MIT/Rust Parallel: A Command-line CPU Load Balancer Written in Rust
This is an attempt at recreating the functionality of GNU Parallel, a work-stealer for the command-line, in Rust under a MIT license. The end goal will be to support much of the functionality of GNU Parallel
and then to extend the functionality further for the next generation of command-line utilities written in Rust. While functionality is important, with the application being developed in Rust, the goal is to also be as fast and efficient as possible.
Note
A nightly version of the Rust compiler is required to build this application.
This is required to disable jemalloc
because it incurs a performance overhead on Linux systems.
In addition to using a nightly compiler, it is recommended to compile with a musl target.
rustup install nightly
rustup target add x86_64-unknown-linux-musl
rustup run nightly cargo build --release --target x86_64-unknown-linux-musl
Additionally, transparent_hugepages
causes serious performance issues with the implementation
of parallel, so it is recommended to ensure that your Linux distribution has the parameter set
to madvise
:
sudo sh -c "echo madvise > /sys/kernel/mm/transparent_hugepage/enabled"
It's a good idea to install the dash
shell as this implementation of Parallel will try to use it by default.
Dash is basically a superior implementation of sh
that's many times faster and more secure.
If dash
cannot be found, it will default to sh
. Although on Windows it will default to cmd
.
See the to-do list for features and improvements that have yet to be done. If you want to contribute, pull requests are welcome. If you have an idea for improvement which isn't listed in the to-do list, feel free to email me and I will consider implementing that idea.
Benchmark Comparison to GNU Parallel
GNU Parallel
Printing 1 to 10,000 in parallel
~/D/parallel (master) $ seq 1 10000 | time -v /usr/bin/parallel echo > /dev/null
Command being timed: "/usr/bin/parallel echo"
User time (seconds): 194.73
System time (seconds): 66.49
Percent of CPU this job got: 230%
Elapsed (wall clock) time (h:mm:ss or m:ss): 1:53.08
Average shared text size (kbytes): 0
Average unshared data size (kbytes): 0
Average stack size (kbytes): 0
Average total size (kbytes): 0
Maximum resident set size (kbytes): 16140
Average resident set size (kbytes): 0
Major (requiring I/O) page faults: 0
Minor (reclaiming a frame) page faults: 10329017
Voluntary context switches: 102095
Involuntary context switches: 229910
Swaps: 0
File system inputs: 0
File system outputs: 0
Socket messages sent: 0
Socket messages received: 0
Signals delivered: 0
Page size (bytes): 4096
Exit status: 0
Cat the contents of every binary in /usr/bin
~/D/parallel (master) $ time -v /usr/bin/parallel cat ::: /usr/bin/* > /dev/null
User time (seconds): 71.71
System time (seconds): 27.67
Percent of CPU this job got: 222%
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:44.62
Average shared text size (kbytes): 0
Average unshared data size (kbytes): 0
Average stack size (kbytes): 0
Average total size (kbytes): 0
Maximum resident set size (kbytes): 17576
Average resident set size (kbytes): 0
Major (requiring I/O) page faults: 8
Minor (reclaiming a frame) page faults: 4112045
Voluntary context switches: 51331
Involuntary context switches: 101494
Swaps: 0
File system inputs: 1512
File system outputs: 0
Socket messages sent: 0
Socket messages received: 0
Signals delivered: 0
Page size (bytes): 4096
Exit status: 5
Rust Parallel (Built with MUSL target)
It's highly recommend to compile Parallel with MUSL instead of glibc, as this reduces memory consumption in half and doubles performance.
Printing 1 to 10,000 in parallel
~/D/parallel (master) $ seq 1 10000 | time -v target/release/x86_64-unknown-linux-musl/parallel echo > /dev/null
Command being timed: "target/x86_64-unknown-linux-musl/release/parallel echo"
User time (seconds): 0.48
System time (seconds): 2.85
Percent of CPU this job got: 104%
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:03.20
Average shared text size (kbytes): 0
Average unshared data size (kbytes): 0
Average stack size (kbytes): 0
Average total size (kbytes): 0
Maximum resident set size (kbytes): 1768
Average resident set size (kbytes): 0
Major (requiring I/O) page faults: 0
Minor (reclaiming a frame) page faults: 823754
Voluntary context switches: 82723
Involuntary context switches: 64834
Swaps: 0
File system inputs: 0
File system outputs: 320
Socket messages sent: 0
Socket messages received: 0
Signals delivered: 0
Page size (bytes): 4096
Exit status: 0
Cat the contents of every binary in /usr/bin
~/D/parallel (master) $ time -v target/release/x86_64-unknown-linux-musl/release/parallel cat ::: /usr/bin/* > /dev/null
User time (seconds): 0.96
System time (seconds): 4.61
Percent of CPU this job got: 192%
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:02.89
Average shared text size (kbytes): 0
Average unshared data size (kbytes): 0
Average stack size (kbytes): 0
Average total size (kbytes): 0
Maximum resident set size (kbytes): 1868
Average resident set size (kbytes): 0
Major (requiring I/O) page faults: 0
Minor (reclaiming a frame) page faults: 350216
Voluntary context switches: 68772
Involuntary context switches: 40085
Swaps: 0
File system inputs: 368
File system outputs: 416
Socket messages sent: 0
Socket messages received: 0
Signals delivered: 0
Page size (bytes): 4096
Exit status: 0
Syntax Examples
The following syntax is supported:
|
|
#!/usr/bin/parallel --shebang echo // Ability to use within a shebang line.
Manual
Parallel parallelizes otherwise non-parallel command-line tasks. When there are a number of commands that need to be executed, which may be executed in parallel, the Parallel application will evenly distribute tasks to all available CPU cores. There are three basic methods for how commands are supplied:
-
A COMMAND may be defined, followed by an which denotes that all following arguments will be usde as INPUTS for the command.
-
If no COMMAND is provided, then the INPUTS will be interpreted as COMMANDS.
-
If no INPUTS are provided, then standard input will be read for INPUTS.
Parallel groups the standard output and error of each child process so that
outputs are printed in the order that they are given, as if the tasks were
executed serially in a traditional for loop. In addition, commands are
executed in the platform's preferred shell by default, which is sh -c
on
Unix systems, and cmd /C
on Windows. This comes at a performance cost, so
it can be disabled with the --no-shell option.
INPUT MODES
Input modes are used to determine whether the following inputs are files that contain inputs or inputs themselves. Files with inputs have each input stored on a separate line, and each line is considered an entire input.When there are multiple collected lists of inputs, each individual input list will be permutated together into a single list.
- :::
Denotes that the input arguments that follow are input arguments. Additionally, those arguments will be collected into a new list.
- :::+
Denotes that the input arguments that follow are input arguments. Additionally, those arguments will be added to the current list.
- ::::
Denotes that the input arguments that follow are files with inputs. Additionally, those arguments will be collected into a new list.
- ::::+
Denotes that the input arguments that follow are files with inputs. Additionally, those arguments will be added to the current list.
INPUT TOKENS
COMMANDs are typically formed the same way that you would normally in the shell, only that you will replace your input arguments with placeholder tokens like {}, {.}, {/}, {//} and {/.}. If no tokens are provided, it is inferred that the final argument in the command will be {}. These tokens will perform text manipulation on the inputs to mangle them in the way you like. Ideas for more tokens are welcome.
- {}: Each occurrence will be replaced with the name of the input.
- {.}: Each occurrence will be replaced with the input, with the extension removed.
- {/}: Each occurrence will be replaced with the base name of the input.
- {/.}: Each occurrence will be replaced with the base name of the input, with the extension removed.
- {//}: Each occurrence will be replaced with the directory name of the input.
- {%}: Each occurrence will be replaced with the slot number.
- {#}: Each occurrence will be replaced with the job number.
- {##}: Each occurrence will be replaced with the total number of jobs.
- {N}: Where N is a number, display the associated job number.
- {N.}: will remove the extension from the Nth job.
- {N/}: Displays the base name (file name) of the Nth job.
- {N//}: Displays the directory name of the Nth job.
- {N/.}: Displays the base name of the Nth job with the extension removed.
OPTIONS
Options may also be supplied to the program to change how the program operates:
- --delay: Delays starting the next job for N amount of seconds, where the seconds can be fractional.
- --dry-run: Prints the jobs that will be run to standard output, without running them.
- --eta: Prints the estimated time to complete based on average runtime of running processes.
- -h, --help: Prints the manual for the application (recommended to pipe it to
less
). - -j, --jobs: Defines the number of jobs/threads to run in parallel.
- --joblog: Logs job statistics to a designated file as they are completed.
- --joblog-8601: Writes the start time in the ISO 8601 format:
YYYY-MM-DD hh:mm:ss
- --memfree: Defines the minimum amount of memory available before starting the next job.
- -n, --max-args: Groups up to a certain number of arguments together in the same command line.
- --num-cpu-cores: Prints the number of CPU cores in the system and exits.
- -p, --pipe: Instead of supplying arguments as arguments to child processes, instead supply the arguments directly to the standard input of each child process.
- -s, --silent, --quiet: Disables printing the standard output of running processes.
- --shebang: Grants ability to utilize the parallel command as an interpreter via calling it within a shebang line.
- --shellquote: Prints commands that will be executed, with the commands quoted.
- --tmpdir: Defines the directory to use for temporary files
- --timeout: If a command runs for longer than a specified number of seconds, it will be killed with a SIGKILL.
- -v, --verbose: Prints information about running processes.
- --version: Prints the current version of the application and it's dependencies.
Useful Examples
Transcoding FLAC music to Opus
ffmpeg is a highly useful application for converting music and videos. However, audio transcoding is limited to a a single core. If you have a large FLAC archive and you wanted to compress it into the efficient Opus codec, it would take forever with the fastest processor to complete, unless you were to take advantage of all cores in your CPU.
Transcoding Videos to VP9
VP9 has one glaring flaw in regards to encoding: it can only use about three cores at any given point in time. If you have an eight core processor and a dozen or more episodes of a TV series to transcode, you can use the parallel program to run three jobs at the same time, provided you also have enough memory for that.
vp9_params="-c:v libvpx-vp9 -tile-columns 6 -frame-parallel 1 -rc_lookahead 25 -threads 4 -speed 1 -b:v 0 -crf 18"
opus_params="-c:a libopus -b:a 128k"
Installation Instructions
There are a number of methods that you can use to install the application. I provide binary packages for AMD64 systems that are available for download:
Gentoo
I have a personal Gentoo layman overlay that provides this application for installation.
Arch Linux
A PKGBUILD is available for Arch Linux users from the AUR.
Everyone Else