icecream - A distributed compile system
Icecream is a distributed compile system for C and C++.
Icecream is created by SUSE and is based on ideas and code by distcc. Like
distcc it takes compile jobs from your build and distributes it to remote
machines allowing a parallel build on several machines you have got. But
unlike distcc Icecream uses a central server that schedules the compile jobs
to the fastest free server and is as this dynamic. This advantage pays off
mostly for shared computers, if you are the only user on X machines, you have
full control over them anyway.
- One machine that runs the scheduler ( icecc-scheduler
- Many machines that run the daemon ( iceccd
If you want to compile using icecream, make sure $prefix/lib/icecc/bin
the first entry in your path, e.g. type export
(Hint: put this in ~/.bashrc
to not have to type it in everytime)
Then you just compile with make -j num
, where num is
the amount of jobs you want to compile in parallel. Do not exaggerate. Too
large numbers can overload your machine or the compile cluster and make the
build in fact slower.
Never use icecream in untrusted environments. Run the daemons and the scheduler
as unpriviliged user in such networks if you have to! But you will have to
rely on homogeneous networks then (see below).
If you want an overview of your icecream compile cluster, or if you just want
funny stats, you might want to run icemon.
If you are running icecream daemons in the same icecream network but on machines
with incompatible compiler versions, icecream needs to send your build
environment to remote machines (note: they all
must be running as root
if compiled without libcap-ng support. In the future icecream might gain the
ability to know when machines cannot accept a different environment, but for
now it is all or nothing).
Under normal circumstances this is handled transparently by the icecream daemon,
which will prepare a tarball with the environment when needed. This is the
recommended way, as the daemon will also automatically update the tarball
whenever your compiler changes.
If you want to handle this manually for some reason, you have to tell icecream
which environment you are using. Use icecc --build-native
create an archive file containing all the files necessary to setup the
compiler environment. The file will have a random unique name like
per default. Rename it to
something more expressive for your convenience, e.g.
in the shell environment where you start the compile jobs and the file
will be transfered to the daemons where your compile jobs run and installed to
a chroot environment for executing the compile jobs in the environment fitting
to the environment of the client. This requires that the icecream daemon runs
SUSE got quite some good machines not having a processor from Intel or AMD, so
icecream is pretty good in using cross-compiler environments similar to the
above way of spreading compilers. There the ICECC_VERSION variable looks like
for example like this:
How to package such a cross compiler is pretty straightforward if you look what
is inside the tarballs generated by icecc --build-native
When building for embedded targets like ARM often you will have a toolchain that
runs on your host and produces code for the target. In these situations you
can exploit the power of icecream as well.
Create symlinks from where icecc is to the name of your cross compilers (e.g.
), make sure that these symlinks
are in the path and before the path of your toolchain, with $ICECC_CC and
$ICECC_CXX you need to tell icecream which compilers to use for preprocessing
and local compiling. e.g. set it to ICECC_CC=arm-linux-gcc
As the next step you need to create a .tar.bz2 of your cross compiler, check the
result of build-native to see what needs to be present.
Finally one needs to set ICECC_VERSION and point it to the .tar.bz2 you have
created. When you start compiling your toolchain will be used.
With ICECC_VERSION you point out on which platforms your toolchain runs, you do
not indicate for which target code will be generated.
When working with toolchains for multiple targets, icecream can be configured to
support multiple toolchains in the same environment.
Multiple toolchains can be configured by appending =<target> to the
tarball filename in the ICECC_VERSION variable. Where the <target> is
the cross compiler prefix. There the ICECC_VERSION variable will look like
Below an example of how to configure icecream to use two toolchains,
/work/toolchain2/bin/arm-linux-androideabi-[gcc,g++], for the same host
- Create symbolic links with the cross compilers names (e.g.
arm-eabi-[gcc,g++] and arm-linux-androideabi-[gcc,g++]) pointing to where
the icecc binary is. Make sure these symbolic links are in the $PATH and
before the path of the toolchains.
- Create a tarball file for each toolchain that you want to
use with icecream. The /usr/lib/icecc/icecc-create-env script can
be used to create the tarball file for each toolchain, for example:
- Set ICECC_VERSION to point to the native tarball file and
for each tarball file created to the toolchains (e.g
With these steps the icecrem will use /work/arm-eabi-toolchain1.tar.gz
file to cross compilers with the prefix arm-eabi (e.g. arm-eabi-gcc
), use /work/arm-linux-androideabi-toolchain2.tar.gz
file to cross compilers with the prefix arm-linux-androideabi (e.g.
) and use
file to compilers without prefix, the native
The easiest way to use ccache with icecream is to set CCACHE_PREFIX to icecc
(the actual icecream client wrapper)
This will make ccache prefix any compilation command it needs to do with icecc,
making it use icecream for the compilation (but not for preprocessing alone).
To actually use ccache, the mechanism is the same like with using icecream
alone. Since ccache does not provide any symlinks in /opt/ccache/bin
you can create them manually:
ln -s /usr/bin/ccache /opt/ccache/bin/gcc
ln -s /usr/bin/ccache /opt/ccache/bin/g++
And then compile with
Note however that ccache is not really worth the trouble if you are not
recompiling your project three times a day from scratch (it adds quite some
overhead in comparing the preprocessor output and uses quite some disc space
and I found a cache hit of 18% a bit too few, so I disabled it again).
You can use the environment variable ICECC_DEBUG to control if icecream gives
debug output or not. Set it to debug to get debug output. The other possible
values are error, warning and info (the -v
option for daemon and
scheduler raise the level per -v
on the command line - so use
for full debug).
It is possible that compilation on some hosts fails because they are too old
(typically the kernel on the remote host is too old for the glibc from the
local host). Recent icecream versions should automatically detect this and
avoid such hosts when compilation would fail. If some hosts are running old
icecream versions and it is not possible to upgrade them for some reason, use
Numbers of my test case (some STL C++ genetic algorithm)
- g++ on my machine: 1.6s
- g++ on fast machine: 1.1s
- icecream using my machine as remote machine: 1.9s
- icecream using fast machine: 1.8s
The icecream overhead is quite huge as you might notice, but the compiler cannot
interleave preprocessing with compilation and the file needs to be
read/written once more and in between the file is transfered.
But even if the other computer is faster, using g++
on my local machine
is faster. If you are (for whatever reason) alone in your network at some
point, you lose all advantages of distributed compiling and only add the
overhead. So icecream got a special case for local compilations (the same
special meaning that localhost got within $DISTCC_HOSTS). This makes compiling
on my machine using icecream down to 1.7s (the overhead is actually less than
0.1s in average).
As the scheduler is aware of that meaning, it will prefer your own computer if
it is free and got not less than 70% of the fastest available computer.
Keep in mind, that this affects only the first compile job, the second one is
distributed anyway. So if I had to compile two of my files, I would get
- g++ -j1 on my machine: 3.2s
- g++ -j1 on the fast machine: 2.2s
- using icecream -j2 on my machine:
- (using icecream -j2 on the other machine:
The math is a bit tricky and depends a lot on the current state of the
compilation network, but make sure you are not blindly assuming make
halves your compilation time.
In most requirements icecream is not special, e.g. it does not matter what
distributed compile system you use, you will not have fun if your nodes are
connected through than less or equal to 10MBit. Note that icecream compresses
input and output files (using lzo), so you can calc with ~1MBit per compile
job - i.e. more than make -j10
will not be possible without
Remember that more machines are only good if you can use massive
parallelization, but you will for sure get the best result if your submitting
machine (the one you called g++
on) will be fast enough to feed the
others. Especially if your project consists of many easy to compile files, the
preprocessing and file I/O will be job enough to need a quick machine.
The scheduler will try to give you the fastest machines available, so even if
you add old machines, they will be used only in exceptional situations, but
still you can have bad luck - the scheduler does not know how long a job will
take before it started. So if you have 3 machines and two quick to compile and
one long to compile source file, you are not safe from a choice where everyone
has to wait on the slow machine. Keep that in mind.
A short overview of the ports icecream requires:
- TCP/10245 on the daemon computers (required)
- TCP/8765 for the the scheduler computer (required)
- TCP/8766 for the telnet interface to the scheduler
- UDP/8765 for broadcast to find the scheduler
If the monitor cannot find the scheduler, use
icecc-scheduler(1), iceccd(1), icemon(1)
Stephan Kulow <email@example.com>
Michael Matz <firstname.lastname@example.org>
Cornelius Schumacher <email@example.com>
...and various other contributors.