I worked on this pretty extensively around 2011-2013, submitting faster Clojure programs for the problems on the Computer Language Benchmarks Game web site. Several other people, most notably Alex Miller, then spent a chunk of time improving on most of my solutions, and submitting those to that site as well.
In April 2017, the maintainer of that site decided not to include the
Clojure programs and performance measurements any longer. You can see
Clojure programs from an archived copy of the site from 2017-Mar-31
here.
They are not present in the next dated snapshot of the site from
2017-Apr-29
here.
I have copied the latest versions of the Clojure programs from the
archive into the directory
2017-mar-31-benchmarks-game-site-versions
of this repository.
Below are some articles by Laurence Tratt, who has done a lot more careful work than I have in analyzing why measuring run time of benchmarks can vary, and how to at least attempt to eliminate some of those causes of variation. That team found cases where VM performance when running a benchmark was definitely not of the form "some number of initial 'warm up' repetitions are slow, then are consistently faster for all later runs". For example, some got consistently slower, others got repeatedly faster and slow as you continued running it.
- "Minimum Times Tend to Mislead When Benchmarking", August 15 2019
- "Why Aren't More Users More Happy With Our VMs?" Part 1 and Part 2 September 2018
None of the performance data I have collected went to anywhere near that kind of effort to control the variations of results. My best effort was not to run any other significant programs intentionally on the computer where the measurements were performed, and run each benchmark 3 times, many hours apart from each other, with different JVM processes started for each, in hopes of avoiding issues like cron jobs taking significant CPU on one of those runs.
Quick start instructions (see below for a few more details of what these do if you are curious):
(1) Check below for install instructions specific to your OS.
Either: (2a) Install Leiningen (see next paragraph), or (2b) edit
env.sh to specify the location of your Clojure JAR files. Look
for the CLOJURE_CLASSPATH variable settings.
(2a) Leiningen version 2.x installation instructions:
http://github.com/technomancy/leiningen
Save Leiningen 2.x as 'lein' somewhere in your command path.
Then use the following script to install multiple Clojure JAR
files in your $HOME/lein directory.
% ./lein-init.sh
(3) Use init.sh to create largish input and expected output files
(total about 850 Mbytes of storage), by running some Java
programs:
% ./init.sh output
Time required on some of my systems: 6-7 mins, or 18 mins on a
Windows + Cygwin running in a VM.
(4) Run Clojure 1.3 versions of all of the benchmark programs:
% ./run-all.sh clj-1.3
Time required on some of my systems: 25 - 35 mins (about half the
time was spent running the long fannkuch benchmark, and about 1/4
on the long knucleotide benchmmark.)
Another example: Run Java, Clojure 1.3, Clojure 1.4 alpha1, and
Clojure 1.4 alpha3 versions of all of the benchmark programs.
% ./run-all.sh long java clj-1.3 clj-1.4-alpha1 clj-1.4-alpha3
(5) If you want all results recorded to an XML file, look in
env.sh for MP_COMMON_ARGS and the comments before it on how to
enable this. Then do a ./run-all.sh command, or individual
batch.sh commands as described below. You can then convert the
XML file to a CSV file for easier importing into spreadsheets
with:
% bin/xmltable2csv $HOME/results.xml > results.csv
Systems on which this has been tested:
- Mac OS X 10.5.8
- Mac OS X 10.6.8, both with and without MacPorts installed
- Mac OS X 10.7
- Ubuntu 10.4 LTS, 32-bit and 64-bit (also Ubuntu 11.10 and 12.04) Windows XP SP3/Vista/7 + Cygwin
You need a Java Development Kit installed. A Java Virtual Machine with no Java compiler (javac) is not enough. You need a compiler for Java source files that are used in the "./init.sh output" step above to create the input files. You also need a JVM to run Leiningen, if you use that.
Tested with recent Java 1.6.0.X and 1.7.0.X HotSpot JVMs from Sun, and JRockit from Oracle on Windows XP and Vista.
The executable program bin/timemem-darwin must be run as root on Mac OS X 10.6 (Snow Leopard). This is not needed for 10.5 (Leopard). See below for more details if you are curious. One way to do this is to make the program setuid root by running the commands below, entering your password when prompted:
% sudo chown root bin/timemem-darwin
% sudo chmod 4555 bin/timemem-darwinThe following MacPorts packages can be useful. There may be equivalents under Homebrew -- I haven't tried them.
- gmp - If you want to run those pidigits programs that use the GNU MP library.
- git-core - optional, but useful for getting updated versions of clojure-contrib from github.com
- gcc44 - Big, takes a while to install, and only useful if you want to compile some of the parallel C memory bandwidth benchmarking programs in src/stream.
- p5-xml-libxml - Required if you have installed some version of perl within MacPorts (or more likely, it was installed as a dependency required by something else you did choose to install). You can tell if the output of 'which perl' shows /opt/local/bin/perl or similar. Without this package, using a MacPorts-installed Perl will likely result in error messages like the one below when you try to run programs with measureproc (invoked by batch.sh):
Can't locate XML/LibXML.pm in @INC (@INC contains: /opt/local/lib/perl5/site_perl/5.12.3/darwin-multi-2level /opt/local/lib/perl5/site_perl/5.12.3 /opt/local/lib/perl5/vendor_perl/5.12.3/darwin-multi-2level /opt/local/lib/perl5/vendor_perl/5.12.3 /opt/local/lib/perl5/5.12.3/darwin-multi-2level /opt/local/lib/perl5/5.12.3 /opt/local/lib/perl5/site_perl /opt/local/lib/perl5/vendor_perl .) at ../bin/measureproc line 22.
BEGIN failed--compilation aborted at ../bin/measureproc line 22.
You should have a Sun/Oracle/Apple Hotspot JDK installed by default already with java* commands in /usr/bin and some other support files (headers for JNI, etc.) in subdirectories beneath /System/Library/Frameworks/JavaVM.framework
More details on bin/timemem-darwin root permissions:
It needs this permission in order to measure the memory usage of the benchmark programs, for the same reasons that top(1) and ps(1) need to run as root. You can examine the source code of the program in src/timemem-darwin/timemem-darwin.c if you want to know what it is doing. The only thing it needs root permission for is using Mach OS calls to get the current memory usage of a child process. It seems getrusage() and wait4() system calls no longer measure the maximum resident set size of a child process in OS X 10.6, like they used to in 10.5. I've filed a bug on Apple's developer web site in January 2011, but this behavior change may be intentional for all I know.
On Ubuntu 11.10 (and perhaps others), the package libxml-libxml-perl must be installed. See below.
If you want the Sun/Oracle JDK, not OpenJDK (which is the default on Ubuntu as of 10.4, or perhaps even earlier versions of Ubuntu), follow these steps:
% sudo add-apt-repository "deb http://archive.canonical.com/ lucid partner"After that, you can either run the following commands, or use the graphical Synaptic Package Manager to add the sun-java6-jdk package.
% sudo apt-get update
% sudo apt-get install sun-java6-jdkThe following packages are also useful to add on Ubuntu Linux:
- libxml-libxml-perl - Required for measureproc to work.
- libgmp3-dev - If you want to run those pidigits programs that use the GNU MP library.
- gcc - To compile the C pidigits program.
- git-core - optional, but useful for getting updated versions of clojure-contrib from github.com
Source: http://happy-coding.com/install-sun-java6-jdk-on-ubuntu-10-04-lucid/
Cygwin can be obtained from http://www.cygwin.com
Besides the basic packages added with every Cygwin installation, you must install at least some of the following for this software to work:
- perl (category Interpeters and Perl) and libxml2 (many categories, including Devel) - needed if you want to run any of the included Perl scripts, which includes bin/measureproc, the recommended way to make time and memory measurements of the benchmark programs.
- wget (in category Web) - needed if you use Leiningen
- git (category Devel) - optional, but useful for getting updated versions of clojure-contrib from github.com
- gcc - If you want to run the C version of the pidigits program.
- libgmp-devel - If you want to run those pidigits programs that use the GNU MP library.
Edit env.sh so that JAVA_BIN points at the directory where your java and javac commands are installed. Examples are there already for the default install locations of JRockit and Hotspot JVMs (at least particular version numbers of them). If you install the Sun/Oracle Hotspot JDK, its default location is C:\Program Files\Java\jdk. For Oracle JRockit the default install location is C:\Program Files\Java\jrmc-.
I've tried to make all bash scripts handle spaces in filenames by using "${BASH_VARIABLE_NAME}" syntax to substitute the values of shell variables. There is some weirdness in some places in the scripts because Cygwin commands accept paths in Unix notation with / separators, but when invoking the JVM commands, they often require Windows path names with \ separators. The Cygwin command 'cygpath' is useful in converting between these two formats.
Caveat: I don't yet know how to compile the JNI interface to the GMP library so that the Java, Scala, or Clojure programs using that native library can run. The gcc4-core (category Devel) and libgmp-devel (category Libs and Math) Cygwin packages can certainly help in compiling the files, but so far I haven't compiled them in a way that they can be used by a JVM running on Windows. Please let me know if you figure out how to do this.
The Perl, SBCL, and Haskell programs have been well tested on Mac OS X, and more lightly tested on Linux and Windows. The main issue is getting the relatively recent versions of SBCL and GHC, and some benchmark programs require additional libraries on top of that (e.g. for regex matching).
The non-Clojure versions of these benchmark programs have been downloaded from this web site:
http://benchmarksgame.alioth.debian.org
See the file COPYING for the licenses under which these files are distributed.
So far I have Clojure implementations for the benchmarks noted in the table below.
Clojure
Clojure Easy to program is
Benchmark program parallel- parallel-
written? ize? ized?
------------------- ---------- ----------- ----------
binarytrees yes yes yes
chameneos-redux - yes -
------------------- ---------- ----------- ----------
fannkuch yes yes yes
(^^^ deprecated on benchmarks web site)
fannkuchredux yes yes yes
fasta yes no no
------------------- ---------- ----------- ----------
knucleotide yes yes yes
mandelbrot yes yes yes
meteor-contest - ? -
------------------- ---------- ----------- ----------
nbody yes no no
pidigits yes no no
regexdna yes part of it yes
------------------- ---------- ----------- ----------
revcomp yes no no
spectralnorm yes yes yes
thread-ring - yes -
------------------- ---------- ----------- ----------
I have hacked up some shell scripts to automate the compilation and running of some of these programs. They have been tested on Mac OS X 10.5.8 and 10.6.8 with recent versions of the Glasgow Haskell Compiler ghc, SBCL Common Lisp, Perl, several versions of Sun's Java VM (for Windows XP, Linux, and Mac), Oracle's JRockit JVM for Windows, and Clojure 1.2.0 and 1.3.0 alpha1 (and earlier 1.0.0 or shortly before that release). Some of the benchmarks use Clojure transients and thus require 1.0 or later, and some use deftype, requiring Clojure 1.2.0 or later (or whenever deftype was introduced). See below for the steps I used to install SBCL and ghc.
You should edit the file env.sh to point at the location of your Clojure JAR files, and perhaps give explicit path names to some of the other language implementations, if you wish.
If you want to create Clojure JAR files in the locations already specified in the file env.sh, you can install Leiningen and then run the following script. Doing so will create directories named 'lein' and '.m2' in your home directory, if they do not already exist, and fill them with a few dozen files each.
Note: This step is optional, but if you do not do it, you must either edit env.sh to point at your Clojure JAR files, or put Clojure JAR files in the locations mentioned in that file.
% ./lein-init.shThe knucleotide, regexdna, and revcomp benchmark input files are quite large (250 Mbytes), and not included in the distribution. These input files must be generated by running this shell script:
% ./init.shThat will generate the input files only. You can also choose to generate those plus the "expected output files" (which are just the output of the Java versions of the benchmark programs) by running this:
% ./init.sh outputIf you have the input files generated, you can run all of the implementations of the knucleotide benchmark, with the quick, medium, and long test input sizes, using these commands:
% cd knucleotide
% ./batch.shYou can also pick and choose which benchmark lengths to run, and which language implementations to run, by giving options like these on the command line:
% ./batch.sh java clj-1.2 clj-1.3-alpha1 quick mediumThat would run the Java, Clojure 1.2, and Clojure 1.3 alpha1 versions of the benchmark, each with the quick and medium size input files. Note that the order of the command line arguments is not important. The following command would run only the Clojure 1.2 version with the long input file:
% ./batch.sh clj-1.2 longYou can also run the following command from the root directory of this package, and it will run a batch.sh command with the same command line arguments in each of the benchmark subdirectories, e.g.
./run-all.sh long java clj-1.2will run the long benchmark for Java and Clojure 1.2 in all of the subdirectories mentioned in the 'for' line you can see for yourself in the run-all.sh script.
Note that all of the benchmarks game web site results are for what are called the 'long' benchmarks in this package. The short and medium tests are primarily for quicker cycling through the edit-compile-run-debug loop, when you are developing a benchmark program yourself.
If you find any improvements to the Clojure versions, I'd love to hear about them.
The files RESULTS-clj-1.1 and RESULTS-clj-1.2 in the results directory contains some summarized execution times from running these programs on my home iMac. The file results-java-clj-1.2-1.3a1.xls is an Excel spreadsheet containing run time results for several different JVMs and operating systems, all on the same hardware as each other (but not the same hardware as the RESULTS-clj-1.1 and RESULTS-clj-1.2 files above, so don't go comparing results between these files to each other directly).
Andy Fingerhut [email protected]
On a Mac OS X machine (tested on 10.5.8 and 10.6.4 at least), download and install MacPorts from here:
After following the instructions there for installing MacPorts, you can install the Glasgow Haskell compiler with the command:
% sudo port install ghcAnd SBCL with the threads (i.e. multi-threading) option enabled with this command:
% sudo port install sbcl@+threadsHere are the versions I currently have installed used to produce some of my benchmark results:
% port installed ghc sbcl
The following ports are currently installed:
ghc @6.10.1_8+darwin_9_i386 (active)
sbcl @1.0.24_0+darwin_9_i386+html+test+threads (active)
% java -version
java version "1.6.0_13"
Java(TM) SE Runtime Environment (build 1.6.0_13-b03-211)
Java HotSpot(TM) 64-Bit Server VM (build 11.3-b02-83, mixed mode)
% javac -version
javac 1.6.0_13
% sbcl --version
SBCL 1.0.29
% ghc --version
The Glorious Glasgow Haskell Compilation System, version 6.10.1I've also done some testing of this set of scripts on an Ubuntu 10.04 Desktop i386 installation, in a VMWare Fusion virtual machine running on my Mac (and earlier tested on Ubuntu 9.04).
I've tried it with these versions of packages installed using Ubuntu's Synaptic Package Manager.
sun-java6-jdk 6-14-0ubuntu1.9.04
sbcl 1.0.42 (and earlier tested with 1.0.18.0-2)
ghc 6.8.2dfsg1-1ubuntu1
% java -version
java version "1.6.0_14"
Java(TM) SE Runtime Environment (build 1.6.0_14-b08)
Java HotSpot(TM) Client VM (build 14.0-b16, mixed mode, sharing)
% javac -version
javac 1.6.0_14
% sbcl --version
SBCL 1.0.18.debian
% ghc --version
The Glorious Glasgow Haskell Compilation System, version 6.8.2Apparently the Haskell knucleotide program requires GHC 6.10 or later, and gives a compilation error with GHC 6.8.x, as is currently the latest available through Ubuntu's distribution system. If you really want to run that benchmark, you could try installing GHC 6.10 yourself. This web page may provide the right recipe for doing so. I haven't tried it myself.
http://www.johnmacfarlane.net/Gitit%20on%20Ubuntu
... and whether these Clojure programs use those features
The following notes come from a reply to a question on Stack Overflow about why the Clojure programs are slower than the Scala programs.
http://stackoverflow.com/questions/4148382/why-does-clojure-do-worse-than-scala-in-the-alioth-benchmarks
Note that significant speed improvements were made to many of the Clojure programs on the Computer Language Benchmarks Game web site after this Stack Overflow discussion occurred, so what the discussion participants saw then was noticeably slower (relative to Java -6 server) than what is on the web site now.
For example, the following features in Clojure are all very cool and useful for development convenience, but incur some runtime performance overhead:
- Lazy sequences and lists
- Dynamic Java interoperability using reflection
- Runtime function composition / first class functions
- Multimethods / dynamic dispatch
- Dynamic compilation with eval or on the REPL
- BigInteger arithmetic
If you want absolute maximum performance (at the cost of some extra complexity), you would want to rewrite code to avoid these and use things like:
- Static type hinting (to avoid reflection)
- Transients
- Macros (for compile time code manipulation)
- Protocols
- Java primitives and arrays
- loop / recur for iteration
I have attempted to categorize which Clojure programs in this collection use the features above, and which do not, in an Excel spreadsheet. See the file results/clojure-features-used.xls. This spreadsheet should be easily readable using Open Office or one of its derivatives, as no "fancy features" are used in it. It is simply a more convenient way for me to record this information in table form, and later rearrange and view them in different ways, than a text file.
Cliff Click's slides for his talk "How NOT To Write A Microbenchmark", given at JavaOne 2002 (Sun's 2002 Worldwide Java Developer Conference).
http://www.azulsystems.com/events/javaone_2002/microbenchmarks.pdf
I mention this because (1) it is worth reading for anyone interested in benchmarks, to avoid pitfalls, and (2) so I can make a list of pitfalls mentioned in there, and for each Clojure program describe whether I believe it has fallen into that pitfall.
Are the benchmarks game programs microbenchmarks? Dr. Click gives his definition on p. 8, and several examples throughout his talk. The benchmarks game programs are relatively small, and in some cases the datasets are "large" (i.e. hundreds of megabytes, but none are a gigabyte). However, there is a significant difference between these benchmarks vs. the examples in Dr. Click's slides: the goal with these benchmarks is to measure the entire run time of the whole program, not some smaller part of it. We aren't trying to see whether a particular function call takes 7 microsec versus 15 microsec. We are trying to find out how much time is required for all of the calls made together during the entire run, and find a program that is as fast as possible for that total time.
Now it is true that for most of these programs, 99% of the time is spent in relatively small fraction of the lines of code. For example, in knucleotide.clj-8b.clj, almost all of the compute time is spent in the last 13 lines of the function tally-dna-subs-with-len. In nbody.clj-12.clj, most of the time is spent in the last 10 lines of the function advance, which calls the 5-line p-dt! and the 18-line v-dt!, which in turn calls the 4-line v+!, for a total of 27 lines of "hot code" in a 252-line file.
One piece of advice from Dr. Click's slides that is most relevant is "Know what you test" (title of p. 32). These benchmarks game programs are not intended to discover some particular fact, and remove all other variables, unless your goal is to discover how well this particular program performs on the task at hand. For example, the knucleotide benchmark programs do not measure only the performance of the hash table implementation (because there is at least some work in doing other things, like reading in a file that is hundreds of megabytes long), but that is certainly the most significant portion of the program's run time.
In general, when comparing Java and Clojure run times, or any two languages implemented on the Java Virtual Machine in this collection of benchmarks, realize that the following run times are included in the times that you see reported:
- JVM initialization time
- loading classes from compiled class files
- time spent reading the input data file (if any) and writing output
- any time spent doing JIT compilation
- any time spent performing garbage collection
The following times are not included in the times you see reported:
- Compiling source files to class files
All of the Clojure programs are ahead of time (or AOT) compiled, and compiled class files are saved on disk, before the measurement clock is started.
Is this fair? In the sense that the same kinds of processing time are included in measurements of both Java and Clojure, it is fair. Is it what you personally want to measure? That depends upon what you want to measure. You are welcome to measure other things about these programs if you wish, and report the times you get.
Because the goal here is to measure the entire run time of these programs, "warming up the JVM" via JIT compilation is all included. There is no reason (with this goal) to first run the JVM until it is warmed up, and then run "the real test". Warmup is part of the real test.
In all of these programs, there should be little or no dead code to be eliminated. All results calculated are printed out in some form. Sometimes only a small summary of the the calculated results are actually printed, but it is always intended to be a summary such that there are no known compilers that are "smart" enough to figure out that they could reduce the processing required in order to print the desird results. Could a human do that? In some cases, it is pretty easy for a human to change the program to produce the same output in a similar way. For example, in the knucleotide benchmark the program must print out the number of occurrences of the substring "GGT" in a large input string. However, the instructions for the benchmark require that first a table is made that maps all length 3 substrings of the input string to a count of the number of occurrences, then extract out the entry for "GGT" and print its count. Obviously a person can change the program to reduce the time and space required if the goal is only to calculate counts for the substring "GGT", but it is breaking the rules of the benchmark to do so.
Cliff Click's advice (summarized greatly):
-
Is there loop unrolling going on behind the scenes? (pp. 17-18)
-
"Beware 'hidden' Cache Blowout" - Does the data set fit into L1 cache? (p. 28-29)
-
Explicitly Handle GC: Either don't allocate in main loops, so no GC pauses, or run long enough to reach GC steady state. (p. 30) TBD: Consider adding instrumentation to measure total GC time in each run involving a JVM.
-
Know what you test (p. 32) - use profiler to measure where most of the time is spent, and report that with the results.
-
Be aware of system load - How to verify total CPU time taken by all other processes on the system? I can eyeball it if I watch the Activity Monitor on a Mac, or top on Linux, Windows Task Manager on Windows, but is there an automated way to easily determine the total CPU time used by all other processes during a particular time interval?
-
Be aware of clock granularity (p. 33): This should not be an issue since all of the Clojure programs run for at least 10 sec, and clock granularity is definitely better than 0.1 sec. (TBD: Verify these numbers)
-
Run to steady state, at least 10 sec (p. 33): In many of these programs, each iteration is doing something a bit different, but the general idea of running at least 10 sec is a good one.
-
Fixed size datasets (p. 34): Every run with input data has the same input file. For each one, state the size of the input data that actually needs to be "worked on".
-
Constant amount of work per iteration (p. 34): This is not really relevant for these benchmarks, if we are only paying attention to the total time to complete the benchmark run. It is only relevant if we are trying to calculate a time per iteration of some loop, which we are not.
-
Avoid the 'eqntott Syndrome' (p. 34): Again, we are not looking at run times of separate iterations and comparing them to each other. There is some warmup time of doing JIT compilation in all of these runs, but it is there every single time, and it is there for Clojure and Java, and any other JVM-based language implementation.
-
Avoid 'dead' loops (p. 35): The final answer is always printed, and the computation is non-trivial in all cases I can recall. Good to verify this for each benchmark program individually.
-
Be explicit about GC (p. 36): Does this mean to measure it and report it separately?
-
JIT performance may change over time. Warmup loop + test code before ANY timing (p. 36): This is relevant if we want to measure the time per iteration of some loop. That is not the goal for these benchmark programs. We only care about the time to get the final result.
TBD: Record a summary for the fastest known Clojure programs, or maybe all of them, where they spend the most time, i.e. the call stacks that (combined) take up at least half of the program execution time, according to a Java profiler.