parallel - build and execute shell command lines from standard input in parallel
parallel [options] [command [arguments]] < list_of_arguments
parallel [options] [command [arguments]] ::: arguments
parallel [options] [command [arguments]] :::: argfile(s)
parallel --semaphore [options] command
GNU parallel is a shell tool for executing jobs in parallel locally or using remote computers. A job is typically a single command or a small script that has to be run for each of the lines in the input. The typical input is a list of files, a list of hosts, a list of users, a list of URLs, or a list of tables.
If you use xargs today you will find GNU parallel very easy to use as GNU parallel is written to have the same options as xargs. If you write loops in shell, you will find GNU parallel may be able to replace most of the loops and make them run faster by running several jobs in parallel. If you use ppss or pexec you will find GNU parallel will often make the command easier to read.
GNU parallel makes sure output from the commands is the same output as you would get had you run the commands sequentially. This makes it possible to use output from GNU parallel as input for other programs.
For each line of input GNU parallel will execute command with the line as arguments. If no command is given, the line of input is executed. Several lines will be run in parallel. GNU parallel can often be used as a substitute for xargs or cat | bash.
Before looking at the options you may want to check out the examples after the list of options. That will give you an idea of what GNU parallel is capable of.
You can also watch the intro video for a quick introduction: http://www.youtube.com/watch?v=OpaiGYxkSuQ or at http://tinyogg.com/watch/TORaR/ and http://tinyogg.com/watch/hfxKj/
Command to execute. If command or the following arguments contain {} every instance will be substituted with the input line. Setting a command also invokes --file.
If command is given, GNU parallel will behave similar to xargs. If command is not given GNU parallel will behave similar to cat | sh.
Input line. This is the default replacement string and will normally be used for putting the argument in the command line. It can be changed with -I.
Input line without extension. This is a specialized replacement string with the extension removed. If the input line contains . after the last / the last . till the end of the string will be removed and {.} will be replaced with the remaining. E.g. foo.jpg becomes foo, subdir/foo.jpg becomes subdir/foo, sub.dir/foo.jpg becomes sub.dir/foo, sub.dir/bar remains sub.dir/bar. If the input line does not contain . it will remain unchanged.
{.} can be used the same places as {}. The replacement string {.} can be changed with -U.
Argument from argument file n or the n'th argument. See -a and -N.
{n} can be used the same places as {}.
Argument from argument file n or the n'th argument without extension. Similar to {.}. See -a and -N.
{n.} can be used the same places as {.}.
Use arguments from the command line as input instead of from stdin (standard input). Unlike other options for GNU parallel ::: is placed after the command and before the arguments.
The following are equivalent:
(echo file1; echo file2) | parallel gzip parallel gzip ::: file1 file2 parallel gzip {} ::: file1 file2 parallel --arg-sep ,, gzip {} ,, file1 file2 parallel --arg-sep ,, gzip ,, file1 file2 parallel ::: "gzip file1" "gzip file2"
To avoid treating ::: as special use --arg-sep to set the argument separator to something else. See also --arg-sep.
stdin (standard input) will be passed to the first process run.
If --arg-file is set arguments from that file will be appended.
Another way to write -a argfile1 -a argfile2 ...
See -a.
Use NUL as delimiter. Normally input lines will end in \n (newline). If they end in \0 (NUL), then use this option. It is useful for processing arguments that may contain \n (newline).
Read items from the file input-file instead of stdin (standard input). If you use this option, stdin is given to the first process run. Otherwise, stdin is redirected from /dev/null.
If multiple -a are given, one line will be read from each of the files. The arguments can be accessed in the command as {1} .. {n}, so {1} will be a line from the first file, and {6} will refer to the line with the same line number from the 6th file.
Use sep-str instead of :::: as separator string between command and argument files. Useful if :::: is used for something else by the command.
See also: ::::.
Use sep-str instead of ::: as separator string. Useful if ::: is used for something else by the command.
Also useful if you command uses ::: but you still want to read arguments from stdin (standard input): Simply change --arg-sep to a string that is not in the command line.
See also: :::.
file will be transferred to each sshlogin before a jobs is started. It will be removed if --cleanup is active. The file may be a script to run or some common base data needed for the jobs. Multiple -B can be specified to transfer more basefiles. The file will be transferred the same way as --transfer.
Run command in background thus GNU parallel will not wait for completion of the command before exiting. This is the default if --semaphore is set.
See also: --fg
Implies --semaphore.
Remove transferred files. --cleanup will remove the transferred files on the remote server after processing is done.
find log -name '*gz' | parallel \ --sshlogin server.example.com --transfer --return {.}.bz2 \ --cleanup "zcat {} | bzip -9 >{.}.bz2"
With --transfer the file transferred to the remote server will be removed on the remote server. Directories created will not be removed - even if they are empty.
With --return the file transferred from the remote server will be removed on the remote server. Directories created will not be removed - even if they are empty.
--cleanup is ignored when not used with --transfer or --return.
Column separator. The input will be treated as a table with regexp separating the columns. The n'th column can be access using {n} or {n.}. E.g. {3} is the 3rd column.
--colsep implies --trim rl.
regexp is a Perl Regular Expression: http://perldoc.perl.org/perlre.html
Line is a command. The input line contains more than one argument or the input line needs to be evaluated by the shell. This is the default if command is not set. Can be reversed with --file.
Most people will never need this because GNU parallel normally selects the correct --file or --command.
Input items are terminated by the specified character. Quotes and backslash are not special; every character in the input is taken literally. Disables the end-of-file string, which is treated like any other argument. This can be used when the input consists of simply newline-separated items, although it is almost always better to design your program to use --null where this is possible. The specified delimiter may be a single character, a C-style character escape such as \n, or an octal or hexadecimal escape code. Octal and hexadecimal escape codes are understood as for the printf command. Multibyte characters are not supported.
Set the end of file string to eof-str. If the end of file string occurs as a line of input, the rest of the input is ignored. If neither -E nor -e is used, no end of file string is used.
This option is a synonym for the -E option. Use -E instead, because it is POSIX compliant for xargs while this option is not. If eof-str is omitted, there is no end of file string. If neither -E nor -e is used, no end of file string is used.
Show the estimated number of seconds before finishing. This forces GNU parallel to read all jobs before starting to find the number of jobs. GNU parallel normally only reads the next job to run. Implies --progress.
Run command in foreground thus GNU parallel will wait for completion of the command before exiting.
See also: --bg
Implies --semaphore.
Line is a filename. The input line contains a filename that will be quoted so it is not evaluated by the shell. This is the default if command is set. Can be reversed with --command.
Most people will never need this because GNU parallel normally selects the correct --file or --command.
Group output. Output from each jobs is grouped together and is only printed when the command is finished. STDERR first followed by STDOUT. -g is the default. Can be reversed with -u.
Print a summary of the options to GNU parallel and exit.
Do not halt if a job fails. Exit status will be the number of jobs failed. This is the default.
Do not start new jobs if a job fails, but complete the running jobs including cleanup. The exit status will be the exit status from the last failing job.
Kill off all jobs immediately and exit without cleanup. The exit status will be the exit status from the failing job.
Use the replacement string replace-str instead of {}.
This option is a synonym for -Ireplace-str if replace-str is specified, and for -I{} otherwise. This option is deprecated; use -I instead.
Run up to N jobs in parallel. 0 means as many as possible. Default is 9.
If --semaphore is set default is 1 thus making a mutex.
Add N to the number of CPU cores. Run this many jobs in parallel. For compute intensive jobs -j +0 is useful as it will run number-of-cpu-cores jobs in parallel. See also --use-cpus-instead-of-cores.
Subtract N from the number of CPU cores. Run this many jobs in parallel. If the evaluated number is less than 1 then 1 will be used. See also --use-cpus-instead-of-cores.
Multiply N% with the number of CPU cores. Run this many jobs in parallel. If the evaluated number is less than 1 then 1 will be used. See also --use-cpus-instead-of-cores.
Keep sequence of output same as the order of input. If jobs 1 2 3 4 end in the sequence 3 1 4 2 the output will still be 1 2 3 4.
Use at most max-lines nonblank input lines per command line. Trailing blanks cause an input line to be logically continued on the next input line.
Implies -X unless -m is set.
Synonym for the -L option. Unlike -L, the max-lines argument is optional. If max-lines is not specified, it defaults to one. The -l option is deprecated since the POSIX standard specifies -L instead.
Implies -X unless -m is set.
Use ssh's ControlMaster to make ssh connections faster. Useful if jobs run remote and are very fast to run. This is disabled for sshlogins that specify their own ssh command.
Multiple. Insert as many arguments as the command line length permits. If {} is not used the arguments will be appended to the line. If {} is used multiple times each {} will be replaced with all the arguments.
Support for -m with --sshlogin is limited and may fail.
See also -X for context replace.
Show progress of computations. List the computers involved in the task with number of CPU cores detected and the max number of jobs to run. After that show progress for each computer: number of running jobs, number of completed jobs, and percentage of all jobs done by this computer. The percentage will only be available after all jobs have been scheduled as GNU parallel only read the next job when ready to schedule it - this is to avoid wasting time and memory by reading everything at startup.
By sending GNU parallel SIGUSR2 you can toggle turning on/off --progress on a running GNU parallel process.
Use at most max-args arguments per command line. Fewer than max-args arguments will be used if the size (see the -s option) is exceeded, unless the -x option is given, in which case GNU parallel will exit.
Implies -X unless -m is set.
Use at most max-args arguments per command line. Like -n but also makes replacement strings {1} .. {max-args} that represents argument 1 .. max-args. If too few args the {n} will be empty.
This will set the owner of the homedir to the user:
tr ':' '\012' < /etc/passwd | parallel -N7 chown {1} {6}
Implies -X unless -m is set.
Print the maximal number characters allowed on the command line and exit (used by GNU parallel itself to determine the line length on remote computers).
Print the number of physical CPUs and exit (used by GNU parallel itself to determine the number of physical CPUs on remote computers).
Print the number of CPU cores and exit (used by GNU parallel itself to determine the number of CPU cores on remote computers).
Prompt the user about whether to run each command line and read a line from the terminal. Only run the command line if the response starts with 'y' or 'Y'. Implies -t.
Quote command. This will quote the command line so special characters are not interpreted by the shell. See the section QUOTING. Most people will never need this. Quoting is disabled by default.
If the stdin (standard input) only contains whitespace, do not run the command.
Transfer files from remote servers. --return is used with --sshlogin when the arguments are files on the remote servers. When processing is done the file filename will be transferred from the remote server using rsync and will be put relative to the default login dir. E.g.
echo foo/bar.txt | parallel \ --sshlogin server.example.com --return {.}.out touch {.}.out
This will transfer the file $HOME/foo/bar.out from the server server.example.com to the file foo/bar.out after running touch foo/bar.out on server.example.com.
echo /tmp/foo/bar.txt | parallel \ --sshlogin server.example.com --return {.}.out touch {.}.out
This will transfer the file /tmp/foo/bar.out from the server server.example.com to the file /tmp/foo/bar.out after running touch /tmp/foo/bar.out on server.example.com.
Multiple files can be transferred by repeating the options multiple times:
echo /tmp/foo/bar.txt | \ parallel --sshlogin server.example.com \ --return {.}.out --return {.}.out2 touch {.}.out {.}.out2
--return is often used with --transfer and --cleanup.
--return is ignored when used with --sshlogin : or when not used with --sshlogin.
Use at most max-chars characters per command line, including the command and initial-arguments and the terminating nulls at the ends of the argument strings. The largest allowed value is system-dependent, and is calculated as the argument length limit for exec, less the size of your environment. The default value is the maximum.
Implies -X unless -m is set.
Display the limits on the command-line length which are imposed by the operating system and the -s option. Pipe the input from /dev/null (and perhaps specify --no-run-if-empty) if you don't want GNU parallel to do anything.
Work as a counting semaphore. --semaphore will cause GNU parallel to start command in the background. When the number of simultaneous jobs is reached, GNU parallel will wait for one of these to complete before starting another command.
--semaphore implies --bg unless --fg is specified.
--semaphore implies --semaphorename `tty` unless --semaphorename is specified.
Used with --fg, --wait, and --semaphorename.
The command sem is an alias for parallel --semaphore.
The name of the semaphore to use. The semaphore can be shared between multiple processes.
Implies --semaphore.
If the semaphore is not released within secs seconds, take it anyway.
Implies --semaphore.
Distribute jobs to remote servers. The jobs will be run on a list of remote servers. GNU parallel will determine the number of CPU cores on the remote servers and run the number of jobs as specified by -j. If the number ncpu is given GNU parallel will use this number for number of CPU cores on the host. Normally ncpu will not be needed.
An sshlogin is of the form:
[sshcommand [options]][username@]hostname
The sshlogin must not require a password.
The sshlogin ':' is special, it means 'no ssh' and will therefore run on the local computer.
The sshlogin '..' is special, it read sshlogins from ~/.parallel/sshloginfile
To specify more sshlogins separate the sshlogins by comma or repeat the options multiple times.
For examples: see --sshloginfile.
The remote host must have GNU parallel installed.
--sshlogin is known to cause problems with -m and -X.
File with sshlogins. The file consists of sshlogins on separate lines. Empty lines and lines starting with '#' are ignored. Example:
server.example.com username@server2.example.com 8/my-8-core-server.example.com 2/my_other_username@my-dualcore.example.net # This server has SSH running on port 2222 ssh -p 2222 server.example.net 4/ssh -p 2222 quadserver.example.net # Use a different ssh program myssh -p 2222 -l myusername hexacpu.example.net # Use a different ssh program with default number of cores //usr/local/bin/myssh -p 2222 -l myusername hexacpu.example.net # Use a different ssh program with 6 cores 6//usr/local/bin/myssh -p 2222 -l myusername hexacpu.example.net # Assume 16 cores on the local computer 16/:
When using a different ssh program the last argument must be the hostname.
The sshloginfile '..' is special, it read sshlogins from ~/.parallel/sshloginfile
Silent. The job to be run will not be printed. This is the default. Can be reversed with -v.
Print the command line on the standard error output before executing it.
See also -v.
Transfer files to remote servers. --transfer is used with --sshlogin when the arguments are files and should be transferred to the remote servers. The files will be transferred using rsync and will be put relative to the default login dir. E.g.
echo foo/bar.txt | parallel \ --sshlogin server.example.com --transfer wc
This will transfer the file foo/bar.txt to the server server.example.com to the file $HOME/foo/bar.txt before running wc foo/bar.txt on server.example.com.
echo /tmp/foo/bar.txt | parallel \ --sshlogin server.example.com --transfer wc
This will transfer the file foo/bar.txt to the server server.example.com to the file /tmp/foo/bar.txt before running wc /tmp/foo/bar.txt on server.example.com.
--transfer is often used with --return and --cleanup.
--transfer is ignored when used with --sshlogin : or when not used with --sshlogin.
Transfer, Return, Cleanup. Short hand for:
--transfer --return filename --cleanup
Trim white space in input.
No trim. Input is not modified. This is the default.
Left trim. Remove white space from start of input. E.g. " a bc " -> "a bc ".
Right trim. Remove white space from end of input. E.g. " a bc " -> " a bc".
Both trim. Remove white space from both start and end of input. E.g. " a bc " -> "a bc". This is the default if --colsep is used.
Ungroup output. Output is printed as soon as possible. This may cause output from different commands to be mixed. GNU parallel runs faster with -u. Can be reversed with -g.
Use the replacement string replace-str instead of {.} for input line without extension.
Count the number of physical CPUs instead of CPU cores. When computing how many jobs to run in parallel relative to the number of CPU cores you can ask GNU parallel to instead look at the number of physical CPUs. This will make sense for computers that have hyperthreading as two jobs running on one CPU with hyperthreading will run slower than two jobs running on two physical CPUs. Some multi-core CPUs can run faster if only one thread is running per physical CPU. Most users will not need this option.
Verbose. Print the job to be run on STDOUT. Can be reversed with --silent. See also -t.
Print the version GNU parallel and exit.
Wait for all commands to complete.
Implies --semaphore.
xargs with context replace. This works like -m except if {} is part of a word (like pic{}.jpg) then the whole word will be repeated. Normally -X will do the right thing, whereas -m can give surprising results if {} is used as part of a word.
Support for -X with --sshlogin is limited and may fail.
Exit if the size (see the -s option) is exceeded.
GNU parallel can work similar to xargs -n1.
To compress all html files using gzip run:
find . -name '*.html' | parallel gzip
If the file names may contain a newline use -0. Substitute FOO BAR with FUBAR in all files in this dir and subdirs:
find . -type f -print0 | parallel -q0 perl -i -pe 's/FOO BAR/FUBAR/g'
Note -q is needed because of the space in 'FOO BAR'.
GNU parallel can take the arguments from command line instead of stdin (standard input). To compress all html files in the current dir using gzip run:
parallel gzip ::: *.html
To convert *.wav to *.mp3 using LAME running one process per CPU core run:
parallel -j+0 lame {} -o {.}.mp3 ::: *.wav
When moving a lot of files like this: mv * destdir you will sometimes get the error:
bash: /bin/mv: Argument list too long
because there are too many files. You can instead do:
ls | parallel mv {} destdir
This will run mv for each file. It can be done faster if mv gets as many arguments that will fit on the line:
ls | parallel -m mv {} destdir
To remove the files pict0000.jpg .. pict9999.jpg you could do:
seq -w 0 9999 | parallel rm pict{}.jpg
You could also do:
seq -w 0 9999 | perl -pe 's/(.*)/pict$1.jpg/' | parallel -m rm
The first will run rm 10000 times, while the last will only run rm as many times needed to keep the command line length short enough to avoid Argument list too long (it typically runs 1-2 times).
You could also run:
seq -w 0 9999 | parallel -X rm pict{}.jpg
This will also only run rm as many times needed to keep the command line length short enough.
If ImageMagick is installed this will generate a thumbnail of a jpg file:
convert -geometry 120 foo.jpg thumb_foo.jpg
If the system has more than 1 CPU core it can be run with number-of-cpu-cores jobs in parallel (-j +0). This will do that for all jpg files in a directory:
ls *.jpg | parallel -j +0 convert -geometry 120 {} thumb_{}
To do it recursively use find:
find . -name '*.jpg' | parallel -j +0 convert -geometry 120 {} {}_thumb.jpg
Notice how the argument has to start with {} as {} will include path (e.g. running convert -geometry 120 ./foo/bar.jpg thumb_./foo/bar.jpg would clearly be wrong). The command will generate files like ./foo/bar.jpg_thumb.jpg.
Use {.} to avoid the extra .jpg in the file name. This command will make files like ./foo/bar_thumb.jpg:
find . -name '*.jpg' | parallel -j +0 convert -geometry 120 {} {.}_thumb.jpg
This will generate an uncompressed version of .gz-files next to the .gz-file:
parallel zcat {} ">"{.} ::: *.gz
Quoting of > is necessary to postpone the redirection. Another solution is to quote the whole command:
parallel "zcat {} >{.}" ::: *.gz
Other special shell charaters (such as * ; $ > < | >> <<) also needs to be put in quotes, as they may otherwise be interpreted by the shell and not given to GNU parallel.
A job can consist of several commands. This will print the number of files in each directory:
ls | parallel 'echo -n {}" "; ls {}|wc -l'
To put the output in a file called <name>.dir:
ls | parallel '(echo -n {}" "; ls {}|wc -l) > {}.dir'
Even small shell scripts can be run by GNU parallel:
find . | parallel 'a={}; name=${a##*/}; upper=$(echo "$name" | tr "[:lower:]" "[:upper:]"); echo "$name - $upper"'
Given a list of URLs, list all URLs that fail to download. Print the line number and the URL.
cat urlfile | parallel "wget {} 2>/dev/null || grep -n {} urlfile"
When processing files removing the file extension using {.} is often useful.
Create a directory for each zip-file and unzip it in that dir:
parallel 'mkdir {.}; cd {.}; unzip ../{}' ::: *.zip
Recompress all .gz files in current directory using bzip2 running 1 job per CPU core in parallel:
parallel -j+0 "zcat {} | bzip2 >{.}.bz2 && rm {}" ::: *.gz
Convert all WAV files to MP3 using LAME:
find sounddir -type f -name '*.wav' | parallel -j+0 lame {} -o {.}.mp3
If you have directory with tar.gz files and want these extracted in the corresponding dir (e.g foo.tar.gz will be extracted in the dir foo) you can do:
ls *.tar.gz| parallel -U {tar} 'echo {tar}|parallel "mkdir -p {.} ; tar -C {.} -xf {.}.tar.gz"'
Let us assume a website stores images like:
http://www.website.com/path/to/YYYYMMDD_##.jpg
where YYYYMMDD is the date and ## is the number 01-10. This will generate the past 30 days as YYYYMMDD:
seq 1 30 | parallel date -d '"today -{} days"' +%Y%m%d
Based on this we can let GNU parallel generate 10 wgets per day:
the above | parallel -I {o} seq -w 1 10 "|" parallel wget http://www.website.com/path/to/{o}_{}.jpg
for-loops like this:
(for x in `cat list` ; do do_something $x done) | process_output
and while-loops like this:
cat list | (while read x ; do do_something $x done) | process_output
can be written like this:
cat list | parallel do_something | process_output
If the processing requires more steps the for-loop like this:
(for x in `cat list` ; do no_extension=${x%.png}; do_something $x scale $no_extension.jpg do_step2 <$x $no_extension done) | process_output
and while-loops like this:
cat list | (while read x ; do no_extension=${x%.png}; do_something $x scale $no_extension.jpg do_step2 <$x $no_extension done) | process_output
can be written like this:
cat list | parallel "do_something {} scale {.}.jpg ; do_step2 <{} {.}" | process_output
When runnning jobs that output data, you often do not want the output of multiple jobs to run together. GNU parallel defaults to grouping the output of each job, so the output is printed when the job finishes. If you want the output to be printed while the job is running you can use -u.
Compare the output of:
parallel traceroute ::: foss.org.my debian.org freenetproject.org
to the output of:
parallel -u traceroute ::: foss.org.my debian.org freenetproject.org
Normally the output of a job will be printed as soon as it completes. Sometimes you want the order of the output to remain the same as the order of the input. This is often important, if the output is used as input for another system. -k will make sure the order of output will be in the same order as input even if later jobs end before earlier jobs.
Append a string to every line in a text file:
cat textfile | parallel -k echo {} append_string
If you remove -k some of the lines may come out in the wrong order.
Another example is traceroute:
parallel traceroute ::: foss.org.my debian.org freenetproject.org
will give traceroute of foss.org.my, debian.org and freenetproject.org, but it will be sorted according to which job completed first.
To keep the order the same as input run:
parallel -k traceroute ::: foss.org.my debian.org freenetproject.org
This will make sure the traceroute to foss.org.my will be printed first.
grep -r greps recursively through directories. On multicore CPUs GNU parallel can often speed this up.
find . -type f | parallel -k -j150% -n 1000 -m grep -H -n STRING {}
This will run 1.5 job per core, and give 1000 arguments to grep.
To run commands on a remote computer SSH needs to be set up and you must be able to login without entering a password (ssh-agent may be handy).
To run echo on server.example.com:
seq 1 10 | parallel --sshlogin server.example.com echo
To run commands on more than one remote computer run:
seq 1 10 | parallel --sshlogin server.example.com,server2.example.net echo
Or:
seq 1 10 | parallel --sshlogin server.example.com \ --sshlogin server2.example.net echo
If the login username is foo on server2.example.net use:
seq 1 10 | parallel --sshlogin server.example.com \ --sshlogin foo@server2.example.net echo
To distribute the commands to a list of computers, make a file mycomputers with all the computers:
server.example.com foo@server2.example.com server3.example.com
Then run:
seq 1 10 | parallel --sshloginfile mycomputers echo
To include the local computer add the special sshlogin ':' to the list:
server.example.com foo@server2.example.com server3.example.com :
GNU parallel will try to determine the number of CPU cores on each of the remote computers, so -j+0 will run one job per CPU core - even if the remote computers do not have the same number of CPU cores.
If the number of CPU cores on the remote servers is not identified correctly the number of CPU cores can be added in front. Here the server has 8 CPU cores.
seq 1 10 | parallel --sshlogin 8/server.example.com echo
To recompress gzipped files with bzip2 using a remote server run:
find logs/ -name '*.gz' | \ parallel --sshlogin server.example.com \ --transfer "zcat {} | bzip2 -9 >{.}.bz2"
This will list the .gz-files in the logs directory and all directories below. Then it will transfer the files to server.example.com to the corresponding directory in $HOME/logs. On server.example.com the file will be recompressed using zcat and bzip2 resulting in the corresponding file with .gz replaced with .bz2.
If you want the resulting bz2-file to be transferred back to the local computer add --return {.}.bz2:
find logs/ -name '*.gz' | \ parallel --sshlogin server.example.com \ --transfer --return {.}.bz2 "zcat {} | bzip2 -9 >{.}.bz2"
After the recompressing is done the .bz2-file is transferred back to the local computer and put next to the original .gz-file.
If you want to delete the transferred files on the remote computer add --cleanup. This will remove both the file transferred to the remote computer and the files transferred from the remote computer:
find logs/ -name '*.gz' | \ parallel --sshlogin server.example.com \ --transfer --return {.}.bz2 --cleanup "zcat {} | bzip2 -9 >{.}.bz2"
If you want run on several servers add the servers to --sshlogin either using ',' or multiple --sshlogin:
find logs/ -name '*.gz' | \ parallel --sshlogin server.example.com,server2.example.com \ --sshlogin server3.example.com \ --transfer --return {.}.bz2 --cleanup "zcat {} | bzip2 -9 >{.}.bz2"
You can add the local computer using --sshlogin :. This will disable the removing and transferring for the local computer only:
find logs/ -name '*.gz' | \ parallel --sshlogin server.example.com,server2.example.com \ --sshlogin server3.example.com \ --sshlogin : \ --transfer --return {.}.bz2 --cleanup "zcat {} | bzip2 -9 >{.}.bz2"
Often --transfer, --return and --cleanup are used together. They can be shortened to --trc:
find logs/ -name '*.gz' | \ parallel --sshlogin server.example.com,server2.example.com \ --sshlogin server3.example.com \ --sshlogin : \ --trc {.}.bz2 "zcat {} | bzip2 -9 >{.}.bz2"
With the file mycomputers containing the list of computers it becomes:
find logs/ -name '*.gz' | parallel --sshloginfile mycomputers \ --trc {.}.bz2 "zcat {} | bzip2 -9 >{.}.bz2"
Convert *.mp3 to *.ogg running one process per CPU core on local computer and server2:
parallel --trc {.}.ogg -j+0 -S server2,: \ 'mpg321 -w - {} | oggenc -q0 - -o {.}.ogg' ::: *.mp3
Copy files like foo.es.ext to foo.ext:
ls *.es.* | perl -pe 'print; s/\.es//' | parallel -N2 cp {1} {2}
The perl command spits out 2 lines for each input. GNU parallel takes 2 inputs (using -N2) and replaces {1} and {2} with the inputs.
Content of table_file.tsv:
foo<TAB>bar baz <TAB> quux
To run:
cmd -o bar -i foo cmd -o quux -i baz
you can run:
parallel -a table_file.tsv --colsep '\t' cmd -o {2} -i {1}
Note: The default for GNU parallel is to remove the spaces around the columns. To keep the spaces:
parallel -a table_file.tsv --trim n --colsep '\t' cmd -o {2} -i {1}
GNU parallel can work similar to cat | sh.
A ressource inexpensive job is a job that takes very little CPU, disk I/O and network I/O. Ping is an example of a ressource inexpensive job. wget is too - if the webpages are small.
The content of the file jobs_to_run:
ping -c 1 10.0.0.1 wget http://status-server/status.cgi?ip=10.0.0.1 ping -c 1 10.0.0.2 wget http://status-server/status.cgi?ip=10.0.0.2 ... ping -c 1 10.0.0.255 wget http://status-server/status.cgi?ip=10.0.0.255
To run 100 processes simultaneously do:
parallel -j 100 < jobs_to_run
As there is not a command the option --command is default because the jobs needs to be evaluated by the shell.
The command sem is an alias for parallel --semaphore.
A counting semaphore will allow a given number of jobs to be started in the background. When the number of jobs are running in the background, GNU sem will wait for one of these to complete before starting another command. sem --wait will wait for all jobs to complete.
Run 10 jobs in parallel in the background:
for i in `ls *.log` ; do echo $i sem -j10 gzip $i ";" echo done done sem --wait
A mutex is a counting semaphore allowing only one job to run. This will edit the file myfile and prepends the file with lines with the numbers 1 to 3.
seq 1 3 | parallel sem sed -i -e 'i{}' myfile
As myfile can be very big it is important only one process edits the file at the same time.
Name the semaphore to have multiple different semaphores active at the same time:
seq 1 3 | parallel sem --id mymutex sed -i -e 'i{}' myfile
For more advanced use quoting may be an issue. The following will print the filename for each line that has exactly 2 columns:
perl -ne '/^\S+\s+\S+$/ and print $ARGV,"\n"' file
This can be done by GNU parallel using:
ls | parallel "perl -ne '/^\\S+\\s+\\S+$/ and print \$ARGV,\"\\n\"'"
Notice how \'s, "'s, and $'s needs to be quoted. GNU parallel can do the quoting by using option -q:
ls | parallel -q perl -ne '/^\S+\s+\S+$/ and print $ARGV,"\n"'
However, this means you cannot make the shell interpret special characters. For example this will not work:
ls *.gz | parallel -q "zcat {} >{.}"
ls *.gz | parallel -q "zcat {} | bzip2 >{.}.bz2"
because > and | need to be interpreted by the shell.
If you get errors like:
sh: -c: line 0: syntax error near unexpected token
then you might try using -q.
If you are using bash process substitution like <(cat foo) then you may try -q and prepending command with bash -c:
ls | parallel -q bash -c 'wc -c <(echo {})'
Or for substituting output:
ls | parallel -q bash -c 'tar c {} | tee >(gzip >{}.tar.gz) | bzip2 >{}.tar.bz2'
Conclusion: To avoid dealing with the quoting problems it may be easier just to write a small script and have GNU parallel call that script.
If you want a list of the jobs currently running you can run:
killall -USR1 parallel
GNU parallel will then print the currently running jobs on STDERR.
If you regret starting a lot of jobs you can simply break GNU parallel, but if you want to make sure you do not have halfcompleted jobs you should send the signal SIGTERM to GNU parallel:
killall -TERM parallel
This will tell GNU parallel to not start any new jobs, but wait until the currently running jobs are finished before exiting.
The environment variable $PARALLEL_PID is set by GNU parallel and is visible to the jobs started from GNU parallel. This makes it possible for the jobs to communicate directly to GNU parallel.
Example: If each of the jobs tests a solution and one of jobs finds the solution the job can tell GNU parallel not to start more jobs by: kill -TERM $PARALLEL_PID. This only works on the local computer.
The environment variable $PARALLEL will be used as default options for GNU parallel. However, because some options take arguments the options need to be split into groups in which only the last option takes an argument. Each group of options should be put on a line of its own.
Example:
cat list | parallel -j1 -k -v ls
can be written as:
cat list | PARALLEL="-kvj1" parallel ls
cat list | parallel -j1 -k -v -S"myssh user@server" ls
can be written as:
cat list | PARALLEL="-kvj1
-Smyssh user@server" parallel echo
Notice the newline in the middle is needed because both -S and -j take an argument and thus both need to be at the end of a group.
The file ~/.parallel/config will be read if it exists. It should be formatted like the environment variable $PARALLEL. Lines starting with '#' will be ignored.
Options on the command line takes precedence over the environment variable $PARALLEL which takes precedence over the file ~/.parallel/config.
If --halt-on-error 0 or not specified:
All jobs ran without error.
Some of the jobs failed. The exit status gives the number of failed jobs
More than 253 jobs failed.
Other error.
If --halt-on-error 1 or 2: Exit status of the failing job.
There are a lot programs with some of the functionality of GNU parallel. GNU parallel strives to include the best of the functionality without sacrifying ease of use.
The following features are in some of the comparable tools:
Execution E1. Running jobs in parallel E2. List running jobs E3. Finish running jobs, but do not start new jobs E4. Number of running jobs can depend on number of cpus E5. Finish running jobs, but do not start new jobs after first failure
Outputs O1. Grouping output so output from different jobs do not mix O2. Send stderr to stderr O3. Send stdout to stdout O4. Order of output can be same as order of input O5. Stdout only from the command O6. Stderr only from the command
Inputs I1. Arguments can be read from stdin I2. Arguments can be read from a file I3. Arguments can be read from multiple files I4. Arguments can be read from command line I5. Arguments can be read from a table I6. Line oriented input as default (Quoting of special chars not needed)
Manipulation of input M1. Composed command M2. Multiple arguments can fill up an execution line M3. Arguments can be put anywhere in the execution line M4. Multiple arguments can be put anywhere in the execution line M5. Arguments can be replaced with context M6. Input can be treated as complete execution line
Remote execution R1. Jobs can be run on remote computers R2. Basefiles can be transferred R3. Argument files can be transferred R4. Result files can be transferred R5. Cleanup of transferred files R6. No config files needed R7. Do not run more than SSHD's MaxStartup can handle R8. Configurable SSH command
Semaphore S1. Possibility to work as a mutex S2. Possibility to work as a counting semaphore
Legend - = no x = not applicable ID = yes
As every new version of the programs are not tested the table may be outdated. Please file a bug-report if you find errors.
parallel: E1 E2 E3 E4 E5 O1 O2 O3 O4 O5 O6 I1 I2 I3 I4 I5 I6 M1 M2 M3 M4 M5 M6 R1 R2 R3 R4 R5 R6 R7 R8 S1 S2
xargs: E1 - - - - - O2 O3 - O5 O6 I1 I2 - - - - - M2 M3 - - - - - - - - x - - -
find -exec: - - - x - x O2 O3 O4 O5 O6 - - - - - - - M2 M3 - - - - - - - - - - x x
make -j: E1 - - - E5 O1 O2 O3 - x O6 - - - - - - - - - - - - - - - - - - - -
ppss: E1 E2 ?E3 E4 - O1 - - x - - I1 I2 - - - I6 M1 - M3 - - M6 R1 R2 R3 R4 - - ?R7 - -
pexec: E1 - - E4 - O1 O2 O3 - O5 O6 I1 I2 - I4 I5 - M1 - M3 - - M6 R1 - - - - R6 - S1 -
xjobs: TODO
prll: TODO
dxargs: TODO
mdm/middelman: TODO
xapply: TODO
ClusterSSH: TODO
xargs offer some of the same possibilites as GNU parallel.
xargs deals badly with special characters (such as space, ' and "). To see the problem try this:
touch important_file touch 'not important_file' ls not* | xargs rm mkdir -p '12" records' ls | xargs rmdir
You can specify -0 or -d "\n", but many input generators are not optimized for using NUL as separator but are optimized for newline as separator. E.g head, tail, awk, ls, echo, sed, tar -v, perl (-0 and \0 instead of \n), locate (requires using -0), find (requires using -print0), grep (requires user to use -z or -Z), sort (requires using -z).
So GNU parallel's newline separation can be emulated with:
cat | xargs -d "\n" -n1 command
xargs can run a given number of jobs in parallel, but has no support for running number-of-cpu-cores jobs in parallel.
xargs has no support for grouping the output, therefore output may run together, e.g. the first half of a line is from one process and the last half of the line is from another process. The example Parallel grep cannot be done reliably with xargs because of this.
xargs has no support for keeping the order of the output, therefore if running jobs in parallel using xargs the output of the second job cannot be postponed till the first job is done.
xargs has no support for running jobs on remote computers.
xargs has no support for context replace, so you will have to create the arguments.
If you use a replace string in xargs (-I) you can not force xargs to use more than one argument.
Quoting in xargs works like -q in GNU parallel. This means composed commands and redirection require using bash -c.
ls | parallel "wc {} > {}.wc"
becomes
ls | xargs -d "\n" -P9 -I {} bash -c "wc {} > {}.wc"
and
ls | parallel "echo {}; ls {}|wc"
becomes
ls | xargs -d "\n" -P9 -I {} bash -c "echo {}; ls {}|wc"
find -exec offer some of the same possibilites as GNU parallel.
find -exec only works on files. So processing other input (such as hosts or URLs) will require creating these inputs as files. find -exec has no support for running commands in parallel.
make -j can run jobs in parallel, but requires a crafted Makefile to do this. That results in extra quoting to get filename containing newline to work correctly.
make -j has no support for grouping the output, therefore output may run together, e.g. the first half of a line is from one process and the last half of the line is from another process. The example Parallel grep cannot be done reliably with make -j because of this.
(Very early versions of GNU parallel were coincidently implemented using make -j).
ppss is also a tool for running jobs in parallel.
The output of ppss is status information and thus not useful for using as input for another command. The output from the jobs are put into files.
The argument replace string ($ITEM) cannot be changed. Arguments must be quoted - thus arguments containing special characters (space '"&!*) may cause problems. More than one argument is not supported. File names containing newlines are not processed correctly. When reading input from a file null cannot be used terminator. ppss needs to read the whole input file before starting any jobs.
Output and status information is stored in ppss_dir and thus requires cleanup when completed. If the dir is not removed before running ppss again it may cause nothing to happen as ppss thinks the task is already done. GNU parallel will normally not need cleaning up if running locally and will only need cleaning up if stopped abnormally and running remote (--cleanup may not complete if stopped abnormally). The example Parallel grep would require extra postprocessing if written using ppss.
For remote systems PPSS requires 3 steps: config, deploy, and start. GNU parallel only requires one step.
Here are the examples from ppss's manual page with the equivalent using GNU parallel:
1 ./ppss.sh standalone -d /path/to/files -c 'gzip '
1 find /path/to/files -type f | parallel -j+0 gzip
2 ./ppss.sh standalone -d /path/to/files -c 'cp "$ITEM" /destination/dir '
2 find /path/to/files -type f | parallel -j+0 cp {} /destination/dir
3 ./ppss.sh standalone -f list-of-urls.txt -c 'wget -q '
3 parallel -a list-of-urls.txt wget -q
4 ./ppss.sh standalone -f list-of-urls.txt -c 'wget -q "$ITEM"'
4 parallel -a list-of-urls.txt wget -q {}
5 ./ppss config -C config.cfg -c 'encode.sh ' -d /source/dir -m 192.168.1.100 -u ppss -k ppss-key.key -S ./encode.sh -n nodes.txt -o /some/output/dir --upload --download ; ./ppss deploy -C config.cfg ; ./ppss start -C config
5 # parallel does not use configs. If you want a different username put it in nodes.txt: user@hostname
5 find source/dir -type f | parallel --sshloginfile nodes.txt --trc {.}.mp3 lame -a {} -o {.}.mp3 --preset standard --quiet
6 ./ppss stop -C config.cfg
6 killall -TERM parallel
7 ./ppss pause -C config.cfg
7 Press: CTRL-Z or killall -SIGTSTP parallel
8 ./ppss continue -C config.cfg
8 Enter: fg or killall -SIGCONT parallel
9 ./ppss.sh status -C config.cfg
9 killall -SIGUSR2 parallel
pexec is also a tool for running jobs in parallel.
Here are the examples from pexec's info page with the equivalent using GNU parallel:
1 pexec -o sqrt-%s.dat -p "$(seq 10)" -e NUM -n 4 -c -- \ 'echo "scale=10000;sqrt($NUM)" | bc'
1 seq 10 | parallel -j4 'echo "scale=10000;sqrt({})" | bc > sqrt-{}.dat'
2 pexec -p "$(ls myfiles*.ext)" -i %s -o %s.sort -- sort
2 ls myfiles*.ext | parallel sort {} ">{}.sort"
3 pexec -f image.list -n auto -e B -u star.log -c -- \ 'fistar $B.fits -f 100 -F id,x,y,flux -o $B.star'
3 parallel -a image.list -j+0 \ 'fistar {}.fits -f 100 -F id,x,y,flux -o {}.star' 2>star.log
4 pexec -r *.png -e IMG -c -o - -- \ 'convert $IMG ${IMG%.png}.jpeg ; "echo $IMG: done"'
4 ls *.png | parallel 'convert {} {.}.jpeg; echo {}: done'
5 pexec -r *.png -i %s -o %s.jpg -c 'pngtopnm | pnmtojpeg'
5 ls *.png | parallel 'pngtopnm < {} | pnmtojpeg > {}.jpg'
6 for p in *.png ; do echo ${p%.png} ; done | \ pexec -f - -i %s.png -o %s.jpg -c 'pngtopnm | pnmtojpeg'
6 ls *.png | parallel 'pngtopnm < {} | pnmtojpeg > {.}.jpg'
7 LIST=$(for p in *.png ; do echo ${p%.png} ; done) pexec -r $LIST -i %s.png -o %s.jpg -c 'pngtopnm | pnmtojpeg'
7 ls *.png | parallel 'pngtopnm < {} | pnmtojpeg > {.}.jpg'
8 pexec -n 8 -r *.jpg -y unix -e IMG -c \ 'pexec -j -m blockread -d $IMG | \ jpegtopnm | pnmscale 0.5 | pnmtojpeg | \ pexec -j -m blockwrite -s th_$IMG'
8 Combining GNU parallel and GNU sem.
8 ls *jpg | parallel -j8 'sem --id blockread cat {} | jpegtopnm |' \ 'pnmscale 0.5 | pnmtojpeg | sem --id blockwrite cat > th_{}'
8 If reading and writing is done to the same disk, this may be faster as only one process will be either reading or writing:
8 ls *jpg | parallel -j8 'sem --id diskio cat {} | jpegtopnm |' \ 'pnmscale 0.5 | pnmtojpeg | sem --id diskio cat > th_{}'
xjobs is also a tool for running jobs in parallel. It only supports running jobs on your local computer.
xjobs deals badly with special characters just like xargs. See the section DIFFERENCES BETWEEN xargs AND GNU Parallel.
Here are the examples from xjobs's man page with the equivalent using GNU parallel:
1 ls -1 *.zip | xjobs unzip
1 ls *.zip | parallel unzip
2 ls -1 *.zip | xjobs -n unzip
2 ls *.zip | parallel unzip >/dev/null
3 find . -name '*.bak' | xjobs gzip
3 find . -name '*.bak' | parallel gzip
4 ls -1 *.jar | sed 's/\(.*\)/\1 > \1.idx/' | xjobs jar tf
4 ls *.jar | parallel jar tf {} '>' {}.idx
5 xjobs -s script
5 cat script | parallel
6 mkfifo /var/run/my_named_pipe; xjobs -s /var/run/my_named_pipe & echo unzip 1.zip >> /var/run/my_named_pipe; echo tar cf /backup/myhome.tar /home/me >> /var/run/my_named_pipe
6 mkfifo /var/run/my_named_pipe; cat /var/run/my_named_pipe | parallel & echo unzip 1.zip >> /var/run/my_named_pipe; echo tar cf /backup/myhome.tar /home/me >> /var/run/my_named_pipe
prll is also a tool for running jobs in parallel. It does not support running jobs on remote computers.
prll encourages using BASH aliases and BASH functions instead of scripts. GNU parallel will never support running aliases and functions (see why http://www.perlmonks.org/index.pl?node_id=484296) but scripts or composed commands work just fine.
prll generates a lot of status information on STDERR which makes it harder to use the STDERR output of the job directly as input for another program.
Here is the example from prll's man page with the equivalent using GNU parallel:
prll -s 'mogrify -flip $1' *.jpg
parallel mogrify -flip ::: *.jpg
dxargs is also a tool for running jobs in parallel.
dxargs does not deal well with more simultaneous jobs than SSHD's MaxStartup. dxargs is only built for remote run jobs, but does not support transferring of files.
middleman(mdm) is also a tool for running jobs in parallel.
Here are the shellscripts of http://mdm.berlios.de/usage.html ported to GNU parallel:
seq 1 19 | parallel -j+0 buffon -o - | sort -n > result
cat files | parallel -j+0 cmd
find dir -execdir sem -j+0 cmd {} \;
xapply can run jobs in parallel on the local computer.
Here are the examples from xapply's man page with the equivalent using GNU parallel:
1 xapply '(cd %1 && make all)' */
1 parallel 'cd {} && make all' ::: */
2 xapply -f 'diff %1 ../version5/%1' manifest | more
2 parallel diff {} ../version5/{} < manifest | more
3 xapply -p/dev/null -f 'diff %1 %2' manifest1 checklist1
3 parallel diff {1} {2} :::: manifest1 checklist1
4 xapply 'indent' *.c
4 parallel indent ::: *.c
5 find ~ksb/bin -type f ! -perm -111 -print | xapply -f -v 'chmod a+x' -
5 find ~ksb/bin -type f ! -perm -111 -print | parallel -v chmod a+x
6 find */ -... | fmt 960 1024 | xapply -f -i /dev/tty 'vi' -
6 sh <(find */ -... | parallel -s 1024 echo vi)
7 find ... | xapply -f -5 -i /dev/tty 'vi' - - - - -
7 sh <(find ... |parallel -n5 echo vi)
8 xapply -fn "" /etc/passwd
8 parallel -k echo < /etc/passwd
9 tr ':' '\012' < /etc/passwd | xapply -7 -nf 'chown %1 %6' - - - - - - -
9 tr ':' '\012' < /etc/passwd | parallel -N7 chown {1} {6}
10 xapply '[ -d %1/RCS ] || echo %1' */
10 parallel '[ -d {}/RCS ] || echo {}' ::: */
11 xapply -f '[ -f %1 ] && echo %1' List | ...
11 parallel '[ -f {} ] && echo {}' < List | ...
ClusterSSH solves a different problem than GNU parallel.
ClusterSSH runs the same command with the same arguments on a list of machines - one per machine. This is typically used for administrating several machines that are almost identical.
GNU parallel runs the same (or different) commands with different arguments in parallel possibly using remote machines to help computing. If more than one machine is listed in -S GNU parallel may only use one of these (e.g. if there are 8 jobs to be run and one machine has 8 cores).
GNU parallel can be used as a poor-mans version of ClusterSSH:
cat hostlist | parallel ssh {} do_stuff
Because of the way newline is quoted this will not work:
echo 1,2,3 | parallel -vkd, "echo 'a{}'"
However, this will work:
echo 1,2,3 | parallel -vkd, echo a{}
GNU parallel is slow at starting up. Half of the startup time is spent finding the maximal length of a command line. Setting -s will remove this part of the startup time.
Report bugs to <bug-parallel@gnu.org>.
Copyright (C) 2007-10-18 Ole Tange, http://ole.tange.dk
Copyright (C) 2008,2009,2010 Ole Tange, http://ole.tange.dk
Copyright (C) 2010 Ole Tange, http://ole.tange.dk and Free Software Foundation, Inc.
Parts of the manual concerning xargs compatibility is inspired by the manual of xargs from GNU findutils 4.4.2.
Copyright (C) 2007,2008,2009,2010 Free Software Foundation, Inc.
This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3 of the License, or at your option any later version.
This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with this program. If not, see <http://www.gnu.org/licenses/>.
Permission is granted to copy, distribute and/or modify this documentation under the terms of the GNU Free Documentation License, Version 1.3 or any later version published by the Free Software Foundation; with no Invariant Sections, with no Front-Cover Texts, and with no Back-Cover Texts. A copy of the license is included in the file fdl.txt.
You are free:
to copy, distribute and transmit the work
to adapt the work
Under the following conditions:
You must attribute the work in the manner specified by the author or licensor (but not in any way that suggests that they endorse you or your use of the work).
If you alter, transform, or build upon this work, you may distribute the resulting work only under the same, similar or a compatible license.
With the understanding that:
Any of the above conditions can be waived if you get permission from the copyright holder.
Where the work or any of its elements is in the public domain under applicable law, that status is in no way affected by the license.
In no way are any of the following rights affected by the license:
Your fair dealing or fair use rights, or other applicable copyright exceptions and limitations;
The author's moral rights;
Rights other persons may have either in the work itself or in how the work is used, such as publicity or privacy rights.
For any reuse or distribution, you must make clear to others the license terms of this work.
A copy of the full license is included in the file as cc-by-sa.txt.
GNU parallel uses Perl, and the Perl modules Getopt::Long, IPC::Open3, Symbol, IO::File, POSIX, and File::Temp.
find(1), xargs(1), make(1), pexec(1), ppss(1), xjobs(1), prll(1), dxargs(1), mdm(1)