PARALLEL_ALTERNATIVES(7) parallel PARALLEL_ALTERNATIVES(7)

parallel_alternatives - Alternatives to GNU parallel

There are a lot programs with some of the functionality of GNU parallel. GNU parallel strives to include the best of the functionality without sacrificing ease of use.

parallel has existed since 2002 and as GNU parallel since 2010. A lot of the alternatives have not had the vitality to survive that long, but have come and gone during that time.

GNU parallel is actively maintained with a new release every month since 2010. Most other alternatives are fleeting interests of the developers with irregular releases and only maintained for a few years.

The following features are in some of the comparable tools:

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. Arguments can be read from the same file using #! (shebang) I7. 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 the complete command line

Outputs O1. Grouping output so output from different jobs do not mix O2. Send stderr (standard error) to stderr (standard error) O3. Send stdout (standard output) to stdout (standard output) O4. Order of output can be same as order of input O5. Stdout only contains stdout (standard output) from the command O6. Stderr only contains stderr (standard error) from the command O7. Buffering on disk O8. Cleanup of file if killed O9. Test if disk runs full during run

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 E6. Number of running jobs can be adjusted while running E7. Only spawn new jobs if load is less than a limit

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 MaxStartups can handle R8. Configurable SSH command R9. Retry if connection breaks occasionally

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 (See REPORTING BUGS).

parallel: I1 I2 I3 I4 I5 I6 I7 M1 M2 M3 M4 M5 M6 O1 O2 O3 O4 O5 O6 O7 O8 O9 E1 E2 E3 E4 E5 E6 E7 R1 R2 R3 R4 R5 R6 R7 R8 R9 S1 S2

find -exec: - - - x - x - - M2 M3 - - - - - O2 O3 O4 O5 O6 - - - - - - - - - - - - - - - - x x

make -j: - - - - - - - - - - - - - O1 O2 O3 - x O6 E1 - - - E5 - - - - - - - - - - - -

xjobs, prll, dxargs, mdm/middelman, xapply, paexec, ladon, jobflow, ClusterSSH: TODO - Please file a bug-report if you know what features they support (See REPORTING BUGS).

Summary table (see legend above): I1 I2 - - - - - - M2 M3 - - - - O2 O3 - O5 O6 E1 - - - - - - - - - - - x - - - - -

xargs offers some of the same possibilities 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 "My brother's 12\" records"
ls | xargs rmdir
touch 'c:\windows\system32\clfs.sys'
echo 'c:\windows\system32\clfs.sys' | xargs ls -l

You can specify -0, but many input generators are not optimized for using NUL as separator but are optimized for newline as separator. E.g. awk, ls, echo, tar -v, head (requires using -z), tail (requires using -z), sed (requires using -z), perl (-0 and \0 instead of \n), locate (requires using -0), find (requires using -print0), grep (requires using -z or -Z), sort (requires using -z).

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. To see this in action try:

parallel perl -e '\$a=\"1\".\"{}\"x10000000\;print\ \$a,\"\\n\"' \
  '>' {} ::: a b c d e f g h
# Serial = no mixing = the wanted result
# 'tr -s a-z' squeezes repeating letters into a single letter
echo a b c d e f g h | xargs -P1 -n1 grep 1 | tr -s a-z
# Compare to 8 jobs in parallel
parallel -kP8 -n1 grep 1 ::: a b c d e f g h | tr -s a-z
echo a b c d e f g h | xargs -P8 -n1 grep 1 | tr -s a-z
echo a b c d e f g h | xargs -P8 -n1 grep --line-buffered 1 | \
  tr -s a-z

Or try this:

slow_seq() {
  echo Count to "$@"
  seq "$@" |
    perl -ne '$|=1; for(split//){ print; select($a,$a,$a,0.100);}'
}
export -f slow_seq
# Serial = no mixing = the wanted result
seq 8 | xargs -n1 -P1 -I {} bash -c 'slow_seq {}'
# Compare to 8 jobs in parallel
seq 8 | parallel -P8 slow_seq {}
seq 8 | xargs -n1 -P8 -I {} bash -c 'slow_seq {}'

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"
ls | parallel "echo {}; ls {}|wc"

becomes (assuming you have 8 cores and that none of the filenames contain space, " or ').

ls | xargs -d "\n" -P8 -I {} bash -c "wc {} >{}.wc"
ls | xargs -d "\n" -P8 -I {} bash -c "echo {}; ls {}|wc"

https://www.gnu.org/software/findutils/

find -exec offers some of the same possibilities as GNU parallel.

find -exec only works on files. 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.

https://www.gnu.org/software/findutils/ (Last checked: 2019-01)

make -j can run jobs in parallel, but requires a crafted Makefile to do this. That results in extra quoting to get filenames containing newlines to work correctly.

make -j computes a dependency graph before running jobs. Jobs run by GNU parallel does not depend on each other.

(Very early versions of GNU parallel were coincidentally implemented using make -j).

https://www.gnu.org/software/make/ (Last checked: 2019-01)

Summary table (see legend above): I1 I2 - - - - I7 M1 - M3 - - M6 O1 - - x - - E1 E2 ?E3 E4 - - - R1 R2 R3 R4 - - ?R7 ? ? - -

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. Filenames containing newlines are not processed correctly. When reading input from a file null cannot be used as a 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.

EXAMPLES FROM ppss MANUAL

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 gzip

2 ./ppss.sh standalone -d /path/to/files -c 'cp "$ITEM" /destination/dir '

2 find /path/to/files -type f | parallel 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

https://github.com/louwrentius/PPSS

Summary table (see legend above): I1 I2 - I4 I5 - - M1 - M3 - - M6 O1 O2 O3 - O5 O6 E1 - - E4 - E6 - R1 - - - - R6 - - - S1 -

pexec is also a tool for running jobs in parallel.

EXAMPLES FROM pexec MANUAL

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 \ '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_{}'

https://www.gnu.org/software/pexec/

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

http://www.maier-komor.de/xjobs.html (Last checked: 2019-01)

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 supports scripts directly, functions if they are exported using export -f, and aliases if using env_parallel.

prll generates a lot of status information on stderr (standard error) which makes it harder to use the stderr (standard error) 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

https://github.com/exzombie/prll (Last checked: 2019-01)

dxargs is also a tool for running jobs in parallel.

dxargs does not deal well with more simultaneous jobs than SSHD's MaxStartups. dxargs is only built for remote run jobs, but does not support transferring of files.

https://web.archive.org/web/20120518070250/http://www. semicomplete.com/blog/geekery/distributed-xargs.html (Last checked: 2019-01)

middleman(mdm) is also a tool for running jobs in parallel.

Here are the shellscripts of https://web.archive.org/web/20110728064735/http://mdm. berlios.de/usage.html ported to GNU parallel:

seq 19 | parallel buffon -o - | sort -n > result
cat files | parallel cmd
find dir -execdir sem cmd {} \;

https://github.com/cklin/mdm (Last checked: 2019-01)

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 --link 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)

6 find */ -... | parallel -s 1024 -Xuj1 vi

7 find ... | xapply -f -5 -i /dev/tty 'vi' - - - - -

7 sh <(find ... |parallel -n5 echo vi)

7 find ... |parallel -n5 -uj1 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 | ...

https://web.archive.org/web/20160702211113/ http://carrera.databits.net/~ksb/msrc/local/bin/xapply/xapply.html

apply can build command lines based on a template and arguments - very much like GNU parallel. apply does not run jobs in parallel. apply does not use an argument separator (like :::); instead the template must be the first argument.

Here are the examples from IBM's Knowledge Center and the corresponding command using GNU parallel:

1. To obtain results similar to those of the ls command, enter:

apply echo *
parallel echo ::: *

2. To compare the file named a1 to the file named b1, and the file named a2 to the file named b2, enter:

apply -2 cmp a1 b1 a2 b2
parallel -N2 cmp ::: a1 b1 a2 b2

3. To run the who command five times, enter:

apply -0 who 1 2 3 4 5
parallel -N0 who ::: 1 2 3 4 5

4. To link all files in the current directory to the directory /usr/joe, enter:

apply 'ln %1 /usr/joe' *
parallel ln {} /usr/joe ::: *

https://www-01.ibm.com/support/knowledgecenter/ ssw_aix_71/com.ibm.aix.cmds1/apply.htm (Last checked: 2019-01)

paexec can run jobs in parallel on both the local and remote computers.

paexec requires commands to print a blank line as the last output. This means you will have to write a wrapper for most programs.

paexec has a job dependency facility so a job can depend on another job to be executed successfully. Sort of a poor-man's make.

Here are the examples from paexec's example catalog with the equivalent using GNU parallel:

1_div_X_run:
../../paexec -s -l -c "`pwd`/1_div_X_cmd" -n +1 <<EOF [...]
parallel echo {} '|' `pwd`/1_div_X_cmd <<EOF [...]
all_substr_run:
../../paexec -lp -c "`pwd`/all_substr_cmd" -n +3 <<EOF [...]
parallel echo {} '|' `pwd`/all_substr_cmd <<EOF [...]
cc_wrapper_run:
../../paexec -c "env CC=gcc CFLAGS=-O2 `pwd`/cc_wrapper_cmd" \
           -n 'host1 host2' \
           -t '/usr/bin/ssh -x' <<EOF [...]
parallel echo {} '|' "env CC=gcc CFLAGS=-O2 `pwd`/cc_wrapper_cmd" \
           -S host1,host2 <<EOF [...]
# This is not exactly the same, but avoids the wrapper
parallel gcc -O2 -c -o {.}.o {} \
           -S host1,host2 <<EOF [...]
toupper_run:
../../paexec -lp -c "`pwd`/toupper_cmd" -n +10 <<EOF [...]
parallel echo {} '|' ./toupper_cmd <<EOF [...]
# Without the wrapper:
parallel echo {} '| awk {print\ toupper\(\$0\)}' <<EOF [...]

https://github.com/cheusov/paexec

map sees it as a feature to have less features and in doing so it also handles corner cases incorrectly. A lot of GNU parallel's code is to handle corner cases correctly on every platform, so you will not get a nasty surprise if a user, for example, saves a file called: My brother's 12" records.txt

map's example showing how to deal with special characters fails on special characters:

echo "The Cure" > My\ brother\'s\ 12\"\ records

ls | \
  map 'echo -n `gzip < "%" | wc -c`; echo -n '*100/'; wc -c < "%"' |
  bc

It works with GNU parallel:

ls | \
  parallel \
    'echo -n `gzip < {} | wc -c`; echo -n '*100/'; wc -c < {}' | bc

And you can even get the file name prepended:

ls | \
  parallel --tag \
    '(echo -n `gzip < {} | wc -c`'*100/'; wc -c < {}) | bc'

map has no support for grouping. So this gives the wrong results without any warnings:

parallel perl -e '\$a=\"1{}\"x10000000\;print\ \$a,\"\\n\"' '>' {} \
  ::: a b c d e f
ls -l a b c d e f
parallel -kP4 -n1 grep 1 > out.par ::: a b c d e f
map -p 4 'grep 1' a b c d e f > out.map-unbuf
map -p 4 'grep --line-buffered 1' a b c d e f > out.map-linebuf
map -p 1 'grep --line-buffered 1' a b c d e f > out.map-serial
ls -l out*
md5sum out*

The documentation shows a workaround, but not only does that mix stdout (standard output) with stderr (standard error) it also fails completely for certain jobs (and may even be considered less readable):

parallel echo -n {} ::: 1 2 3

map -p 4 'echo -n % 2>&1 | sed -e "s/^/$$:/"' 1 2 3 | \
  sort | cut -f2- -d:

maps replacement strings (% %D %B %E) can be simulated in GNU parallel by putting this in ~/.parallel/config:

--rpl '%'
--rpl '%D $_=Q(::dirname($_));'
--rpl '%B s:.*/::;s:\.[^/.]+$::;'
--rpl '%E s:.*\.::'

map does not have an argument separator on the command line, but uses the first argument as command. This makes quoting harder which again may affect readability. Compare:

map -p 2 'perl -ne '"'"'/^\S+\s+\S+$/ and print $ARGV,"\n"'"'" *

parallel -q perl -ne '/^\S+\s+\S+$/ and print $ARGV,"\n"' ::: *

map can do multiple arguments with context replace, but not without context replace:

parallel --xargs echo 'BEGIN{'{}'}END' ::: 1 2 3

map "echo 'BEGIN{'%'}END'" 1 2 3

map requires Perl v5.10.0 making it harder to use on old systems.

map has no way of using % in the command (GNU parallel has -I to specify another replacement string than {}).

By design map is option incompatible with xargs, it does not have remote job execution, a structured way of saving results, multiple input sources, progress indicator, configurable record delimiter (only field delimiter), logging of jobs run with possibility to resume, keeping the output in the same order as input, --pipe processing, and dynamically timeouts.

https://github.com/sitaramc/map

ladon can run multiple jobs on files in parallel.

ladon only works on files and the only way to specify files is using a quoted glob string (such as \*.jpg). It is not possible to list the files manually.

As replacement strings it uses FULLPATH DIRNAME BASENAME EXT RELDIR RELPATH

These can be simulated using GNU parallel by putting this in ~/.parallel/config:

--rpl 'FULLPATH $_=Q($_);chomp($_=qx{readlink -f $_});'
--rpl 'DIRNAME $_=Q(::dirname($_));chomp($_=qx{readlink -f $_});'
--rpl 'BASENAME s:.*/::;s:\.[^/.]+$::;'
--rpl 'EXT s:.*\.::'
--rpl 'RELDIR $_=Q($_);chomp(($_,$c)=qx{readlink -f $_;pwd});
       s:\Q$c/\E::;$_=::dirname($_);'
--rpl 'RELPATH $_=Q($_);chomp(($_,$c)=qx{readlink -f $_;pwd});
       s:\Q$c/\E::;'

ladon deals badly with filenames containing " and newline, and it fails for output larger than 200k:

ladon '*' -- seq 36000 | wc

EXAMPLES FROM ladon MANUAL

It is assumed that the '--rpl's above are put in ~/.parallel/config and that it is run under a shell that supports '**' globbing (such as zsh):

1 ladon "**/*.txt" -- echo RELPATH

1 parallel echo RELPATH ::: **/*.txt

2 ladon "~/Documents/**/*.pdf" -- shasum FULLPATH >hashes.txt

2 parallel shasum FULLPATH ::: ~/Documents/**/*.pdf >hashes.txt

3 ladon -m thumbs/RELDIR "**/*.jpg" -- convert FULLPATH -thumbnail 100x100^ -gravity center -extent 100x100 thumbs/RELPATH

3 parallel mkdir -p thumbs/RELDIR\; convert FULLPATH -thumbnail 100x100^ -gravity center -extent 100x100 thumbs/RELPATH ::: **/*.jpg

4 ladon "~/Music/*.wav" -- lame -V 2 FULLPATH DIRNAME/BASENAME.mp3

4 parallel lame -V 2 FULLPATH DIRNAME/BASENAME.mp3 ::: ~/Music/*.wav

https://github.com/danielgtaylor/ladon (Last checked: 2019-01)

jobflow can run multiple jobs in parallel.

Just like xargs output from jobflow jobs running in parallel mix together by default. jobflow can buffer into files (placed in /run/shm), but these are not cleaned up if jobflow dies unexpectedly (e.g. by Ctrl-C). If the total output is big (in the order of RAM+swap) it can cause the system to slow to a crawl and eventually run out of memory.

jobflow gives no error if the command is unknown, and like xargs redirection and composed commands require wrapping with bash -c.

Input lines can at most be 4096 bytes. You can at most have 16 {}'s in the command template. More than that either crashes the program or simple does not execute the command.

jobflow has no equivalent for --pipe, or --sshlogin.

jobflow makes it possible to set resource limits on the running jobs. This can be emulated by GNU parallel using bash's ulimit:

jobflow -limits=mem=100M,cpu=3,fsize=20M,nofiles=300 myjob

parallel 'ulimit -v 102400 -t 3 -f 204800 -n 300 myjob'

EXAMPLES FROM jobflow README

1 cat things.list | jobflow -threads=8 -exec ./mytask {}

1 cat things.list | parallel -j8 ./mytask {}

2 seq 100 | jobflow -threads=100 -exec echo {}

2 seq 100 | parallel -j100 echo {}

3 cat urls.txt | jobflow -threads=32 -exec wget {}

3 cat urls.txt | parallel -j32 wget {}

4 find . -name '*.bmp' | jobflow -threads=8 -exec bmp2jpeg {.}.bmp {.}.jpg

4 find . -name '*.bmp' | parallel -j8 bmp2jpeg {.}.bmp {.}.jpg

https://github.com/rofl0r/jobflow

gargs can run multiple jobs in parallel.

Older versions cache output in memory. This causes it to be extremely slow when the output is larger than the physical RAM, and can cause the system to run out of memory.

See more details on this in man parallel_design.

Newer versions cache output in files, but leave files in $TMPDIR if it is killed.

Output to stderr (standard error) is changed if the command fails.

Here are the two examples from gargs website.

1 seq 12 -1 1 | gargs -p 4 -n 3 "sleep {0}; echo {1} {2}"

1 seq 12 -1 1 | parallel -P 4 -n 3 "sleep {1}; echo {2} {3}"

2 cat t.txt | gargs --sep "\s+" -p 2 "echo '{0}:{1}-{2}' full-line: \'{}\'"

2 cat t.txt | parallel --colsep "\\s+" -P 2 "echo '{1}:{2}-{3}' full-line: \'{}\'"

https://github.com/brentp/gargs

orgalorg can run the same job on multiple machines. This is related to --onall and --nonall.

orgalorg supports entering the SSH password - provided it is the same for all servers. GNU parallel advocates using ssh-agent instead, but it is possible to emulate orgalorg's behavior by setting SSHPASS and by using --ssh "sshpass ssh".

To make the emulation easier, make a simple alias:

alias par_emul="parallel -j0 --ssh 'sshpass ssh' --nonall --tag --lb"

If you want to supply a password run:

SSHPASS=`ssh-askpass`

or set the password directly:

SSHPASS=P4$$w0rd!

If the above is set up you can then do:

orgalorg -o frontend1 -o frontend2 -p -C uptime
par_emul -S frontend1 -S frontend2 uptime

orgalorg -o frontend1 -o frontend2 -p -C top -bid 1
par_emul -S frontend1 -S frontend2 top -bid 1

orgalorg -o frontend1 -o frontend2 -p -er /tmp -n \
  'md5sum /tmp/bigfile' -S bigfile
par_emul -S frontend1 -S frontend2 --basefile bigfile \
  --workdir /tmp md5sum /tmp/bigfile

orgalorg has a progress indicator for the transferring of a file. GNU parallel does not.

https://github.com/reconquest/orgalorg

Rust parallel focuses on speed. It is almost as fast as xargs. It implements a few features from GNU parallel, but lacks many functions. All these fail:

# Read arguments from file
parallel -a file echo
# Changing the delimiter
parallel -d _ echo ::: a_b_c_

These do something different from GNU parallel

# -q to protect quoted $ and space
parallel -q perl -e '$a=shift; print "$a"x10000000' ::: a b c
# Generation of combination of inputs
parallel echo {1} {2} ::: red green blue ::: S M L XL XXL
# {= perl expression =} replacement string
parallel echo '{= s/new/old/ =}' ::: my.new your.new
# --pipe
seq 100000 | parallel --pipe wc
# linked arguments
parallel echo ::: S M L :::+ sml med lrg ::: R G B :::+ red grn blu
# Run different shell dialects
zsh -c 'parallel echo \={} ::: zsh && true'
csh -c 'parallel echo \$\{\} ::: shell && true'
bash -c 'parallel echo \$\({}\) ::: pwd && true'
# Rust parallel does not start before the last argument is read
(seq 10; sleep 5; echo 2) | time parallel -j2 'sleep 2; echo'
tail -f /var/log/syslog | parallel echo

Most of the examples from the book GNU Parallel 2018 do not work, thus Rust parallel is not close to being a compatible replacement.

Rust parallel has no remote facilities.

It uses /tmp/parallel for tmp files and does not clean up if terminated abruptly. If another user on the system uses Rust parallel, then /tmp/parallel will have the wrong permissions and Rust parallel will fail. A malicious user can setup the right permissions and symlink the output file to one of the user's files and next time the user uses Rust parallel it will overwrite this file.

attacker$ mkdir /tmp/parallel
attacker$ chmod a+rwX /tmp/parallel
# Symlink to the file the attacker wants to zero out
attacker$ ln -s ~victim/.important-file /tmp/parallel/stderr_1
victim$ seq 1000 | parallel echo
# This file is now overwritten with stderr from 'echo'
victim$ cat ~victim/.important-file

If /tmp/parallel runs full during the run, Rust parallel does not report this, but finishes with success - thereby risking data loss.

https://github.com/mmstick/parallel

rush (https://github.com/shenwei356/rush) is written in Go and based on gargs.

Just like GNU parallel rush buffers in temporary files. But opposite GNU parallel rush does not clean up, if the process dies abnormally.

rush has some string manipulations that can be emulated by putting this into ~/.parallel/config (/ is used instead of %, and % is used instead of ^ as that is closer to bash's ${var%postfix}):

--rpl '{:} s:(\.[^/]+)*$::'
--rpl '{:%([^}]+?)} s:$$1(\.[^/]+)*$::'
--rpl '{/:%([^}]*?)} s:.*/(.*)$$1(\.[^/]+)*$:$1:'
--rpl '{/:} s:(.*/)?([^/.]+)(\.[^/]+)*$:$2:'
--rpl '{@(.*?)} /$$1/ and $_=$1;'

Here are the examples from rush's website with the equivalent command in GNU parallel.

EXAMPLES

1. Simple run, quoting is not necessary

$ seq 1 3 | rush echo {}

$ seq 1 3 | parallel echo {}

2. Read data from file (`-i`)

$ rush echo {} -i data1.txt -i data2.txt

$ cat data1.txt data2.txt | parallel echo {}

3. Keep output order (`-k`)

$ seq 1 3 | rush 'echo {}' -k

$ seq 1 3 | parallel -k echo {}

4. Timeout (`-t`)

$ time seq 1 | rush 'sleep 2; echo {}' -t 1

$ time seq 1 | parallel --timeout 1 'sleep 2; echo {}'

5. Retry (`-r`)

$ seq 1 | rush 'python unexisted_script.py' -r 1

$ seq 1 | parallel --retries 2 'python unexisted_script.py'

Use -u to see it is really run twice:

$ seq 1 | parallel -u --retries 2 'python unexisted_script.py'

6. Dirname (`{/}`) and basename (`{%}`) and remove custom suffix (`{^suffix}`)

$ echo dir/file_1.txt.gz | rush 'echo {/} {%} {^_1.txt.gz}'

$ echo dir/file_1.txt.gz |
    parallel --plus echo {//} {/} {%_1.txt.gz}

7. Get basename, and remove last (`{.}`) or any (`{:}`) extension

$ echo dir.d/file.txt.gz | rush 'echo {.} {:} {%.} {%:}'

$ echo dir.d/file.txt.gz | parallel 'echo {.} {:} {/.} {/:}'

8. Job ID, combine fields index and other replacement strings

$ echo 12 file.txt dir/s_1.fq.gz |
    rush 'echo job {#}: {2} {2.} {3%:^_1}'

$ echo 12 file.txt dir/s_1.fq.gz |
    parallel --colsep ' ' 'echo job {#}: {2} {2.} {3/:%_1}'

9. Capture submatch using regular expression (`{@regexp}`)

$ echo read_1.fq.gz | rush 'echo {@(.+)_\d}'

$ echo read_1.fq.gz | parallel 'echo {@(.+)_\d}'

10. Custom field delimiter (`-d`)

$ echo a=b=c | rush 'echo {1} {2} {3}' -d =

$ echo a=b=c | parallel -d = echo {1} {2} {3}

11. Send multi-lines to every command (`-n`)

$ seq 5 | rush -n 2 -k 'echo "{}"; echo'

$ seq 5 |
    parallel -n 2 -k \
      'echo {=-1 $_=join"\n",@arg[1..$#arg] =}; echo'

$ seq 5 | rush -n 2 -k 'echo "{}"; echo' -J ' '

$ seq 5 | parallel -n 2 -k 'echo {}; echo'

12. Custom record delimiter (`-D`), note that empty records are not used.

$ echo a b c d | rush -D " " -k 'echo {}'

$ echo a b c d | parallel -d " " -k 'echo {}'

$ echo abcd | rush -D "" -k 'echo {}'

Cannot be done by GNU Parallel

$ cat fasta.fa
>seq1
tag
>seq2
cat
gat
>seq3
attac
a
cat

$ cat fasta.fa | rush -D ">" \
    'echo FASTA record {#}: name: {1} sequence: {2}' -k -d "\n"
# rush fails to join the multiline sequences

$ cat fasta.fa | (read -n1 ignore_first_char;
    parallel -d '>' --colsep '\n' echo FASTA record {#}: \
      name: {1} sequence: '{=2 $_=join"",@arg[2..$#arg]=}'
  )

13. Assign value to variable, like `awk -v` (`-v`)

$ seq 1 |
    rush 'echo Hello, {fname} {lname}!' -v fname=Wei -v lname=Shen

$ seq 1 |
    parallel -N0 \
      'fname=Wei; lname=Shen; echo Hello, ${fname} ${lname}!'

$ for var in a b; do \
$   seq 1 3 | rush -k -v var=$var 'echo var: {var}, data: {}'; \
$ done

In GNU parallel you would typically do:

$ seq 1 3 | parallel -k echo var: {1}, data: {2} ::: a b :::: -

If you really want the var:

$ seq 1 3 |
    parallel -k var={1} ';echo var: $var, data: {}' ::: a b :::: -

If you really want the for-loop:

$ for var in a b; do
>   export var;
>   seq 1 3 | parallel -k 'echo var: $var, data: {}';
> done

Contrary to rush this also works if the value is complex like:

My brother's 12" records

14. Preset variable (`-v`), avoid repeatedly writing verbose replacement strings

# naive way
$ echo read_1.fq.gz | rush 'echo {:^_1} {:^_1}_2.fq.gz'

$ echo read_1.fq.gz | parallel 'echo {:%_1} {:%_1}_2.fq.gz'

# macro + removing suffix
$ echo read_1.fq.gz |
    rush -v p='{:^_1}' 'echo {p} {p}_2.fq.gz'

$ echo read_1.fq.gz |
    parallel 'p={:%_1}; echo $p ${p}_2.fq.gz'

# macro + regular expression
$ echo read_1.fq.gz | rush -v p='{@(.+?)_\d}' 'echo {p} {p}_2.fq.gz'

$ echo read_1.fq.gz | parallel 'p={@(.+?)_\d}; echo $p ${p}_2.fq.gz'

Contrary to rush GNU parallel works with complex values:

echo "My brother's 12\"read_1.fq.gz" |
  parallel 'p={@(.+?)_\d}; echo $p ${p}_2.fq.gz'

15. Interrupt jobs by `Ctrl-C`, rush will stop unfinished commands and exit.

$ seq 1 20 | rush 'sleep 1; echo {}'
^C

$ seq 1 20 | parallel 'sleep 1; echo {}'
^C

16. Continue/resume jobs (`-c`). When some jobs failed (by execution failure, timeout, or canceling by user with `Ctrl + C`), please switch flag `-c/--continue` on and run again, so that `rush` can save successful commands and ignore them in NEXT run.

$ seq 1 3 | rush 'sleep {}; echo {}' -t 3 -c
$ cat successful_cmds.rush
$ seq 1 3 | rush 'sleep {}; echo {}' -t 3 -c

$ seq 1 3 | parallel --joblog mylog --timeout 2 \
    'sleep {}; echo {}'
$ cat mylog
$ seq 1 3 | parallel --joblog mylog --retry-failed \
    'sleep {}; echo {}'

Multi-line jobs:

$ seq 1 3 | rush 'sleep {}; echo {}; \
  echo finish {}' -t 3 -c -C finished.rush
$ cat finished.rush
$ seq 1 3 | rush 'sleep {}; echo {}; \
  echo finish {}' -t 3 -c -C finished.rush

$ seq 1 3 |
    parallel --joblog mylog --timeout 2 'sleep {}; echo {}; \
  echo finish {}'
$ cat mylog
$ seq 1 3 |
    parallel --joblog mylog --retry-failed 'sleep {}; echo {}; \
      echo finish {}'

17. A comprehensive example: downloading 1K+ pages given by three URL list files using `phantomjs save_page.js` (some page contents are dynamically generated by Javascript, so `wget` does not work). Here I set max jobs number (`-j`) as `20`, each job has a max running time (`-t`) of `60` seconds and `3` retry changes (`-r`). Continue flag `-c` is also switched on, so we can continue unfinished jobs. Luckily, it's accomplished in one run :)

$ for f in $(seq 2014 2016); do \
$    /bin/rm -rf $f; mkdir -p $f; \
$    cat $f.html.txt | rush -v d=$f -d = \
       'phantomjs save_page.js "{}" > {d}/{3}.html' \
       -j 20 -t 60 -r 3 -c; \
$ done

GNU parallel can append to an existing joblog with '+':

$ rm mylog
$ for f in $(seq 2014 2016); do
    /bin/rm -rf $f; mkdir -p $f;
    cat $f.html.txt |
      parallel -j20 --timeout 60 --retries 4 --joblog +mylog \
        --colsep = \
        phantomjs save_page.js {1}={2}={3} '>' $f/{3}.html
  done

18. A bioinformatics example: mapping with `bwa`, and processing result with `samtools`:

$ ref=ref/xxx.fa
$ threads=25
$ ls -d raw.cluster.clean.mapping/* \
  | rush -v ref=$ref -v j=$threads -v p='{}/{%}' \
      'bwa mem -t {j} -M -a {ref} {p}_1.fq.gz {p}_2.fq.gz >{p}.sam;\
      samtools view -bS {p}.sam > {p}.bam; \
      samtools sort -T {p}.tmp -@ {j} {p}.bam -o {p}.sorted.bam; \
      samtools index {p}.sorted.bam; \
      samtools flagstat {p}.sorted.bam > {p}.sorted.bam.flagstat; \
      /bin/rm {p}.bam {p}.sam;' \
      -j 2 --verbose -c -C mapping.rush

GNU parallel would use a function:

$ ref=ref/xxx.fa
$ export ref
$ thr=25
$ export thr
$ bwa_sam() {
    p="$1"
    bam="$p".bam
    sam="$p".sam
    sortbam="$p".sorted.bam
    bwa mem -t $thr -M -a $ref ${p}_1.fq.gz ${p}_2.fq.gz > "$sam"
    samtools view -bS "$sam" > "$bam"
    samtools sort -T ${p}.tmp -@ $thr "$bam" -o "$sortbam"
    samtools index "$sortbam"
    samtools flagstat "$sortbam" > "$sortbam".flagstat
    /bin/rm "$bam" "$sam"
  }
$ export -f bwa_sam
$ ls -d raw.cluster.clean.mapping/* |
    parallel -j 2 --verbose --joblog mylog bwa_sam

Other rush features

rush has:

  • awk -v like custom defined variables (-v)

    With GNU parallel you would simply set a shell variable:

    parallel 'v={}; echo "$v"' ::: foo
    echo foo | rush -v v={} 'echo {v}'
    

    Also rush does not like special chars. So these do not work:

    echo does not work | rush -v v=\" 'echo {v}'
    echo "My  brother's  12\"  records" | rush -v v={} 'echo {v}'
    

    Whereas the corresponding GNU parallel version works:

    parallel 'v=\"; echo "$v"' ::: works
    parallel 'v={}; echo "$v"' ::: "My  brother's  12\"  records"
    
  • Exit on first error(s) (-e)

    This is called --halt now,fail=1 (or shorter: --halt 2) when used with GNU parallel.

  • Settable records sending to every command (-n, default 1)

    This is also called -n in GNU parallel.

  • Practical replacement strings
{:} remove any extension
With GNU parallel this can be emulated by:

parallel --plus echo '{/\..*/}' ::: foo.ext.bar.gz
{^suffix}, remove suffix
With GNU parallel this can be emulated by:

parallel --plus echo '{%.bar.gz}' ::: foo.ext.bar.gz
{@regexp}, capture submatch using regular expression
With GNU parallel this can be emulated by:

parallel --rpl '{@(.*?)} /$$1/ and $_=$1;' \
  echo '{@\d_(.*).gz}' ::: 1_foo.gz
{%.}, {%:}, basename without extension
With GNU parallel this can be emulated by:

parallel echo '{= s:.*/::;s/\..*// =}' ::: dir/foo.bar.gz

And if you need it often, you define a --rpl in $HOME/.parallel/config:

--rpl '{%.} s:.*/::;s/\..*//'
--rpl '{%:} s:.*/::;s/\..*//'

Then you can use them as:

parallel echo {%.} {%:} ::: dir/foo.bar.gz
  • Preset variable (macro)

    E.g.

    echo foosuffix | rush -v p={^suffix} 'echo {p}_new_suffix'
    

    With GNU parallel this can be emulated by:

    echo foosuffix |
      parallel --plus 'p={%suffix}; echo ${p}_new_suffix'
    

    Opposite rush GNU parallel works fine if the input contains double space, ' and ":

    echo "1'6\"  foosuffix" |
      parallel --plus 'p={%suffix}; echo "${p}"_new_suffix'
    
  • Commands of multi-lines

    While you can use multi-lined commands in GNU parallel, to improve readability GNU parallel discourages the use of multi-line commands. In most cases it can be written as a function:

    seq 1 3 |
      parallel --timeout 2 --joblog my.log 'sleep {}; echo {}; \
        echo finish {}'
    

    Could be written as:

    doit() {
      sleep "$1"
      echo "$1"
      echo finish "$1"
    }
    export -f doit
    seq 1 3 | parallel --timeout 2 --joblog my.log doit
    

    The failed commands can be resumed with:

    seq 1 3 |
      parallel --resume-failed --joblog my.log 'sleep {}; echo {};\
        echo finish {}'
    

https://github.com/shenwei356/rush

ClusterSSH solves a different problem than GNU parallel.

ClusterSSH opens a terminal window for each computer and using a master window you can run the same command on all the computers. This is typically used for administrating several computers that are almost identical.

GNU parallel runs the same (or different) commands with different arguments in parallel possibly using remote computers to help computing. If more than one computer is listed in -S GNU parallel may only use one of these (e.g. if there are 8 jobs to be run and one computer has 8 cores).

GNU parallel can be used as a poor-man's version of ClusterSSH:

parallel --nonall -S server-a,server-b do_stuff foo bar

https://github.com/duncs/clusterssh

coshell only accepts full commands on standard input. Any quoting needs to be done by the user.

Commands are run in sh so any bash/tcsh/zsh specific syntax will not work.

Output can be buffered by using -d. Output is buffered in memory, so big output can cause swapping and therefore be terrible slow or even cause out of memory.

https://github.com/gdm85/coshell (Last checked: 2019-01)

spread runs commands on all directories.

It can be emulated with GNU parallel using this Bash function:

spread() {
  _cmds() {
    perl -e '$"=" && ";print "@ARGV"' "cd {}" "$@"
  }
  parallel $(_cmds "$@")'|| echo exit status $?' ::: */
}

This works except for the --exclude option.

(Last checked: 2017-11)

pyargs deals badly with input containing spaces. It buffers stdout, but not stderr. It buffers in RAM. {} does not work as replacement string. It does not support running functions.

pyargs does not support composed commands if run with --lines, and fails on pyargs traceroute gnu.org fsf.org.

Examples

seq 5 | pyargs -P50 -L seq
seq 5 | parallel -P50 --lb seq

seq 5 | pyargs -P50 --mark -L seq
seq 5 | parallel -P50 --lb \
  --tagstring OUTPUT'[{= $_=$job->replaced()=}]' seq
# Similar, but not precisely the same
seq 5 | parallel -P50 --lb --tag seq

seq 5 | pyargs -P50  --mark command
# Somewhat longer with GNU Parallel due to the special
#   --mark formatting
cmd="$(echo "command" | parallel --shellquote)"
wrap_cmd() {
   echo "MARK $cmd $@================================" >&3
   echo "OUTPUT START[$cmd $@]:"
   eval $cmd "$@"
   echo "OUTPUT END[$cmd $@]"
}
(seq 5 | env_parallel -P2 wrap_cmd) 3>&1
# Similar, but not exactly the same
seq 5 | parallel -t --tag command

(echo '1  2  3';echo 4 5 6) | pyargs  --stream seq
(echo '1  2  3';echo 4 5 6) | perl -pe 's/\n/ /' |
  parallel -r -d' ' seq
# Similar, but not exactly the same
parallel seq ::: 1 2 3 4 5 6

https://github.com/robertblackwell/pyargs (Last checked: 2019-01)

concurrently runs jobs in parallel.

The output is prepended with the job number, and may be incomplete:

$ concurrently 'seq 100000' | (sleep 3;wc -l)
7165

When pretty printing it caches output in memory. Output mixes by using test MIX below whether or not output is cached.

There seems to be no way of making a template command and have concurrently fill that with different args. The full commands must be given on the command line.

There is also no way of controlling how many jobs should be run in parallel at a time - i.e. "number of jobslots". Instead all jobs are simply started in parallel.

https://github.com/kimmobrunfeldt/concurrently (Last checked: 2019-01)

map does not run jobs in parallel by default. The README suggests using:

... | map t 'sleep $t && say done &'

But this fails if more jobs are run in parallel than the number of available processes. Since there is no support for parallelization in map itself, the output also mixes:

seq 10 | map i 'echo start-$i && sleep 0.$i && echo end-$i &'

The major difference is that GNU parallel is built for parallelization and map is not. So GNU parallel has lots of ways of dealing with the issues that parallelization raises:

  • Keep the number of processes manageable
  • Make sure output does not mix
  • Make Ctrl-C kill all running processes

Here are the 5 examples converted to GNU Parallel:

1$ ls *.c | map f 'foo $f'
1$ ls *.c | parallel foo

2$ ls *.c | map f 'foo $f; bar $f'
2$ ls *.c | parallel 'foo {}; bar {}'

3$ cat urls | map u 'curl -O $u'
3$ cat urls | parallel curl -O

4$ printf "1\n1\n1\n" | map t 'sleep $t && say done'
4$ printf "1\n1\n1\n" | parallel 'sleep {} && say done'
4$ parallel 'sleep {} && say done' ::: 1 1 1

5$ printf "1\n1\n1\n" | map t 'sleep $t && say done &'
5$ printf "1\n1\n1\n" | parallel -j0 'sleep {} && say done'
5$ parallel -j0 'sleep {} && say done' ::: 1 1 1

https://github.com/soveran/map (Last checked: 2019-01)

loop mixes stdout and stderr:

loop 'ls /no-such-file' >/dev/null

loop's replacement string $ITEM does not quote strings:

echo 'two  spaces' | loop 'echo $ITEM'

loop cannot run functions:

myfunc() { echo joe; }
export -f myfunc
loop 'myfunc this fails'

Some of the examples from https://github.com/Miserlou/Loop/ can be emulated with GNU parallel:

# A couple of functions will make the code easier to read
$ loopy() {
    yes | parallel -uN0 -j1 "$@"
  }
$ export -f loopy
$ time_out() {
    parallel -uN0 -q --timeout "$@" ::: 1
  }
$ match() {
    perl -0777 -ne 'grep /'"$1"'/,$_ and print or exit 1'
  }
$ export -f match

$ loop 'ls' --every 10s
$ loopy --delay 10s ls

$ loop 'touch $COUNT.txt' --count-by 5
$ loopy touch '{= $_=seq()*5 =}'.txt

$ loop --until-contains 200 -- \
    ./get_response_code.sh --site mysite.biz`
$ loopy --halt now,success=1 \
    './get_response_code.sh --site mysite.biz | match 200'

$ loop './poke_server' --for-duration 8h
$ time_out 8h loopy ./poke_server

$ loop './poke_server' --until-success
$ loopy --halt now,success=1 ./poke_server

$ cat files_to_create.txt | loop 'touch $ITEM'
$ cat files_to_create.txt | parallel touch {}

$ loop 'ls' --for-duration 10min --summary
# --joblog is somewhat more verbose than --summary
$ time_out 10m loopy --joblog my.log ./poke_server; cat my.log

$ loop 'echo hello'
$ loopy echo hello

$ loop 'echo $COUNT'
# GNU Parallel counts from 1
$ loopy echo {#}
# Counting from 0 can be forced
$ loopy echo '{= $_=seq()-1 =}'

$ loop 'echo $COUNT' --count-by 2
$ loopy echo '{= $_=2*(seq()-1) =}'

$ loop 'echo $COUNT' --count-by 2 --offset 10
$ loopy echo '{= $_=10+2*(seq()-1) =}'

$ loop 'echo $COUNT' --count-by 1.1
# GNU Parallel rounds 3.3000000000000003 to 3.3
$ loopy echo '{= $_=1.1*(seq()-1) =}'

$ loop 'echo $COUNT $ACTUALCOUNT' --count-by 2
$ loopy echo '{= $_=2*(seq()-1) =} {#}'

$ loop 'echo $COUNT' --num 3 --summary
# --joblog is somewhat more verbose than --summary
$ seq 3 | parallel --joblog my.log echo; cat my.log

$ loop 'ls -foobarbatz' --num 3 --summary
# --joblog is somewhat more verbose than --summary
$ seq 3 | parallel --joblog my.log -N0 ls -foobarbatz; cat my.log

$ loop 'echo $COUNT' --count-by 2 --num 50 --only-last
# Can be emulated by running 2 jobs
$ seq 49 | parallel echo '{= $_=2*(seq()-1) =}' >/dev/null
$ echo 50| parallel echo '{= $_=2*(seq()-1) =}'

$ loop 'date' --every 5s
$ loopy --delay 5s date

$ loop 'date' --for-duration 8s --every 2s
$ time_out 8s loopy --delay 2s date

$ loop 'date -u' --until-time '2018-05-25 20:50:00' --every 5s
$ seconds=$((`date -d 2019-05-25T20:50:00 +%s` - `date  +%s`))s
$ time_out $seconds loopy --delay 5s date -u

$ loop 'echo $RANDOM' --until-contains "666"
$ loopy --halt now,success=1 'echo $RANDOM | match 666'

$ loop 'if (( RANDOM % 2 )); then
          (echo "TRUE"; true);
        else
          (echo "FALSE"; false);
        fi' --until-success
$ loopy --halt now,success=1 'if (( $RANDOM % 2 )); then
                                (echo "TRUE"; true);
                              else
                                (echo "FALSE"; false);
                              fi'

$ loop 'if (( RANDOM % 2 )); then
    (echo "TRUE"; true);
  else
    (echo "FALSE"; false);
  fi' --until-error
$ loopy --halt now,fail=1 'if (( $RANDOM % 2 )); then
                             (echo "TRUE"; true);
                           else
                             (echo "FALSE"; false);
                           fi'

$ loop 'date' --until-match "(\d{4})"
$ loopy --halt now,success=1 'date | match [0-9][0-9][0-9][0-9]'

$ loop 'echo $ITEM' --for red,green,blue
$ parallel echo ::: red green blue

$ cat /tmp/my-list-of-files-to-create.txt | loop 'touch $ITEM'
$ cat /tmp/my-list-of-files-to-create.txt | parallel touch

$ ls | loop 'cp $ITEM $ITEM.bak'; ls
$ ls | parallel cp {} {}.bak; ls

$ loop 'echo $ITEM | tr a-z A-Z' -i
$ parallel 'echo {} | tr a-z A-Z'
# Or more efficiently:
$ parallel --pipe tr a-z A-Z

$ loop 'echo $ITEM' --for "`ls`"
$ parallel echo {} ::: "`ls`"

$ ls | loop './my_program $ITEM' --until-success;
$ ls | parallel --halt now,success=1 ./my_program {}

$ ls | loop './my_program $ITEM' --until-fail;
$ ls | parallel --halt now,fail=1 ./my_program {}

$ ./deploy.sh;
  loop 'curl -sw "%{http_code}" http://coolwebsite.biz' \
    --every 5s --until-contains 200;
  ./announce_to_slack.sh
$ ./deploy.sh;
  loopy --delay 5s --halt now,success=1 \
  'curl -sw "%{http_code}" http://coolwebsite.biz | match 200';
  ./announce_to_slack.sh

$ loop "ping -c 1 mysite.com" --until-success; ./do_next_thing
$ loopy --halt now,success=1 ping -c 1 mysite.com; ./do_next_thing

$ ./create_big_file -o my_big_file.bin;
  loop 'ls' --until-contains 'my_big_file.bin';
  ./upload_big_file my_big_file.bin
# inotifywait is a better tool to detect file system changes.
# It can even make sure the file is complete
# so you are not uploading an incomplete file
$ inotifywait -qmre MOVED_TO -e CLOSE_WRITE --format %w%f . |
    grep my_big_file.bin

$ ls | loop 'cp $ITEM $ITEM.bak'
$ ls | parallel cp {} {}.bak

$ loop './do_thing.sh' --every 15s --until-success --num 5
$ parallel --retries 5 --delay 15s ::: ./do_thing.sh

https://github.com/Miserlou/Loop/ (Last checked: 2018-10)

lorikeet can run jobs in parallel. It does this based on a dependency graph described in a file, so this is similar to make.

https://github.com/cetra3/lorikeet (Last checked: 2018-10)

spp can run jobs in parallel. spp does not use a command template to generate the jobs, but requires jobs to be in a file. Output from the jobs mix.

https://github.com/john01dav/spp (Last checked: 2019-01)

paral prints a lot of status information and stores the output from the commands run into files. This means it cannot be used the middle of a pipe like this

paral "echo this" "echo does not" "echo work" | wc

Instead it puts the output into files named like out_#_command.out.log. To get a very similar behaviour with GNU parallel use --results 'out_{#}_{=s/[^\sa-z_0-9]//g;s/\s+/_/g=}.log' --eta

paral only takes arguments on the command line and each argument should be a full command. Thus it does not use command templates.

This limits how many jobs it can run in total, because they all need to fit on a single command line.

paral has no support for running jobs remotely.

The examples from README.markdown and the corresponding command run with GNU parallel (--results 'out_{#}_{=s/[^\sa-z_0-9]//g;s/\s+/_/g=}.log' --eta is omitted from the GNU parallel command):

paral "command 1" "command 2 --flag" "command arg1 arg2"
parallel ::: "command 1" "command 2 --flag" "command arg1 arg2"

paral "sleep 1 && echo c1" "sleep 2 && echo c2" \
  "sleep 3 && echo c3" "sleep 4 && echo c4"  "sleep 5 && echo c5"
parallel ::: "sleep 1 && echo c1" "sleep 2 && echo c2" \
  "sleep 3 && echo c3" "sleep 4 && echo c4"  "sleep 5 && echo c5"
# Or shorter:
parallel "sleep {} && echo c{}" ::: {1..5}

paral -n=0 "sleep 5 && echo c5" "sleep 4 && echo c4" \
  "sleep 3 && echo c3" "sleep 2 && echo c2" "sleep 1 && echo c1"
parallel ::: "sleep 5 && echo c5" "sleep 4 && echo c4" \
  "sleep 3 && echo c3" "sleep 2 && echo c2" "sleep 1 && echo c1"
# Or shorter:
parallel -j0 "sleep {} && echo c{}" ::: 5 4 3 2 1

paral -n=1 "sleep 5 && echo c5" "sleep 4 && echo c4" \
  "sleep 3 && echo c3" "sleep 2 && echo c2" "sleep 1 && echo c1"
parallel -j1 "sleep {} && echo c{}" ::: 5 4 3 2 1

paral -n=2 "sleep 5 && echo c5" "sleep 4 && echo c4" \
  "sleep 3 && echo c3" "sleep 2 && echo c2" "sleep 1 && echo c1"
parallel -j2 "sleep {} && echo c{}" ::: 5 4 3 2 1

paral -n=5 "sleep 5 && echo c5" "sleep 4 && echo c4" \
  "sleep 3 && echo c3" "sleep 2 && echo c2" "sleep 1 && echo c1"
parallel -j5 "sleep {} && echo c{}" ::: 5 4 3 2 1

paral -n=1 "echo a && sleep 0.5 && echo b && sleep 0.5 && \
  echo c && sleep 0.5 && echo d && sleep 0.5 && \
  echo e && sleep 0.5 && echo f && sleep 0.5 && \
  echo g && sleep 0.5 && echo h"
parallel ::: "echo a && sleep 0.5 && echo b && sleep 0.5 && \
  echo c && sleep 0.5 && echo d && sleep 0.5 && \
  echo e && sleep 0.5 && echo f && sleep 0.5 && \
  echo g && sleep 0.5 && echo h"

https://github.com/amattn/paral (Last checked: 2019-01)

concurr is built to run jobs in parallel using a client/server model.

The examples from README.md:

concurr 'echo job {#} on slot {%}: {}' : arg1 arg2 arg3 arg4
parallel 'echo job {#} on slot {%}: {}' ::: arg1 arg2 arg3 arg4

concurr 'echo job {#} on slot {%}: {}' :: file1 file2 file3
parallel 'echo job {#} on slot {%}: {}' :::: file1 file2 file3

concurr 'echo {}' < input_file
parallel 'echo {}' < input_file

cat file | concurr 'echo {}'
cat file | parallel 'echo {}'

concurr deals badly empty input files and with output larger than 64 KB.

https://github.com/mmstick/concurr (Last checked: 2019-01)

lesser-parallel is the inspiration for parallel --embed. Both lesser-parallel and parallel --embed define bash functions that can be included as part of a bash script to run jobs in parallel.

lesser-parallel implements a few of the replacement strings, but hardly any options, whereas parallel --embed gives you the full GNU parallel experience.

https://github.com/kou1okada/lesser-parallel (Last checked: 2019-01)

npm-parallel can run npm tasks in parallel.

There are no examples and very little documentation, so it is hard to compare to GNU parallel.

https://github.com/spion/npm-parallel (Last checked: 2019-01)

machma runs tasks in parallel. It gives time stamped output. It buffers in RAM. The examples from README.md:

# Put shorthand for timestamp in config for the examples
echo '--rpl '\
  \''{time} $_=::strftime("%Y-%m-%d %H:%M:%S",localtime())'\' \
  > ~/.parallel/machma
echo '--line-buffer --tagstring "{#} {time} {}"' >> ~/.parallel/machma

find . -iname '*.jpg' |
  machma --  mogrify -resize 1200x1200 -filter Lanczos {}
find . -iname '*.jpg' |
  parallel --bar -Jmachma mogrify -resize 1200x1200 -filter Lanczos {}

cat /tmp/ips | machma -p 2 -- ping -c 2 -q {}
cat /tmp/ips | parallel -j2 -Jmachma ping -c 2 -q {}

cat /tmp/ips |
  machma -- sh -c 'ping -c 2 -q $0 > /dev/null && echo alive' {}
cat /tmp/ips |
  parallel -Jmachma 'ping -c 2 -q {} > /dev/null && echo alive'

find . -iname '*.jpg' |
  machma --timeout 5s -- mogrify -resize 1200x1200 -filter Lanczos {}
find . -iname '*.jpg' |
  parallel --timeout 5s --bar mogrify -resize 1200x1200 \
    -filter Lanczos {}

find . -iname '*.jpg' -print0 |
  machma --null --  mogrify -resize 1200x1200 -filter Lanczos {}
find . -iname '*.jpg' -print0 |
  parallel --null --bar mogrify -resize 1200x1200 -filter Lanczos {}

https://github.com/fd0/machma (Last checked: 2019-06)

Summary table (see legend above): - I2 I3 I4 - - - M1 - M3 - - M6 - O2 O3 - - - - x x E1 E2 - - - - - - - - - - - - - - - -

interlace is built for network analysis to run network tools in parallel.

interface does not buffer output, so output from different jobs mixes.

The overhead for each target is O(n*n), so with 1000 targets it becomes very slow with an overhead in the order of 500ms/target.

Using prips most of the examples from https://github.com/codingo/Interlace can be run with GNU parallel:

Blocker

commands.txt:
  mkdir -p _output_/_target_/scans/
  _blocker_
  nmap _target_ -oA _output_/_target_/scans/_target_-nmap
interlace -tL ./targets.txt -cL commands.txt -o $output

parallel -a targets.txt \
  mkdir -p $output/{}/scans/\; nmap {} -oA $output/{}/scans/{}-nmap

Blocks

commands.txt:
  _block:nmap_
  mkdir -p _target_/output/scans/
  nmap _target_ -oN _target_/output/scans/_target_-nmap
  _block:nmap_
  nikto --host _target_
interlace -tL ./targets.txt -cL commands.txt

_nmap() {
  mkdir -p $1/output/scans/
  nmap $1 -oN $1/output/scans/$1-nmap
}
export -f _nmap
parallel ::: _nmap "nikto --host" :::: targets.txt

Run Nikto Over Multiple Sites

interlace -tL ./targets.txt -threads 5 \
  -c "nikto --host _target_ > ./_target_-nikto.txt" -v

parallel -a targets.txt -P5 nikto --host {} \> ./{}_-nikto.txt

Run Nikto Over Multiple Sites and Ports

interlace -tL ./targets.txt -threads 5 -c \
  "nikto --host _target_:_port_ > ./_target_-_port_-nikto.txt" \
  -p 80,443 -v

parallel -P5 nikto --host {1}:{2} \> ./{1}-{2}-nikto.txt \
  :::: targets.txt ::: 80 443

Run a List of Commands against Target Hosts

commands.txt:
  nikto --host _target_:_port_ > _output_/_target_-nikto.txt
  sslscan _target_:_port_ >  _output_/_target_-sslscan.txt
  testssl.sh _target_:_port_ > _output_/_target_-testssl.txt
interlace -t example.com -o ~/Engagements/example/ \
  -cL ./commands.txt -p 80,443

parallel --results ~/Engagements/example/{2}:{3}{1} {1} {2}:{3} \
  ::: "nikto --host" sslscan testssl.sh ::: example.com ::: 80 443

CIDR notation with an application that doesn't support it

interlace -t 192.168.12.0/24 -c "vhostscan _target_ \
  -oN _output_/_target_-vhosts.txt" -o ~/scans/ -threads 50

prips 192.168.12.0/24 |
  parallel -P50 vhostscan {} -oN ~/scans/{}-vhosts.txt

Glob notation with an application that doesn't support it

interlace -t 192.168.12.* -c "vhostscan _target_ \
  -oN _output_/_target_-vhosts.txt" -o ~/scans/ -threads 50

# Glob is not supported in prips
prips 192.168.12.0/24 |
  parallel -P50 vhostscan {} -oN ~/scans/{}-vhosts.txt

Dash (-) notation with an application that doesn't support it

interlace -t 192.168.12.1-15 -c \
  "vhostscan _target_ -oN _output_/_target_-vhosts.txt" \
  -o ~/scans/ -threads 50

# Dash notation is not supported in prips
prips 192.168.12.1 192.168.12.15 |
  parallel -P50 vhostscan {} -oN ~/scans/{}-vhosts.txt

Threading Support for an application that doesn't support it

interlace -tL ./target-list.txt -c \
  "vhostscan -t _target_ -oN _output_/_target_-vhosts.txt" \
  -o ~/scans/ -threads 50

cat ./target-list.txt |
  parallel -P50 vhostscan -t {} -oN ~/scans/{}-vhosts.txt

alternatively

./vhosts-commands.txt:
  vhostscan -t $target -oN _output_/_target_-vhosts.txt
interlace -cL ./vhosts-commands.txt -tL ./target-list.txt \
  -threads 50 -o ~/scans

./vhosts-commands.txt:
  vhostscan -t "$1" -oN "$2"
parallel -P50 ./vhosts-commands.txt {} ~/scans/{}-vhosts.txt \
  :::: ./target-list.txt

Exclusions

interlace -t 192.168.12.0/24 -e 192.168.12.0/26 -c \
  "vhostscan _target_ -oN _output_/_target_-vhosts.txt" \
  -o ~/scans/ -threads 50

prips 192.168.12.0/24 | grep -xv -Ff <(prips 192.168.12.0/26) |
  parallel -P50 vhostscan {} -oN ~/scans/{}-vhosts.txt

Run Nikto Using Multiple Proxies

interlace -tL ./targets.txt -pL ./proxies.txt -threads 5 -c \
  "nikto --host _target_:_port_ -useproxy _proxy_ > \
   ./_target_-_port_-nikto.txt" -p 80,443 -v

parallel -j5 \
  "nikto --host {1}:{2} -useproxy {3} > ./{1}-{2}-nikto.txt" \
  :::: ./targets.txt ::: 80 443 :::: ./proxies.txt

https://github.com/codingo/Interlace (Last checked: 2019-09)

I have been unable to get the code to run at all. It seems unfinished.

https://github.com/otonvm/Parallel (Last checked: 2019-02)

par requires Haskell to work. This limits the number of platforms this can work on.

par does line buffering in memory. The memory usage is 3x the longest line (compared to 1x for parallel --lb). Commands must be given as arguments. There is no template.

These are the examples from https://github.com/k-bx/par with the corresponding GNU parallel command.

par "echo foo; sleep 1; echo foo; sleep 1; echo foo" \
    "echo bar; sleep 1; echo bar; sleep 1; echo bar" && echo "success"
parallel --lb ::: "echo foo; sleep 1; echo foo; sleep 1; echo foo" \
    "echo bar; sleep 1; echo bar; sleep 1; echo bar" && echo "success"

par "echo foo; sleep 1; foofoo" \
    "echo bar; sleep 1; echo bar; sleep 1; echo bar" && echo "success"
parallel --lb --halt 1 ::: "echo foo; sleep 1; foofoo" \
    "echo bar; sleep 1; echo bar; sleep 1; echo bar" && echo "success"

par "PARPREFIX=[fooechoer] echo foo" "PARPREFIX=[bar] echo bar"
parallel --lb --colsep , --tagstring {1} {2} \
  ::: "[fooechoer],echo foo" "[bar],echo bar"

par --succeed "foo" "bar" && echo 'wow'
parallel "foo" "bar"; true && echo 'wow'

https://github.com/k-bx/par (Last checked: 2019-02)

parallelshell does not allow for composed commands:

# This does not work
parallelshell 'echo foo;echo bar' 'echo baz;echo quuz'

Instead you have to wrap that in a shell:

parallelshell 'sh -c "echo foo;echo bar"' 'sh -c "echo baz;echo quuz"'

It buffers output in RAM. All commands must be given on the command line and all commands are started in parallel at the same time. This will cause the system to freeze if there are so many jobs that there is not enough memory to run them all at the same time.

https://github.com/keithamus/parallelshell (Last checked: 2019-02)

https://github.com/darkguy2008/parallelshell (Last checked: 2019-03)

shell-executor does not allow for composed commands:

# This does not work
sx 'echo foo;echo bar' 'echo baz;echo quuz'

Instead you have to wrap that in a shell:

sx 'sh -c "echo foo;echo bar"' 'sh -c "echo baz;echo quuz"'

It buffers output in RAM. All commands must be given on the command line and all commands are started in parallel at the same time. This will cause the system to freeze if there are so many jobs that there is not enough memory to run them all at the same time.

https://github.com/royriojas/shell-executor (Last checked: 2019-02)

par buffers in memory to avoid mixing of jobs. It takes 1s per 1 million output lines.

par needs to have all commands before starting the first job. The jobs are read from stdin (standard input) so any quoting will have to be done by the user.

Stdout (standard output) is prepended with o:. Stderr (standard error) is sendt to stdout (standard output) and prepended with e:.

For short jobs with little output par is 20% faster than GNU parallel and 60% slower than xargs.

http://savannah.nongnu.org/projects/par (Last checked: 2019-02)

fd does not support composed commands, so commands must be wrapped in sh -c.

It buffers output in RAM.

It only takes file names from the filesystem as input (similar to find).

https://github.com/sharkdp/fd (Last checked: 2019-02)

lateral is very similar to sem: It takes a single command and runs it in the background. The design means that output from parallel running jobs may mix. If it dies unexpectly it leaves a socket in ~/.lateral/socket.PID.

lateral deals badly with too long command lines. This makes the lateral server crash:

lateral run echo `seq 100000| head -c 1000k`

Any options will be read by lateral so this does not work (lateral interprets the -l):

lateral run ls -l

Composed commands do not work:

lateral run pwd ';' ls

Functions do not work:

myfunc() { echo a; }
export -f myfunc
lateral run myfunc

Running emacs in the terminal causes the parent shell to die:

echo '#!/bin/bash' > mycmd
echo emacs -nw >> mycmd
chmod +x mycmd
lateral start
lateral run ./mycmd

Here are the examples from https://github.com/akramer/lateral with the corresponding GNU sem and GNU parallel commands:

1$ lateral start
1$ for i in $(cat /tmp/names); do
1$   lateral run -- some_command $i
1$ done
1$ lateral wait
1$
1$ for i in $(cat /tmp/names); do
1$   sem some_command $i
1$ done
1$ sem --wait
1$
1$ parallel some_command :::: /tmp/names

2$ lateral start
2$ for i in $(seq 1 100); do
2$   lateral run -- my_slow_command < workfile$i > /tmp/logfile$i
2$ done
2$ lateral wait
2$
2$ for i in $(seq 1 100); do
2$   sem my_slow_command < workfile$i > /tmp/logfile$i
2$ done
2$ sem --wait
2$
2$ parallel 'my_slow_command < workfile{} > /tmp/logfile{}' \
     ::: {1..100}

3$ lateral start -p 0 # yup, it will just queue tasks
3$ for i in $(seq 1 100); do
3$   lateral run -- command_still_outputs_but_wont_spam inputfile$i
3$ done
3$ # command output spam can commence
3$ lateral config -p 10; lateral wait
3$
3$ for i in $(seq 1 100); do
3$   echo "command inputfile$i" >> joblist
3$ done
3$ parallel -j 10 :::: joblist
3$
3$ echo 1 > /tmp/njobs
3$ parallel -j /tmp/njobs command inputfile{} \
     ::: {1..100} &
3$ echo 10 >/tmp/njobs
3$ wait

https://github.com/akramer/lateral (Last checked: 2019-03)

The examples from https://github.com/amritb/with-this.git and the corresponding GNU parallel command:

with -v "$(cat myurls.txt)" "curl -L this"
parallel curl -L ::: myurls.txt

with -v "$(cat myregions.txt)" \
  "aws --region=this ec2 describe-instance-status"
parallel aws --region={} ec2 describe-instance-status \
  :::: myregions.txt

with -v "$(ls)" "kubectl --kubeconfig=this get pods"
ls | parallel kubectl --kubeconfig={} get pods

with -v "$(ls | grep config)" "kubectl --kubeconfig=this get pods"
ls | grep config | parallel kubectl --kubeconfig={} get pods

with -v "$(echo {1..10})" "echo 123"
parallel -N0 echo 123 ::: {1..10}

Stderr is merged with stdout. with-this buffers in RAM. It uses 3x the output size, so you cannot have output larger than 1/3rd the amount of RAM. The input values cannot contain spaces. Composed commands do not work.

with-this gives some additional information, so the output has to be cleaned before piping it to the next command.

https://github.com/amritb/with-this.git (Last checked: 2019-03)

Summary table (see legend above): - - - I4 - - I7 - - M3 - - M6 - O2 O3 - O5 O6 - x x E1 - - - - - E7 - x x x x x x x x - -

EXAMPLES FROM Tollef's parallel MANUAL

Tollef parallel sh -c "echo hi; sleep 2; echo bye" -- 1 2 3

GNU parallel "echo hi; sleep 2; echo bye" ::: 1 2 3

Tollef parallel -j 3 ufraw -o processed -- *.NEF

GNU parallel -j 3 ufraw -o processed ::: *.NEF

Tollef parallel -j 3 -- ls df "echo hi"

GNU parallel -j 3 ::: ls df "echo hi"

(Last checked: 2019-08)

Summary table (see legend above): I1 - - - - - I7 - - M3 M4 - - - O2 O3 - O5 O6 - O8 - E1 - - E4 - - - - - - - - - - - - - -

rargs has elegant ways of doing named regexp capture and field ranges.

With GNU parallel you can use --rpl to get a similar functionality as regexp capture gives, and use join and @arg to get the field ranges. But the syntax is longer. This:

--rpl '{r(\d+)\.\.(\d+)} $_=join"$opt::colsep",@arg[$$1..$$2]'

would make it possible to use:

{1r3..6}

for field 3..6.

For full support of {n..m:s} including negative numbers use a dynamic replacement string like this:

PARALLEL=--rpl\ \''{r((-?\d+)?)\.\.((-?\d+)?)((:([^}]*))?)} $a = defined $$2 ? $$2 < 0 ? 1+$#arg+$$2 : $$2 : 1; $b = defined $$4 ? $$4 < 0 ? 1+$#arg+$$4 : $$4 : $#arg+1; $s = defined $$6 ? $$7 : " "; $_ = join $s,@arg[$a..$b]'\' export PARALLEL

You can then do:

head /etc/passwd | parallel --colsep : echo ..={1r..} ..3={1r..3} 4..={1r4..} 2..4={1r2..4} 3..3={1r3..3} ..3:-={1r..3:-} ..3:/={1r..3:/} -1={-1} -5={-5} -6={-6} -3..={1r-3..}

EXAMPLES FROM rargs MANUAL

ls *.bak | rargs -p '(.*)\.bak' mv {0} {1}
ls *.bak | parallel mv {} {.}

cat download-list.csv | rargs -p '(?P<url>.*),(?P<filename>.*)' wget {url} -O {filename}
cat download-list.csv | parallel --csv wget {1} -O {2}
# or use regexps:
cat download-list.csv |
  parallel --rpl '{url} s/,.*//' --rpl '{filename} s/.*?,//' wget {url} -O {filename}

cat /etc/passwd | rargs -d: echo -e 'id: "{1}"\t name: "{5}"\t rest: "{6..::}"'
cat /etc/passwd |
  parallel -q --colsep : echo -e 'id: "{1}"\t name: "{5}"\t rest: "{=6 $_=join":",@arg[6..$#arg]=}"'

https://github.com/lotabout/rargs (Last checked: 2020-01)

https://pypi.org/project/papply/

https://github.com/jreisinger/runp

https://github.com/JeiKeiLim/simple_distribute_job

https://github.com/reggi/pkgrun

https://github.com/benoror/better-npm-run - not obvious how to use

https://github.com/bahmutov/with-package

https://github.com/xuchenCN/go-pssh

https://github.com/flesler/parallel

https://github.com/Julian/Verge

https://github.com/ExpectationMax/simple_gpu_scheduler simple_gpu_scheduler --gpus 0 1 2 < gpu_commands.txt parallel -j3 --shuf CUDA_VISIBLE_DEVICES='{=1 $_=slot()-1 =} {=uq;=}' < gpu_commands.txt

simple_hypersearch "python3 train_dnn.py --lr {lr} --batch_size {bs}" -p lr 0.001 0.0005 0.0001 -p bs 32 64 128 | simple_gpu_scheduler --gpus 0,1,2
parallel --header : --shuf -j3 -v CUDA_VISIBLE_DEVICES='{=1 $_=slot()-1 =}' python3 train_dnn.py --lr {lr} --batch_size {bs} ::: lr 0.001 0.0005 0.0001 ::: bs 32 64 128

simple_hypersearch "python3 train_dnn.py --lr {lr} --batch_size {bs}" --n-samples 5 -p lr 0.001 0.0005 0.0001 -p bs 32 64 128 | simple_gpu_scheduler --gpus 0,1,2
parallel --header : --shuf CUDA_VISIBLE_DEVICES='{=1 $_=slot()-1; seq() > 5 and skip() =}' python3 train_dnn.py --lr {lr} --batch_size {bs} ::: lr 0.001 0.0005 0.0001 ::: bs 32 64 128

touch gpu.queue
tail -f -n 0 gpu.queue | simple_gpu_scheduler --gpus 0,1,2 &
echo "my_command_with | and stuff > logfile" >> gpu.queue

touch gpu.queue
tail -f -n 0 gpu.queue | parallel -j3 CUDA_VISIBLE_DEVICES='{=1 $_=slot()-1 =} {=uq;=}' &
# Needed to fill job slots once
seq 3 | parallel echo true >> gpu.queue
# Add jobs
echo "my_command_with | and stuff > logfile" >> gpu.queue
# Needed to flush output from completed jobs 
seq 3 | parallel echo true >> gpu.queue

There are certain issues that are very common on parallelizing tools. Here are a few stress tests. Be warned: If the tool is badly coded it may overload your machine.

Output from 2 jobs should not mix. If the output is not used, this does not matter; but if the output is used then it is important that you do not get half a line from one job followed by half a line from another job.

If the tool does not buffer, output will most likely mix now and then.

This test stresses whether output mixes.

#!/bin/bash

paralleltool="parallel -j0"

cat <<-EOF > mycommand
#!/bin/bash

# If a, b, c, d, e, and f mix: Very bad
perl -e 'print STDOUT "a"x3000_000," "'
perl -e 'print STDERR "b"x3000_000," "'
perl -e 'print STDOUT "c"x3000_000," "'
perl -e 'print STDERR "d"x3000_000," "'
perl -e 'print STDOUT "e"x3000_000," "'
perl -e 'print STDERR "f"x3000_000," "'
echo
echo >&2
EOF
chmod +x mycommand

# Run 30 jobs in parallel
seq 30 |
  $paralleltool ./mycommand > >(tr -s abcdef) 2> >(tr -s abcdef >&2)

# 'a c e' and 'b d f' should always stay together
# and there should only be a single line per job

Output from stdout and stderr should not be merged, but kept separated.

This test shows whether stdout is mixed with stderr.

#!/bin/bash

paralleltool="parallel -j0"

cat <<-EOF > mycommand
#!/bin/bash

echo stdout
echo stderr >&2
echo stdout
echo stderr >&2
EOF
chmod +x mycommand

# Run one job
echo |
  $paralleltool ./mycommand > stdout 2> stderr
cat stdout
cat stderr

Some tools cache output in RAM. This makes them extremely slow if the output is bigger than physical memory and crash if the output is bigger than the virtual memory.

#!/bin/bash

paralleltool="parallel -j0"

cat <<'EOF' > mycommand
#!/bin/bash

# Generate 1 GB output
yes "`perl -e 'print \"c\"x30_000'`" | head -c 1G
EOF
chmod +x mycommand

# Run 20 jobs in parallel
# Adjust 20 to be > physical RAM and < free space on /tmp
seq 20 | time $paralleltool ./mycommand | wc -c

If caching is done on disk, the disk can run full during the run. Not all programs discover this. GNU Parallel discovers it, if it stays full for at least 2 seconds.

#!/bin/bash

paralleltool="parallel -j0"

# This should be a dir with less than 100 GB free space
smalldisk=/tmp/shm/parallel

TMPDIR="$smalldisk"
export TMPDIR

max_output() {
    # Force worst case scenario:
    # Make GNU Parallel only check once per second
    sleep 10
    # Generate 100 GB to fill $TMPDIR
    # Adjust if /tmp is bigger than 100 GB
    yes | head -c 100G >$TMPDIR/$$
    # Generate 10 MB output that will not be buffered due to full disk
    perl -e 'print "X"x10_000_000' | head -c 10M
    echo This part is missing from incomplete output
    sleep 2
    rm $TMPDIR/$$
    echo Final output
}

export -f max_output
seq 10 | $paralleltool max_output | tr -s X

Some tools do not clean up tmp files if they are killed. If the tool buffers on disk, they may not clean up, if they are killed.

#!/bin/bash

paralleltool=parallel

ls /tmp >/tmp/before
seq 10 | $paralleltool sleep &
pid=$!
# Give the tool time to start up
sleep 1
# Kill it without giving it a chance to cleanup
kill -9 $!
# Should be empty: No files should be left behind
diff <(ls /tmp) /tmp/before

It is not uncommon for users to create files like:

My brother's 12" *** record  (costs $$$).jpg

Some tools break on this.

#!/bin/bash

paralleltool=parallel

touch "My brother's 12\" *** record  (costs \$\$\$).jpg"
ls My*jpg | $paralleltool ls -l

Some tools require you to wrap composed commands into bash -c.

echo bar | $paralleltool echo foo';' echo {}

Some tools can only insert the argument once.

echo bar | $paralleltool echo {} foo {}

Some tools limit the length of the input lines artificially with no good reason. GNU parallel does not:

perl -e 'print "foo."."x"x100_000_000' | parallel echo {.}

GNU parallel limits the command to run to 128 KB due to execve(1):

perl -e 'print "x"x131_000' | parallel echo {} | wc

Some tools become very slow if output lines have many words.

#!/bin/bash

paralleltool=parallel

cat <<-EOF > mycommand
#!/bin/bash

# 10 MB of lines with 1000 words
yes "`seq 1000`" | head -c 10M
EOF
chmod +x mycommand

# Run 30 jobs in parallel
seq 30 | time $paralleltool -j0 ./mycommand > /dev/null

When using GNU parallel for a publication please cite:

O. Tange (2011): GNU Parallel - The Command-Line Power Tool, ;login: The USENIX Magazine, February 2011:42-47.

This helps funding further development; and it won't cost you a cent. If you pay 10000 EUR you should feel free to use GNU Parallel without citing.

Copyright (C) 2007-10-18 Ole Tange, http://ole.tange.dk

Copyright (C) 2008-2010 Ole Tange, http://ole.tange.dk

Copyright (C) 2010-2020 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.

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 Share
to copy, distribute and transmit the work
to Remix
to adapt the work

Under the following conditions:

Attribution
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).
Share Alike
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:

Waiver
Any of the above conditions can be waived if you get permission from the copyright holder.
Public Domain
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.
Other Rights
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.
Notice
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. For remote usage it also uses rsync with ssh.

find(1), xargs(1), make(1), pexec(1), ppss(1), xjobs(1), prll(1), dxargs(1), mdm(1)
2020-01-31 20200122