parallel/src/parallel_alternatives.pod

#!/usr/bin/perl -w

=encoding utf8

=head1 NAME

parallel_alternatives - Alternatives to GNU B<parallel>


=head1 DIFFERENCES BETWEEN GNU Parallel AND ALTERNATIVES

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

B<parallel> has existed since 2002 and as GNU B<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 B<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.


=head2 SUMMARY TABLE

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

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

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
E1 E2 E3 E4 E5 E6
R1 R2 R3 R4 R5 R6 R7 R8 R9
S1 S2

xargs:
I1 I2 -  -  -  -  -
-  M2 M3 -  -  -
-  O2 O3 -  O5 O6
E1 -  -  -  -  -
-  -  -  -  -  x  -  -  -
-  -

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

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

ppss:
I1 I2 -  -  -  -  I7
M1 -  M3 -  -  M6
O1 -  -  x  -  -
E1 E2 ?E3 E4 - -
R1 R2 R3 R4 -  -  ?R7 ? ?
-  -

pexec:
I1 I2 -  I4 I5 -  -
M1 -  M3 -  -  M6
O1 O2 O3 -  O5 O6
E1 -  -  E4 -  E6
R1 -  -  -  -  R6 -  -  -
S1 -

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).


=head2 DIFFERENCES BETWEEN xargs AND GNU Parallel

B<xargs> offers some of the same possibilities as GNU B<parallel>.

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

GNU B<parallel>'s newline separation can be emulated with:

B<cat | xargs -d "\n" -n1 I<command>>

B<xargs> can run a given number of jobs in parallel, but has no
support for running number-of-cpu-cores jobs in parallel.

B<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
B<Parallel grep> cannot be done reliably with B<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 {}'

B<xargs> has no support for keeping the order of the output, therefore
if running jobs in parallel using B<xargs> the output of the second
job cannot be postponed till the first job is done.

B<xargs> has no support for running jobs on remote computers.

B<xargs> has no support for context replace, so you will have to create the
arguments.

If you use a replace string in B<xargs> (B<-I>) you can not force
B<xargs> to use more than one argument.

Quoting in B<xargs> works like B<-q> in GNU B<parallel>. This means
composed commands and redirection require using B<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/


=head2 DIFFERENCES BETWEEN find -exec AND GNU Parallel

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

B<find -exec> only works on files. Processing other input (such as
hosts or URLs) will require creating these inputs as files. B<find
-exec> has no support for running commands in parallel.

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


=head2 DIFFERENCES BETWEEN make -j AND GNU Parallel

B<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.

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

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

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


=head2 DIFFERENCES BETWEEN ppss AND GNU Parallel

B<ppss> is also a tool for running jobs in parallel.

The output of B<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. B<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
B<ppss> again it may cause nothing to happen as B<ppss> thinks the
task is already done. GNU B<parallel> will normally not need cleaning
up if running locally and will only need cleaning up if stopped
abnormally and running remote (B<--cleanup> may not complete if
stopped abnormally). The example B<Parallel grep> would require extra
postprocessing if written using B<ppss>.

For remote systems PPSS requires 3 steps: config, deploy, and
start. GNU B<parallel> only requires one step.

=head3 EXAMPLES FROM ppss MANUAL

Here are the examples from B<ppss>'s manual page with the equivalent
using GNU B<parallel>:

B<1> ./ppss.sh standalone -d /path/to/files -c 'gzip '

B<1> find /path/to/files -type f | parallel gzip

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

B<2> find /path/to/files -type f | parallel cp {} /destination/dir

B<3> ./ppss.sh standalone -f list-of-urls.txt -c 'wget -q '

B<3> parallel -a list-of-urls.txt wget -q

B<4> ./ppss.sh standalone -f list-of-urls.txt -c 'wget -q "$ITEM"'

B<4> parallel -a list-of-urls.txt wget -q {}

B<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

B<5> # parallel does not use configs. If you want a different username put it in nodes.txt: user@hostname

B<5> find source/dir -type f | parallel --sshloginfile nodes.txt --trc {.}.mp3 lame -a {} -o {.}.mp3 --preset standard --quiet

B<6> ./ppss stop -C config.cfg

B<6> killall -TERM parallel

B<7> ./ppss pause -C config.cfg

B<7> Press: CTRL-Z or killall -SIGTSTP parallel

B<8> ./ppss continue -C config.cfg

B<8> Enter: fg or killall -SIGCONT parallel

B<9> ./ppss.sh status -C config.cfg

B<9> killall -SIGUSR2 parallel

https://github.com/louwrentius/PPSS


=head2 DIFFERENCES BETWEEN pexec AND GNU Parallel

B<pexec> is also a tool for running jobs in parallel.

=head3 EXAMPLES FROM pexec MANUAL

Here are the examples from B<pexec>'s info page with the equivalent
using GNU B<parallel>:

B<1> pexec -o sqrt-%s.dat -p "$(seq 10)" -e NUM -n 4 -c -- \
  'echo "scale=10000;sqrt($NUM)" | bc'

B<1> seq 10 | parallel -j4 'echo "scale=10000;sqrt({})" | bc > sqrt-{}.dat'

B<2> pexec -p "$(ls myfiles*.ext)" -i %s -o %s.sort -- sort

B<2> ls myfiles*.ext | parallel sort {} ">{}.sort"

B<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'

B<3> parallel -a image.list \
  'fistar {}.fits -f 100 -F id,x,y,flux -o {}.star' 2>star.log

B<4> pexec -r *.png -e IMG -c -o - -- \
  'convert $IMG ${IMG%.png}.jpeg ; "echo $IMG: done"'

B<4> ls *.png | parallel 'convert {} {.}.jpeg; echo {}: done'

B<5> pexec -r *.png -i %s -o %s.jpg -c 'pngtopnm | pnmtojpeg'

B<5> ls *.png | parallel 'pngtopnm < {} | pnmtojpeg > {}.jpg'

B<6> for p in *.png ; do echo ${p%.png} ; done | \
  pexec -f - -i %s.png -o %s.jpg -c 'pngtopnm | pnmtojpeg'

B<6> ls *.png | parallel 'pngtopnm < {} | pnmtojpeg > {.}.jpg'

B<7> LIST=$(for p in *.png ; do echo ${p%.png} ; done)
  pexec -r $LIST -i %s.png -o %s.jpg -c 'pngtopnm | pnmtojpeg'

B<7> ls *.png | parallel 'pngtopnm < {} | pnmtojpeg > {.}.jpg'

B<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'

B<8> Combining GNU B<parallel> and GNU B<sem>.

B<8> ls *jpg | parallel -j8 'sem --id blockread cat {} | jpegtopnm |' \
  'pnmscale 0.5 | pnmtojpeg | sem --id blockwrite cat > th_{}'

B<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:

B<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/


=head2 DIFFERENCES BETWEEN xjobs AND GNU Parallel

B<xjobs> is also a tool for running jobs in parallel. It only supports
running jobs on your local computer.

B<xjobs> deals badly with special characters just like B<xargs>. See
the section B<DIFFERENCES BETWEEN xargs AND GNU Parallel>.

Here are the examples from B<xjobs>'s man page with the equivalent
using GNU B<parallel>:

B<1> ls -1 *.zip | xjobs unzip

B<1> ls *.zip | parallel unzip

B<2> ls -1 *.zip | xjobs -n unzip

B<2> ls *.zip | parallel unzip >/dev/null

B<3> find . -name '*.bak' | xjobs gzip

B<3> find . -name '*.bak' | parallel gzip

B<4> ls -1 *.jar | sed 's/\(.*\)/\1 > \1.idx/' | xjobs jar tf

B<4> ls *.jar | parallel jar tf {} '>' {}.idx

B<5> xjobs -s script

B<5> cat script | parallel

B<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

B<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


=head2 DIFFERENCES BETWEEN prll AND GNU Parallel

B<prll> is also a tool for running jobs in parallel. It does not
support running jobs on remote computers.

B<prll> encourages using BASH aliases and BASH functions instead of
scripts. GNU B<parallel> supports scripts directly, functions if they
are exported using B<export -f>, and aliases if using B<env_parallel>.

B<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 B<prll>'s man page with the equivalent
using GNU B<parallel>:

  prll -s 'mogrify -flip $1' *.jpg
  parallel mogrify -flip ::: *.jpg

https://github.com/exzombie/prll


=head2 DIFFERENCES BETWEEN dxargs AND GNU Parallel

B<dxargs> is also a tool for running jobs in parallel.

B<dxargs> does not deal well with more simultaneous jobs than SSHD's
MaxStartups. B<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


=head2 DIFFERENCES BETWEEN mdm/middleman AND GNU Parallel

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 B<parallel>:

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

https://github.com/cklin/mdm


=head2 DIFFERENCES BETWEEN xapply AND GNU Parallel

B<xapply> can run jobs in parallel on the local computer.

Here are the examples from B<xapply>'s man page with the equivalent
using GNU B<parallel>:

B<1> xapply '(cd %1 && make all)' */

B<1> parallel 'cd {} && make all' ::: */

B<2> xapply -f 'diff %1 ../version5/%1' manifest | more

B<2> parallel diff {} ../version5/{} < manifest | more

B<3> xapply -p/dev/null -f 'diff %1 %2' manifest1 checklist1

B<3> parallel --link diff {1} {2} :::: manifest1 checklist1

B<4> xapply 'indent' *.c

B<4> parallel indent ::: *.c

B<5> find ~ksb/bin -type f ! -perm -111 -print | xapply -f -v 'chmod a+x' -

B<5> find ~ksb/bin -type f ! -perm -111 -print | parallel -v chmod a+x

B<6> find */ -... | fmt 960 1024 | xapply -f -i /dev/tty 'vi' -

B<6> sh <(find */ -... | parallel -s 1024 echo vi)

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

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

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

B<7> find ... |parallel -n5 -uj1 vi

B<8> xapply -fn "" /etc/passwd

B<8> parallel -k echo < /etc/passwd

B<9> tr ':' '\012' < /etc/passwd | xapply -7 -nf 'chown %1 %6' - - - - - - -

B<9> tr ':' '\012' < /etc/passwd | parallel -N7 chown {1} {6}

B<10> xapply '[ -d %1/RCS ] || echo %1' */

B<10> parallel '[ -d {}/RCS ] || echo {}' ::: */

B<11> xapply -f '[ -f %1 ] && echo %1' List | ...

B<11> parallel '[ -f {} ] && echo {}' < List | ...

http://carrera.databits.net/~ksb/msrc/local/bin/xapply/xapply.html


=head2 DIFFERENCES BETWEEN AIX apply AND GNU Parallel

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

Here are the examples from
https://www-01.ibm.com/support/knowledgecenter/ssw_aix_71/com.ibm.aix.cmds1/apply.htm

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

  apply echo *
  parallel echo ::: *

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

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

3. To run the B<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
B</usr/joe>, enter:

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

https://www.ibm.com/support/knowledgecenter/en/ssw_aix_61/com.ibm.aix.cmds1/apply.htm


=head2 DIFFERENCES BETWEEN paexec AND GNU Parallel

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

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

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

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

=over 1

=item 1_div_X_run:

  ../../paexec -s -l -c "`pwd`/1_div_X_cmd" -n +1 <<EOF [...]
  parallel echo {} '|' `pwd`/1_div_X_cmd <<EOF [...]

=item all_substr_run:

  ../../paexec -lp -c "`pwd`/all_substr_cmd" -n +3 <<EOF [...]
  parallel echo {} '|' `pwd`/all_substr_cmd <<EOF [...]

=item 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 [...]

=item 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 [...]

=back

https://github.com/cheusov/paexec


=head2 DIFFERENCES BETWEEN map(sitaramc) AND GNU Parallel

B<map> sees it as a feature to have less features and in doing so it
also handles corner cases incorrectly. A lot of GNU B<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: I<My
brother's 12" records.txt>

B<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 B<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'

B<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:

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

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

B<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"' ::: *

B<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

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

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

By design B<map> is option incompatible with B<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


=head2 DIFFERENCES BETWEEN ladon AND GNU Parallel

B<ladon> can run multiple jobs on files in parallel.

B<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 B<parallel> by putting this in
B<~/.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::;'

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

    ladon '*' -- seq 36000 | wc

=head3 EXAMPLES FROM ladon MANUAL

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

B<1> ladon "**/*.txt" -- echo RELPATH

B<1> parallel echo RELPATH ::: **/*.txt

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

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

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

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

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

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

https://github.com/danielgtaylor/ladon


=head2 DIFFERENCES BETWEEN jobflow AND GNU Parallel

B<jobflow> can run multiple jobs in parallel.

Just like B<xargs> output from B<jobflow> jobs running in parallel mix
together by default. B<jobflow> can buffer into files (placed in
/run/shm), but these are not cleaned up if B<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.

B<jobflow> gives no error if the command is unknown, and like B<xargs>
redirection and composed commands require wrapping with B<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.

B<jobflow> has no equivalent for B<--pipe>, or B<--sshlogin>.

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

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

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


=head3 EXAMPLES FROM jobflow README

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

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

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

B<2> seq 100 | parallel -j100 echo {}

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

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

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

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

https://github.com/rofl0r/jobflow


=head2 DIFFERENCES BETWEEN gargs AND GNU Parallel

B<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 B<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 B<gargs> website.

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

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

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

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

https://github.com/brentp/gargs


=head2 DIFFERENCES BETWEEN orgalorg AND GNU Parallel

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

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

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

https://github.com/reconquest/orgalorg


=head2 DIFFERENCES BETWEEN Rust parallel AND GNU Parallel

Rust parallel focuses on speed. It is almost as fast as B<xargs>. It
implements a few features from GNU B<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 B<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


=head2 DIFFERENCES BETWEEN Rush AND GNU Parallel

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

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

B<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 B<rush>'s website with the equivalent
command in GNU B<parallel>.

=head3 EXAMPLES

B<1. Simple run, quoting is not necessary>

  $ seq 1 3 | rush echo {}

  $ seq 1 3 | parallel echo {}

B<2. Read data from file (`-i`)>

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

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

B<3. Keep output order (`-k`)>

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

  $ seq 1 3 | parallel -k echo {}


B<4. Timeout (`-t`)>

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

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

B<5. Retry (`-r`)>

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

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

Use B<-u> to see it is really run twice:

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

B<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}

B<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 {.} {:} {/.} {/:}'

B<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}'

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

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

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

B<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}

B<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'


B<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]=}'
    )

B<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 B<parallel> you would typically do:

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

If you I<really> want the var:

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

If you I<really> want the B<for>-loop:

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

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

  My brother's 12" records


B<14. B<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 B<rush> GNU B<parallel> works with complex values:

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

B<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

B<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 I<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 {}'

B<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 B<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

B<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 B<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

=head3 Other B<rush> features

B<rush> has:

=over 4

=item * B<awk -v> like custom defined variables (B<-v>)

With GNU B<parallel> you would simply set a shell variable:

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

Also B<rush> does not like special chars. So these B<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 B<parallel> version works:

   parallel 'v=\"; echo "$v"' ::: works
   parallel 'v={}; echo "$v"' ::: "My  brother's  12\"  records"

=item * Exit on first error(s) (-e)

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

=item * Settable records sending to every command (B<-n>, default 1)

This is also called B<-n> in GNU B<parallel>.

=item * Practical replacement strings

=over 4

=item {:} remove any extension

With GNU B<parallel> this can be emulated by:

  parallel --plus echo '{/\..*/}' ::: foo.ext.bar.gz

=item {^suffix}, remove suffix

With GNU B<parallel> this can be emulated by:

  parallel --plus echo '{%.bar.gz}' ::: foo.ext.bar.gz

=item {@regexp}, capture submatch using regular expression

With GNU B<parallel> this can be emulated by:

  parallel --rpl '{@(.*?)} /$$1/ and $_=$1;' \
    echo '{@\d_(.*).gz}' ::: 1_foo.gz

=item {%.}, {%:}, basename without extension

With GNU B<parallel> this can be emulated by:

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

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

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

Then you can use them as:

  parallel echo {%.} {%:} ::: dir/foo.bar.gz

=back

=item * Preset variable (macro)

E.g.

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

With GNU B<parallel> this can be emulated by:

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

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

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


=item * Commands of multi-lines

While you I<can> use multi-lined commands in GNU B<parallel>, to
improve readability GNU B<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 {}'

=back

https://github.com/shenwei356/rush


=head2 DIFFERENCES BETWEEN ClusterSSH AND GNU Parallel

ClusterSSH solves a different problem than GNU B<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 B<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 B<-S> GNU B<parallel> may
only use one of these (e.g. if there are 8 jobs to be run and one
computer has 8 cores).

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

B<parallel --nonall -S server-a,server-b do_stuff foo bar>

https://github.com/duncs/clusterssh


=head2 DIFFERENCES BETWEEN coshell AND GNU Parallel

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

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

Output can be buffered by using B<-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


=head2 DIFFERENCES BETWEEN spread AND GNU Parallel

B<spread> runs commands on all directories.

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

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

This works except for the B<--exclude> option.


=head2 DIFFERENCES BETWEEN pyargs AND GNU Parallel

B<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.

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

=head3 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


=head2 DIFFERENCES BETWEEN concurrently AND GNU Parallel

B<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
B<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


=head2 DIFFERENCES BETWEEN map(soveran) AND GNU Parallel

B<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
B<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 B<parallel> is built for parallelization
and map is not. So GNU B<parallel> has lots of ways of dealing with the
issues that parallelization raises:

=over 4

=item *

Keep the number of processes manageable

=item *

Make sure output does not mix

=item *

Make Ctrl-C kill all running processes

=back

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


=head2 DIFFERENCES BETWEEN loop AND GNU Parallel

B<loop> mixes stdout and stderr:

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

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

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

B<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 B<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)


=head2 DIFFERENCES BETWEEN lorikeet AND GNU Parallel

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

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


=head2 Todo

Url for spread

https://github.com/john01dav/spp

https://github.com/amritb/with-this.git

https://github.com/fd0/machma Requires Go >= 1.7.

https://github.com/k-bx/par requires Haskell to work. This limits the
number of platforms this can work on.

https://github.com/otonvm/Parallel

https://github.com/flesler/parallel

https://github.com/kou1okada/lesser-parallel

https://github.com/Julian/Verge

https://github.com/amattn/paral

https://github.com/mmstick/concurr

pyargs


=head1 TESTING OTHER TOOLS

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 you machine.


=head2 MIX: Output mixes

Output from 2 jobs should not mix. If the output is not used, this
does not matter; but if the output I<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


=head2 RAM: Output limited by RAM

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


=head2 DISKFULL: Incomplete data if /tmp runs full

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


=head2 CLEANUP: Leaving tmp files at unexpected death

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


=head2 SPCCHAR: Dealing badly with special file names.

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


=head2 COMPOSED: Composed commands do not work

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

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


=head2 ONEREP: Only one replacement string allowed

Some tools can only insert the argument once.

  echo bar | $paralleltool echo {} foo {}


=head2 INPUTSIZE: Length of input should not be limited

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

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

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

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


=head2 NUMWORDS: Speed depends on number of words

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


=head1 AUTHOR

When using GNU B<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) 2010-2018 Ole Tange, http://ole.tange.dk and Free
Software Foundation, Inc.

Parts of the manual concerning B<xargs> compatibility is inspired by
the manual of B<xargs> from GNU findutils 4.4.2.


=head1 LICENSE

Copyright (C) 2007-2018 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/>.

=head2 Documentation license I

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.

=head2 Documentation license II

You are free:

=over 9

=item B<to Share>

to copy, distribute and transmit the work

=item B<to Remix>

to adapt the work

=back

Under the following conditions:

=over 9

=item B<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).

=item B<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.

=back

With the understanding that:

=over 9

=item B<Waiver>

Any of the above conditions can be waived if you get permission from
the copyright holder.

=item B<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.

=item B<Other Rights>

In no way are any of the following rights affected by the license:

=over 2

=item *

Your fair dealing or fair use rights, or other applicable
copyright exceptions and limitations;

=item *

The author's moral rights;

=item *

Rights other persons may have either in the work itself or in
how the work is used, such as publicity or privacy rights.

=back

=back

=over 9

=item B<Notice>

For any reuse or distribution, you must make clear to others the
license terms of this work.

=back

A copy of the full license is included in the file as cc-by-sa.txt.


=head1 DEPENDENCIES

GNU B<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.


=head1 SEE ALSO

B<find>(1), B<xargs>(1), B<make>(1), B<pexec>(1), B<ppss>(1),
B<xjobs>(1), B<prll>(1), B<dxargs>(1), B<mdm>(1)

=cut