Buzz had trouble posting this:
Michael,
I seem to be having trouble with my emails to statlist, I posted this
twice and it does not seem to get through. I posted this to the list
yesterday, but never saw it appear, given your follow up questions I
will post it again this morning. Regarding benchmark comparisons for
-gllamm- on different platforms:
I was fortunate to have some comparisons available where a -gllamm-
model was run on different processors. At the time I was (and still do
part of the time) running a Pentium 4 at 2 GHz. The comparison
processors were an single proc Opteron64 at 2.0 GHz, a dual G5 at 2.0
GHz, and a dual proc Pentium at 2.8GHz.
The G5 and my P4, running at the same speed, ran essentially
similar times. The faster Pentium (2.8 GHz) ran almost exactly
proportionately faster than my P4 at 2.0GHz, which was a significant
increase, but not enough to make a hardware upgrade worth it to me.
The Opteron 64 ran 3.7 times faster than my P4, although the
clockspeed on both my P4 and the AMD Opteron were both 2.0 GHz
The 3.7 times faster was enough incentive for me to make a change, since
my GLLAMM models were running anywhere from several hours to several
days on average. For similar work in the past I had used S-Plus, but I
wanted to stay in Stata if possible. However, the speed of gllamm is
model as well as platform dependent. If I remember correctly, somewhere
I recall Sylvia Rabe-Hesketh describing the relative change in
computation time with different models, perhaps you can find that in the
archives, I can't recall exactly where I saw that. In any case,
different -gllamm- models take quite different amounts of time, even on
the same platform. Of course, all of Stata runs faster on the opteron,
not just -gllamm-, though for me the speed of Stata was only an issue
when I was running -gllamm-.
Related considerations:
I subsequently purchased an Opteron box for almost the sole
purpose of running -gllamm-; it is quite basic in it's configuration to
keep the cost down. I continue, thus far, to do a lot of work on my P4,
the only purpose for the Opteron box is essentially to run GLLAMM. I
bought a single processor Linux box; previous examination of CPU and
memory usage on my P4 WinXP machine suggested that for what I was
running, the bottleneck was the CPU and not memory. One engineer I
spoke with at the computer reseller suggested to me that with two
processors, one could do handle Stata, and the other would handle system
overhead and there would be a resulting speed increase. However, a
second added about another $900 cost to the machine. Since than, I have
monitored my system, and when I have a large GLLAMM model running the
Opteron chip runs at 100% capacity, with ~98% of the load from Stata,
and <1-2% system overhead. That confirms that for my usage, the second
processor would have been an insignificantly negligible increase in
speed. This might not be the case with other patterns of use in other
situations however, and would depend on what else is running
concurrently. Running two instances of Stata on my Opteron does cut the
throughput from each in half, if one were to run two instances regularly
the second processor would perhaps make sense.
If you purchase a Linux box, check with Stata about a
recommended Linux distribution. I had some minor issues with matching
the Linux distribution with the compiled Stata version that temporarily
kept me from running Stata. Both Stata tech support and the computer
reseller (Penguin Computing) were helpful in getting me going, but it
took a week to sort out.
There is a learning curve moving to Linux, but the distribution
I ended up on (SuSE v9) includes a windowed desktop, so Windows and
presumably Apple users can be functional almost immediately if one is
not comfortable operating soley from a terminal window. I am running
XStata Intercooled on my Linux box, and it is essentially not different
from working on my WinXP system. I don't use the Stata GUI much on
either system.
Even with the move to Linux, which I am satisfied was worth it for my
situation, I am guessing that there is opportunity for -gllamm- to run
faster yet, and I continue to hope that this is something that we will
see from Stata Corp at some point. Like you, I agree that -gllamm-
is fantastic; although my hardware move has improved my
productivitymarkedly, in some cases -gllamm- remains quite time-consuming.
Buzz Burhans
At 04:03 PM 4/19/2004 +0200, you wrote:
Dear all
-gllamm- is a fantastic piece of software however, it does not make my
computer look too good .... so we are thinking about investing in new
hardware. I'm wondering if there are any benchmarks available to aid in our
decision.
Which platform is the fastest for analyses with -gllamm- or Stata?
Are there any benchmarks available for -gllamm- (Stata) performance across
platforms?
I'm particularly interested in comparisons with Apples dual G5 computer.
Thanks
Michael
Buzz Burhans
[email protected]
--
______________________________________________________________________
Marcello Pagano
Biostatistics Department Tel: 1-617-432-4911
Harvard School of Public Health Fax: 1-617-739-1781
655 Huntington Avenue email:[email protected]
Boston, MA 02115 http://biosun1.harvard.edu/~bio200
USA
eppur si muove
*
* 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/
