Outrun: Execute local command using processing power of another Linux machine

Last modified on March 18, 2021

Outrun lets you finish a local recount using the processing power of every other Linux machine.

  • No have to first set up the recount on the alternative machine.
  • Reference native recordsdata and paths lots like it's seemingly you will perhaps perhaps presumably typically.
  • Works all of the gadget by wildly assorted Linux distributions, like Ubuntu and Alpine.


  • Set up
  • Usage
  • How it the reality is works
  • Boundaries
  • FAQ
  • Vogue
  • License

Set up

Outrun requires Python 3.7 and may merely be construct in using pip:

pip3 set up outrun

It will doubtless be construct in to your dangle machine and any machines that it's seemingly you will perhaps perhaps be using to bustle directions. On these different machines you like to it's best to with out a doubt set up it globally, such that it may perhaps perhaps additionally even be began with a recount like ssh specific individual@host outrun.

To boot to outrun itself, you moreover mght want to put in the FUSE 3.x library. Most Linux distributions consist of it in a bundle that's merely named fuse3.



It is effective to accumulate root internet admission to to the alternative machine, both by being able to immediately SSH into it as root or by having sudo internet admission to. Right this is important as a result of outrun makes eat of chroot.

For the rationale that a ways away machine will purchase beefy internet admission to to your native file system as your most new specific individual, assure to handiest eat machines that you just belief.


Must it's seemingly you will perhaps perhaps presumably typically bustle:

ffmpeg -i enter.mp4 -vcodec libx265 -crf 28 output.mp4

Then you definately can let outrun finish that very same recount using the processing power of specific individual@host like this:

outrun specific individual@host ffmpeg -i enter.mp4 -vcodec libx265 -crf 28 output.mp4

FFMPEG would not have to be construct in on the alternative machine and enter.mp4/output.mp4 will doubtless be be taught from and written to your native onerous strain as it's seemingly you will perhaps perhaps presumably ask.

The first time you bustle a authentic recount, it may perhaps perhaps additionally trustworthy deal with a while to initiating as a result of its dependencies should first be copied to the alternative machine. This is able to perhaps perhaps additionally trustworthy happen regardless that every other machine has already bustle FFMPEG on it forward of, because it doubtless has a a bit assorted mannequin. The time this takes is extraordinarily relying to your bandwidth and latency.


Survey the output of outrun --advantage for a high degree association of accessible alternate concepts to customize behaviour. The persistent cache on each a ways away machine may perhaps additionally even be configured by making a ~/.outrun/config file like this:

max_size=21474836480 # bytes -> 20 GB

How it the reality is works

Making a recount work on a apparent machine

Let’s deal with into consideration the ffmpeg recount from the occasion and deal with into consideration what’s important for it to bustle on the alternative machine as if it had been operating domestically.

First we'd like a tool of operating directions and gazing their output on every other machine throughout the important construct, so all of it begins with a original SSH session. Are making an attempt operating ssh -tt specific individual@host htop and as well it's seemingly you will perhaps perhaps uncover that or not it's simple to bustle an interactive a ways away program with SSH that affords a the reality is comparable skills to a domestically operating program.

Unnecessary to say, we're not desirous about operating software program that's construct in on the different machine, nonetheless reasonably the ffmpeg program on our dangle machine. A simple gadget to initiating can be to scp the /usr/bin/ffmpeg executable to the alternative machine and check out to bustle it there. Unfortunately it's seemingly you will perhaps perhaps doubtless win that the next will happen when you happen to attempt to finish it:

$ ./ffmpeg
./ffmpeg: not found

That’s for the reason that executable is dynamically linked and it tries to load its library dependencies from the file system, which don’t exist on the alternative machine. It is seemingly you will perhaps perhaps eat ldd to look which shared libraries an ELF executable like this is dependent upon:

$ ldd `which ffmpeg`
    linux-vdso.so.1 (0x00007fffb7796000)
    libavdevice.so.58=> /usr/lib/libavdevice.so.58 (0x00007f2407f2a000)
    libavfilter.so.7=> /usr/lib/libavfilter.so.7 (0x00007f2407be0000)
    libavformat.so.58=> /usr/lib/libavformat.so.58 (0x00007f2407977000)
    libavcodec.so.58=> /usr/lib/libavcodec.so.58 (0x00007f2406434000)
    libpostproc.so.55=> /usr/lib/libpostproc.so.55 (0x00007f2406414000)
    libswresample.so.3=> /usr/lib/libswresample.so.3 (0x00007f24063f4000)

We may perhaps additionally painstakingly replica all of these libraries as neatly, nonetheless that wouldn’t essentially be the pause of it. Tool like Blender, as an example, furthermore tons of comparatively a pair of different dependencies like Python recordsdata after it has began operating. Despite the incontrovertible fact that we had been to litter the a ways away file system with all of these dependencies, we might gentle not be able to bustle the same old recount since enter.mp4 would not exist on the alternative machine.

To deal with the inevitable fact that purposes may perhaps additionally trustworthy want internet admission to to any file throughout the native file system, outrun merely mirrors the overall file system of the native machine by mounting it over the community. Right this is accomplished by mounting a FUSE file system on the a ways away machine that forwards all of its operations, like listing recordsdata in a listing, to an RPC provider on the native machine. This RPC provider merely exposes capabilities like readdir() and stat() to work alongside with the native file system.

It is seemingly you will perhaps perhaps trustworthy marvel why we’re not using an current community file system resolution like NFS or SSHFS. The peril with these is that it’s refined to automate fast setup for ad-hoc periods. NFS needs to be jam up with configuration recordsdata and SSHFS requires the a ways away machine to be able to SSH again to the native machine. In distinction, the included RPC file system is a light-weight TCP server with a per-session token that may perhaps perhaps presumably additionally even be securely tunneled over the SSH session that we're already using. Having a customized resolution additionally opens up alternatives for comparatively a pair of optimizations as we're going to uncover shortly.

To illustrate we mount the native machine's file system at /tmp/local_fs. We are ready to now with out catastrophe internet admission to FFMPEG at /tmp/local_fs/usr/bin/ffmpeg and all of its library dependencies at /tmp/local_fs/usr/lib. Unfortunately it would not appear as if we now purchase made outstanding progress however:

$ /tmp/local_fs/usr/bin/ffmpeg
/tmp/local_fs/usr/bin/ffmpeg: not found

The peril is that we're gentle shopping for for its dependencies in /usr/lib on the a ways away machine. We may perhaps additionally work round this peril by having enjoyable with with ambiance variables like $LD_LIBRARY_PATH to make FFMPEG examine libraries in /tmp/local_fs/usr/lib, nonetheless then we're pleasing again to fixing one shrimp peril at a time over once more. We assign not have to scare about this if we may perhaps additionally fake that /tmp/local_fs is the basis file system in relate that /usr/lib robotically redirects to /tmp/local_fs/usr/lib in its construct. We are ready to finish precisely that using chroot.

$ chroot /tmp/local_fs /bin/bash
# ffmpeg -i enter.mp4 -vcodec libx265 -crf 28 output.mp4
ffmpeg mannequin n4.2.3 Copyright (c) 2000-2020 the FFmpeg builders
enter.mp4: No such file or listing

It the reality is works! Well, practically. We're gentle lacking some of probably the most context through which the same old recount was as soon as to be accomplished. To be prepared to look out enter.mp4 and to retailer output.mp4 inside the perfect construct, we additionally want to modify to the identical common working listing.

# cd /house/specific individual/motion pictures
# ffmpeg -i enter.mp4 -vcodec libx265 -crf 28 output.mp4
# ls
enter.mp4   output.mp4

While FFMPEG already works as anticipated, we have to additionally herald the perfect ambiance variables. For occasion, $HOME may perhaps additionally trustworthy purchase an impress on the construct configuration recordsdata are loaded from and $LANG may perhaps additionally trustworthy purchase an impress on the current language of particular purposes.

Must it's seemingly you will perhaps perhaps presumably now creep again to your dangle machine and uncover on the same old /house/specific individual/motion pictures, it's seemingly you will perhaps perhaps uncover that output.mp4 is right there as if we did not pleasing bustle FFMPEG on a totally assorted machine!


While this arrive works, the effectivity leaves so lots to be desired. Must you'll perchance be connecting to a server over the procure, even one in a construct shut to you, your latency to this might often doubtless perhaps perhaps additionally trustworthy doubtless be on the least 20 ms. Right this is additionally how extended each file system operation will deal with and that fast supplies up. For occasion, on my machine FFMPEG has 110 shared libraries that have to be loaded, which arrive that pleasing making an strive up their attributes will already deal with 2.2 seconds! If we needed to naively finish each single file system operation over the community like this then outrun would not the reality is be seemingly. That is why outrun's community file system comes with two types of optimizations: caching and prefetching.

Outrun time and again caches all recordsdata which might perchance be be taught from system directories like /usr/bin and /lib, which might perchance be recognized to occupy purposes and their dependencies. These directories are assumed to dwell unchanged all by an outrun session. Every authentic session will assessment if any of the cached recordsdata purchase modified and exchange them as important. This simple method is ample to make the identical program initiating outstanding sooner the 2nd time round, and typically additionally helps with different purposes since many dependencies like glibc are shared. An LRU cache safety is broken-down to throw out recordsdata that haven’t been broken-down shortly.

For different directories, outrun lets the kernel make original optimizations like learning recordsdata in big chunks and buffering shrimp writes into one big write, very like NFS.

The construct outrun's file system the reality is deviates from generic community file techniques is its aggressive arrive to prefetching. For occasion, when an executable like /usr/bin/ffmpeg is be taught, it is rather doubtless that that can be accomplished and its shared libraries will doubtless be loaded subsequent. Attributable to this fact, when the originate() demand /usr/bin/ffmpeg is available in, this might often doubtless perhaps perhaps additionally trustworthy not pleasing swap that executable in its entirety forward of time, nonetheless additionally swap all of its 110 shared libraries and related file system metadata in a single operation. If this assumption is factual, which it typically will doubtless be, we now purchase pleasing decreased tons of of stat()/readlink()/originate()/be taught()/shut() calls to a single RPC title with a one-time 20 ms overhead.

One different occasion of this kind of prefetch is the belief that if a .py file is accessed, it's doubtless being interpreted and Python will quickly examine its compiled .pyc companion. Attributable to this incontrovertible fact that file is right now despatched alongside with it. If the __pycache__ listing would not exist, then we merely prefetch the ENOENT error that may perhaps perhaps presumably consequence from searching for to internet admission to it in its construct.

Since compiled recordsdata are inclined to compress neatly, all file contents which might perchance be be taught of their entirety are additionally transferred with LZ4 compression to assign on bandwidth.


In enlighten so that you can be taught extra particulars about outrun and its make selections, then purchase a uncover on the supply code. Every module (file system, RPC, operation orchestration) accommodates extra specific documentation in its docstrings.


There are some issues to deal with in thoughts whereas using outrun:

  • File system effectivity stays a bottleneck, so probably the most applicable workloads are computationally constrained duties like ray tracing and video encoding. Using outrun for one factor like git construct works, nonetheless is not any longer suggested.
  • For the rationale that software program to be accomplished is copied out of your dangle machine to the a ways away machine, it have to be binary like minded. It’s not conceivable to jam up a session from an x86 machine to an ARM machine, as an example.
  • The recount will eat the community and date/time of the a ways away machine.
    • Must you deal with to want to internet admission to native endpoints, then they have to be explicitly forwarded by using the SSH flags parameter to jam up a ways away forwarding.


  • Does outrun reinforce a couple of periods on the identical time?
    • Yes, it's seemingly you will perhaps perhaps presumably bustle software program on many various machines concurrently. Every a ways away machine may perhaps additionally additionally reinforce many various machines connecting to it on the identical time with wildly assorted file techniques.
  • Why was as soon as outrun written in Python?
    • The common prototype for outrun was as soon as written in C++, nonetheless it grew to change into out that Python and its similar outdated library make it outstanding less complicated to glue orchestrate processes and dealing system interactions. Almost about file system effectivity, it doesn’t make outstanding of a inequity since community latency is by a ways the pleasurable bottleneck.


Static analysis and code vogue

Outrun was as soon as written to closely depend upon selection hints for documentation purposes and to allow for static analysis using mypy. To boot to that, flake8 is broken-down to assemble in strain code vogue and pylint is broken-down to choose additional issues.

mypy outrun
pylint outrun


Outrun comes with a verify suite for all particular person modules primarily based on pytest.

pytest --fuse --cov=outrun outrun/assessments

Since so lots of the performance in outrun is dependent upon OS interplay, its verify suite additionally entails beefy integration assessments that simulate utilization with a VM connecting to every other VM. These are jam up using Vagrant and will perhaps additionally even be bustle by along with the --vagrant flag:

pytest --vagrant --cov=outrun outrun/assessments


Outrun is licensed beneath mannequin 2.0 of the Apache License.

Read More

Similar Products:

    None Found

Recent Content