Statalist The Stata Listserver

[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: Stata/MP for parallel processing

From (Vince Wiggins, StataCorp)
Subject   st: Stata/MP for parallel processing
Date   Thu, 30 Mar 2006 16:31:32 -0600

We have just put up the web page announcing Stata/MP:

Stata/MP is the newest flavor of Stata, and it sits just above Stata/SE.
Basically, it's Stata/SE on steroids.  It doesn't do more, it does it faster.

Stata/MP ships April 14th.

The MP part stands for Multi Processing, and Stata/MP runs on dual-core and
multiprocessor computers, from 2 to 32.  Some of you who have attended
meetings have heard Bill Gould talk about "Parallel Stata", which was its
development name.

Anyway, Stata/MP is faster.  On 2-processor computers, Stata runs about 1.4
times faster, or if you prefer, executes in about 1/1.4 = 71% of the time.  On
4 processors, Stata runs twice as faster or in half the time.

That's the median across all commands, and the range (2 processors) is 1 to 2
(same time to half of the time) and (4 processors) 1 to 4 (same to a quarter
of the time).

Estimation commands speed up more, on median, as do most commands that take a
long time to run.  In summary

                         min         median         max
     2 processors
       all commands        1          1.4           2.0
       est. commands       1          1.7           2.0

     4 processors
       all commands        1          2.0           4.0
       est. commands       1          2.8           4.0

We are pretty pleased with these results.

Some commands do not speed up at all, either because they are inherently 
sequential (time series) or because no effort was made parallelizing 
them (graphics, xtmixed, stcox), and some commands speed up perfectly
(regress, most parametric survival models, etc.)

For a complete evaluation of Stata/MP's performance, command by command, and
detailed discussion of theory and outcome, see the white paper at

There's a section of the report on GLLAMM

                         min         median         max
     2 processors
       GLLAMM              1          1.7           2.0

     4 processors
       GLLAMM              1          2.3           3.4

Take those GLLAMM figures with a grain of salt, however.  GLLAMM can fit so
many models that we need to run more tests to have reasonable statistics, 
and then present those statistics according to model type.

We are looking forward to hearing feedback from Stata users.

-- The Stata/MP development team (in no particular order)
        Hua Peng      (
        Alan Riley    (
        David Drukker (
        Vince Wiggins (
        Jeff Pitblado (
	Chinh Nguyen  (
	Kevin Turner  (
	Pete Huckelba (

*   For searches and help try:

© Copyright 1996–2020 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index