Parallel lets you run Stata faster, sometimes faster than MP itself. By organizing your job in several Stata instances, parallel allows you to work with out-of-the-box parallel computing. Using the the parallel prefix, you can get faster simulations, bootstrapping, reshaping big data, etc. without having to know a thing about parallel computing. With no need of having Stata/MP installed on your computer, parallel has showed to dramatically speedup computations up to two, four, or more times depending on how many processors your computer has.
See the HTML version of the program help file for more info. Other resources include the Stata 2017 conference presentation and the SSC page at Boston College (though the SSC version is a bit out-of-date, see below).
When using parallel, please include the following:
Vega Yon GG, Quistorff B. parallel: A command for parallel computing. The Stata Journal. 2019;19(3):667-684. doi:10.1177/1536867X19874242
Or use the following bibtex entry:
@article{
VegaYon2019,
author = {George G. {Vega Yon} and Brian Quistorff},
title ={parallel: A command for parallel computing},
journal = {The Stata Journal},
volume = {19},
number = {3},
pages = {667-684},
year = {2019},
doi = {10.1177/1536867X19874242},
URL = {https://doi.org/10.1177/1536867X19874242},
eprint = {https://doi.org/10.1177/1536867X19874242}
}If you have a previous installation of parallel installed from a different source (SSC, specific folder, specific URL) you should uninstall that first (ado uninstall parallel). Once installed it is suggested to restart Stata.
Stata version >=13 can install the latest stable version using a GitHub URL,
net install parallel, from(https://raw.github.com/gvegayon/parallel/stable/) replace
mata mata mlib indexFor Stata version <13, download as zip, unzip, and then replace the above URL with the full local path to the files.
Latest version (master branch): Use the URL https://raw.github.com/gvegayon/parallel/master/. To get a zip click the "Clone or download" button and choose zip.
Older releases: Go to the Releases Page. You can use the release tag to in the URL (e.g., https://raw.github.com/gvegayon/parallel/v1.15.8.19/). See also the associated zip download option.
An older version is available from the SSC. It does not have all the bug fixes so it is not recommended to install it. If it is required, however, use
ssc install parallel, replace
mata mata mlib indexThe following minimal examples have been written to introduce how to use the module. Please notice that the only examples actually designed to show potential speed gains are parfor and bootstrap.
The examples have been executed on a Dell Vostro 3300 notebook running Ubuntu 14.04 with an Intel Core i5 CPU M 560 (2 physical cores) with 8Gb of RAM, using Stata/IC 12.1 for Unix (Linux 64-bit x86-64).
For more examples and details please refer to the module's help file or the wiki Gallery page.
When conducted over groups, parallelizing egen can be useful. In the following example we show how to use parallel with by: egen.
. parallel initialize 2, f
N Clusters: 2
Stata dir: /usr/local/stata13/stata
. sysuse auto
(1978 Automobile Data)
. parallel, by(foreign): egen maxp = max(price)
-------------------------------------------------------------------------------
Parallel Computing with Stata
Clusters : 2
pll_id : m61jt2abc1
Running at : /home/vegayon/Dropbox/repos/parallel
Randtype : datetime
Waiting for the clusters to finish...
cluster 0001 has exited without error...
cluster 0002 has exited without error...
-------------------------------------------------------------------------------
Enter -parallel printlog #- to checkout logfiles.
-------------------------------------------------------------------------------
. tab maxp
maxp | Freq. Percent Cum.
------------+-----------------------------------
12990 | 22 29.73 29.73
15906 | 52 70.27 100.00
------------+-----------------------------------
Total | 74 100.00
Which is the ``parallel'' way to do:
. sysuse auto
(1978 Automobile Data)
. bysort foreign: egen maxp = max(price)
. tab maxp
maxp | Freq. Percent Cum.
------------+-----------------------------------
12990 | 22 29.73 29.73
15906 | 52 70.27 100.00
------------+-----------------------------------
Total | 74 100.00
In this example we'll evaluate a regression model using bootstrapping which, together with simulations, is one of the best ways to use parallel
. sysuse auto, clear
(1978 Automobile Data)
. parallel initialize 4, f
N Clusters: 4
Stata dir: /usr/local/stata13/stata
. timer on 1
. parallel bs, reps(5000): reg price c.weig##c.weigh foreign rep
-------------------------------------------------------------------------------
Parallel Computing with Stata
Clusters : 4
pll_id : m61jt2abc1
Running at : /home/vegayon/Dropbox/repos/parallel
Randtype : datetime
Waiting for the clusters to finish...
cluster 0001 has exited without error...
cluster 0002 has exited without error...
cluster 0003 has exited without error...
cluster 0004 has exited without error...
-------------------------------------------------------------------------------
Enter -parallel printlog #- to checkout logfiles.
-------------------------------------------------------------------------------
parallel bootstrapping Number of obs = 69
Replications = 5000
command: regress price c.weig##c.weigh foreign rep
------------------------------------------------------------------------------
| Observed Bootstrap Normal-based
| Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
weight | -4.317581 3.033419 -1.42 0.155 -10.26297 1.627811
|
c.weight#|
c.weight | .0012192 .0004827 2.53 0.012 .0002732 .0021653
|
foreign | 3155.969 890.4385 3.54 0.000 1410.742 4901.197
rep78 | -30.11387 327.7725 -0.09 0.927 -672.5361 612.3084
_cons | 6415.187 5047.099 1.27 0.204 -3476.945 16307.32
------------------------------------------------------------------------------
. timer off 1
. timer list
1: 10.59 / 1 = 10.5930
97: 0.07 / 2 = 0.0340
98: 0.00 / 1 = 0.0030
99: 10.52 / 1 = 10.5190
Which is the ``parallel way'' to do:
. sysuse auto, clear
(1978 Automobile Data)
. timer on 2
. bs, reps(5000) nodots: reg price c.weig##c.weigh foreign rep
Linear regression Number of obs = 69
Replications = 5000
Wald chi2(4) = 51.13
Prob > chi2 = 0.0000
R-squared = 0.5622
Adj R-squared = 0.5348
Root MSE = 1986.4039
------------------------------------------------------------------------------
| Observed Bootstrap Normal-based
price | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
weight | -4.317581 3.110807 -1.39 0.165 -10.41465 1.779489
|
c.weight#|
c.weight | .0012192 .0004951 2.46 0.014 .0002489 .0021896
|
foreign | 3155.969 863.9629 3.65 0.000 1462.633 4849.305
rep78 | -30.11387 323.6419 -0.09 0.926 -664.4404 604.2127
_cons | 6415.187 5162.58 1.24 0.214 -3703.285 16533.66
------------------------------------------------------------------------------
. timer off 2
. timer list
2: 17.78 / 1 = 17.7810
From the simulate stata command:
. parallel initialize 2, f
N Clusters: 2
Stata dir: /usr/local/stata13/stata
. program define lnsim, rclass
1. version 12.1
2. syntax [, obs(integer 1) mu(real 0) sigma(real 1) ]
3. drop _all
4. set obs `obs'
5. tempvar z
6. gen `z' = exp(rnormal(`mu',`sigma'))
7. summarize `z'
8. return scalar mean = r(mean)
9. return scalar Var = r(Var)
10. end
. parallel sim, expr(mean=r(mean) var=r(Var)) reps(10000): lnsim, obs(100)
Warning: No data loaded.
-------------------------------------------------------------------------------
> -
Exporting the following program(s): lnsim
lnsim, rclass:
1. version 12.1
2. syntax [, obs(integer 1) mu(real 0) sigma(real 1) ]
3. drop _all
4. set obs `obs'
5. tempvar z
6. gen `z' = exp(rnormal(`mu',`sigma'))
7. summarize `z'
8. return scalar mean = r(mean)
9. return scalar Var = r(Var)
-------------------------------------------------------------------------------
> -
-------------------------------------------------------------------------------
Parallel Computing with Stata
Clusters : 2
pll_id : 93mwp9vps1
Running at : /home/vegayon/Dropbox/repos/parallel
Randtype : datetime
Waiting for the clusters to finish...
cluster 0001 has exited without error...
cluster 0002 has exited without error...
-------------------------------------------------------------------------------
Enter -parallel printlog #- to checkout logfiles.
-------------------------------------------------------------------------------
.
. summ
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
mean | 10000 1.648843 .2165041 1.021977 2.907977
var | 10000 4.650656 4.218584 .6159253 133.9232
which is the parallel way to do
. program define lnsim, rclass
1. version 12.1
2. syntax [, obs(integer 1) mu(real 0) sigma(real 1) ]
3. drop _all
4. set obs `obs'
5. tempvar z
6. gen `z' = exp(rnormal(`mu',`sigma'))
7. summarize `z'
8. return scalar mean = r(mean)
9. return scalar Var = r(Var)
10. end
. simulate mean=r(mean) var=r(Var), reps(10000) nodots: lnsim, obs(100)
command: lnsim, obs(100)
mean: r(mean)
var: r(V.
. summ
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
mean | 10000 1.644006 .2133008 1.061809 2.991108
var | 10000 4.568202 3.984818 .6348574 110.893
In this example we create a short program (parfor) which is intended to work as a parfor program, this is, looping through 1/N in a parallel fashion
. // Cleaning working space
. clear all
. timer clear
.
. // Set up
. set seed 123
. local n = 5e6
. set obs `n'
obs was 0, now 5000000
. gen x = runiform()
. gen y_pll = .
(5000000 missing values generated)
. clonevar y_ser = y_pll
(5000000 missing values generated)
.
. // Loop replacement function
. prog def parfor
1. args var
2. forval i=1/`=_N' {
3. qui replace `var' = sqrt(x) in `i'
4. }
5. end
.
. // Running the algorithm in parallel fashion
. timer on 1
. parallel initialize 4, f
N Clusters: 4
Stata dir: /usr/local/stata13/stata
. parallel, prog(parfor): parfor y_pll
-------------------------------------------------------------------------------
> -
Exporting the following program(s): parfor
parfor:
1. args var
2. forval i=1/`=_N' {
3. qui replace `var' = sqrt(x) in `i'
4. }
-------------------------------------------------------------------------------
> -
-------------------------------------------------------------------------------
Parallel Computing with Stata
Clusters : 4
pll_id : wrusvgqb91
Running at : /home/vegayon/Dropbox/repos/parallel
Randtype : datetime
Waiting for the clusters to finish...
cluster 0001 has exited without error...
cluster 0002 has exited without error...
cluster 0003 has exited without error...
cluster 0004 has exited without error...
-------------------------------------------------------------------------------
Enter -parallel printlog #- to checkout logfiles.
-------------------------------------------------------------------------------
. timer off 1
.
. // Running the algorithm in a serial way
. timer on 2
. parfor y_ser
. timer off 2
.
. // Is there any difference?
. list in 1/10
+--------------------------------+
| x y_pll y_ser |
|--------------------------------|
1. | .912044 .9550099 .9550099 |
2. | .0075452 .0868631 .0868631 |
3. | .2808588 .5299612 .5299612 |
4. | .4602787 .6784384 .6784384 |
5. | .5601059 .7484022 .7484022 |
|--------------------------------|
6. | .6731906 .820482 .820482 |
7. | .6177611 .7859778 .7859778 |
8. | .8656877 .9304234 .9304234 |
9. | 9.57e-06 .0030943 .0030943 |
10. | .4090917 .6396028 .6396028 |
+--------------------------------+
. gen diff = y_pll != y_ser
. tab diff
diff | Freq. Percent Cum.
------------+-----------------------------------
0 | 5,000,000 100.00 100.00
------------+-----------------------------------
Total | 5,000,000 100.00
.
. // Comparing time
. timer list
1: 8.93 / 1 = 8.9260
2: 16.06 / 1 = 16.0580
97: 0.42 / 1 = 0.4240
98: 0.32 / 1 = 0.3150
99: 8.17 / 1 = 8.1740
. di "Parallel is `=round(r(t2)/r(t1),.1)' times faster"
Parallel is 1.8 times faster
.
If you need to use parallel on an older version of Stata than what we build here, you can build and install the package locally.
You will need to install Stata devtools to build the package and log2html to build the html version of the help.
Then you can go to ado/ and either do compile.do or do compile_and_install.do depending on whether you want to just build the package (.mlib) or also install. There are also several build build checks in the makefile that can easily be run from Linux.
The Windows plugins can be built using Visual Studio Community Edition (tested on 2019), which is freee, with the C++ build tools and Windows SDK.
George G. Vega [aut,cre] g.vegayon %at% gmail
Brian Quistorff [aut] brian-work %at% quistorff . com