FTOOLS: A faster Stata for large datasets
Some of the most common Stata commands (collapse, merge, sort, etc.) are not designed for large datasets. This package provides alternative implementations that solves this problem, speeding up these commands by 3x-10x:
Other user commands that are very useful for speeding up Stata with large datasets include:
gtools, a package similar to
ftoolsbut written in C. In most cases it's much faster than both ftools and the standard Stata commands, as shown in the graph above. Try it out!
sumupprovides fast summary statistics, and includes the
fasttabstatcommand, a faster version of
egenmiscintroduces the egen functions
fastwpctile, etc. that provide much faster alternatives to
pctile. Also see the
fastxtilepackage, which provides similar functionality.
randomtagis a much faster alternative to
reghdfeprovides a faster alternative to
areg, as well as multi-way clustering and IV regression.
parallelallows for easier parallel computing in Stata (useful when running simulations, reshaping, etc.)
boottest, for efficiently running wild bootstraps.
rangestatcommands are useful for running commands and collecting statistics on rolling windows of observations.
ftools can also be used to speed up your own commands. For more information, see this presentation from the 2017 Stata Conference (slides 14 and 15 show how to create faster alternatives to
xmiss with only a couple lines of code). Also, see
help ftools for the detailed documentation.
ftools is two things:
- A list of Stata commands optimized for large datasets, replacing commands such as: collapse, contract, merge, egen, sort, levelsof, etc.
- A Mata class (Factor) that focuses on working with categorical variables. This class is what makes the above commands fast, and is also what powers
Currently the following commands are implemented:
contractand most of
join(and its wrapper
sort(although it is rarely faster than sort)
* Stata usage: sysuse auto fsort turn fegen id = group(turn trunk) fcollapse (sum) price (mean) gear, by(turn foreign) freq * Advanced: creating the .mlib library: ftools, compile * Mata usage: sysuse auto, clear mata: F = factor("turn") mata: F.keys, F.counts mata: sorted_price = F.sort(st_data(., "price"))
Other features include:
- Add your own functions to -fcollapse-
- View the levels of each variable with
- Embed -factor()- into your own Mata program. For this, you can
F.sort()and the built-in
(see the test folder for the details of the tests and benchmarks)
Given a dataset with 20 million obs. and 5 variables, we create the following variable, and create IDs based on that:
gen long x = ceil(uniform()*5000)
Then, we compare five different variants of egen group:
|egen id = group(x)||49.17||51.26|
|fegen id = group(x)||1.44||1.53|
|fegen id = group(x), method(hash0)||1.41||1.60|
|fegen id = group(x), method(hash1)||8.87||9.35|
|fegen id = group(x), method(stata)||34.73||35.43|
Our variant takes roughly 3% of the time of egen group.
If we were to choose a more complex hash method, it would take 18% of the time.
We also report the most efficient method based in Stata (that uses
which is still significantly slower than our Mata approach.
- The gap is larger in systems with two or less cores, and smaller in systems with many cores (because our approach does not take much advantage of multicore)
- The gap is larger in datasets with more observations or variables.
- The gap is larger with fewer levels
On a dataset of similar size, we ran
collapse (sum) y1-y15, by(x3) where
x3 takes 100 different values:
|Method||Time||% of Collapse|
|collapse … , fast||81.87||100%|
|fcollapse … , fast||38.54||47%|
|fcollapse … , fast pool(5)||28.32||35%|
We can see that
fcollapse takes roughly a third of the time of
(although it uses more memory when moving data from Stata to Mata).
As a comparison, tabulating the data (one of the most efficient Stata operations) takes 11% of the time of
pool(#) option will use very little memory (similar to
collapse) at also very good speeds.
- The gap is larger if you want to collapse fewer variables
- The gap is larger if you want to collapse to fewer levels
- The gap is larger for more complex stats. (such as median)
compressing the by() identifiers beforehand might lead to significant improvements in speed (by allowing the use of the internal hash0 function instead of hash1).
- In a computer with less memory, it seems
pool(#)might actually be faster.
collapse: alternative benchmark
We can run a more complex query, collapsing means and medians instead of sums, also with 20mm obs.:
|Method||Time||% of Collapse|
|collapse … , fast||81.06||100%|
|fcollapse … , fast||30.93||38%|
|fcollapse … , fast pool(5)||33.85||42%|
sumup might be better for medium-sized datasets, although some benchmarking is needed)
And we can see that the results are similar.
join (and fmerge)
merge but avoids sorting the datasets. It is faster than
for datasets larger than ~ 100,000 obs., and for datasets above 1mm obs. it
takes a third of the time.
|Method||Time||% of merge|
isid, but allowing for
if in and on the other hand not allowing for
In very large datasets, it takes roughly a third of the time of
Provides the same results as
In large datasets, takes up to 20% of the time of
At this stage, you would need a significantly large dataset (50 million+) for
fsort to be faster than
|Method||Avg. 1||Avg. 2|
|sort id, stable||63.74||65.72|
The table above shows the benchmark
on a 50 million obs. dataset.
The unstable sorting is slightly slower (col. 1) or slighlty faster (col. 2)
fsort approach. On the other hand, a stable sort is clearly
fsort (which always produces a stable sort)
Within Stata, type:
cap ado uninstall ftools ssc install ftools
With Stata 13+, type:
cap ado uninstall ftools net install ftools, from(https://github.com/sergiocorreia/ftools/raw/master/src/)
For older versions, first download and extract the zip file, and then run
cap ado uninstall ftools net install ftools, from(SOME_FOLDER)
Where SOME_FOLDER is the folder that contains the stata.toc and related files.
Compiling the mata library
In case of a Mata error, try typing
ftools to create the Mata library (lftools.mlib).
Installing local versions
To install from a git fork, type something like:
cap ado uninstall ftools net install ftools, from("C:/git/ftools/src") ftools, compile
(Changing "C:/git/" to your own folder)
fcollapse function requires the
moremata package for some the median and percentile stats:
ssc install moremata
Users of Stata 11 and 12 need to install the
ssc install boottest
"What features is this missing?"
- You can create levels based on one or more variables, and on numeric or string variables, but not on combinations of both. Thus, you can't do something like
fcollapse price, by(make foreign)because make is string and foreign is numeric. This is due to a limitation in Mata and is probably a hard restriction. As a workaround, just run something like
fegen id = group(make), to create a numeric ID.
- Support for weights is incomplete (datasets that use weights are often relatively small, so this feature has less priority)
- Some commands could also gain large speedups (merge, reshape, etc.)
- Since Mata is ~4 times slower than C, rewriting this in a C plugin should lead to a large speedup.
"How can this be faster than existing commands?"
Existing commands (e.g. sort) are often compiled and don't have to move data from Stata to Mata and viceversa. However, they use inefficient algorithms, so for datasets large enough, they are slower. In particular, creating identifiers can be an ~O(N) operation if we use hashes instead of sorting the data (see the help file). Similarly, once the identifiers are created, sorting other variables by these identifiers can be done as an O(N) operation instead of O(N log N).
"But I already tried to use Mata's
asarray and it was much slower"
asarray() has a key problem: it is very slow with hash collisions (which you see a lot in this use case). Thus, I avoid using
asarray() and instead use
hash1() to create a hash table with open addressing (see a comparision between both approaches here).