available.
It uses a single Atholon 64 FX-51 Processor at 2.4 GHz (front bus 200 Mhz). 
It has 2 GB of memory. Stata ran in a 200 MB partition. The data and 
overhead for the programs used 47.4K of memory. Disks were 2 Raid 0 Sata at 
10,000 RPM.


The Linux setup is running under SuSE v9 on a box supplied by Penguin 
Computing Inc. about a month ago.  It is running an Opteron 246 chip 
at  2.0 MHz (there is a faster Opteron available from Penguin, the 248 @ 
2.2 MHz ).The hard drive was an 80GB Seagate IDE @ 7200 RPM.  The speed 
differences here, it must be remembered, were between two newer chips, my 
original comparison with a Pentium 4 2.0 MHz chip had the Linux Opteron 
running at a 370% improvement in speed.


Although the Windows machine was slower than the Linux machine, it is still 
very fast. Stata graphs that were rather slow in my 4 year old Windows 2000 
machine run at a very acceptable speed on the new machine, about 3.5 times 
faster. Non-Windows users on the list often comment on problems of 
stability and compatibility in the Windows OS. My machine has been rock 
solid, and there have been no problems whatsoever (Windows and its programs 
have improved over the years).

My machine came with "Windows Media Player." I thought I would have little 
interest in it. But it plays music beautifully. Playing Bach Partitas makes 
Stata run 40% faster, or so it seems.

There is now available a beta version of a 64 bit XP. If this version 
becomes available this year we would expect the speed of XP on a 64 bit 
Athlon to increase substantially.

Compared to the data set that Buzz provided for the test, I usually run 
data sets that have 100 to 1,000 time more observations. One would hope 
that future increases in the speed of GLLAMM with DLLs might make it more 
usable.





Buzz Burhans
wsb2@cornell.edu

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/