/icecream

Distributed compiler with a central scheduler to share build load

Primary LanguageC++GNU General Public License v2.0GPL-2.0

Build Status (Linux & macOS) Build Status (FreeBSD) Codacy Badge Code Quality: Cpp Total Alerts

Icecream was created by SUSE based on distcc. Like distcc, Icecream takes compile jobs from a build and distributes them among remote machines allowing a parallel build. But unlike distcc, Icecream uses a central server that dynamically schedules the compile jobs to the fastest free server. This advantage pays off mostly for shared computers, if you're the only user on x machines, you have full control over them.

Table of Contents

Installation

We recommend that you use packages maintained by your distribution if possible. Your distribution should provide customized startup scripts that make icecream fit better into the way your system is configured.

We highly recommend you install icemon or icecream-sundae with icecream.

If you want to install from source see the instructions in the README file provided in the source package.

How to use icecream

You need:

  • At least one machine that runs the scheduler ("./icecc-scheduler -d")
  • Many machines that run the daemon ("./iceccd -d")

It is possible to run the scheduler and the daemon on one machine and only the daemon on another, thus forming a compile cluster with two nodes.

If you want to compile using icecream, make sure $prefix/lib/icecc/bin is the first entry in your path, e.g. type

 export PATH=/usr/lib/icecc/bin:$PATH

(Hint: put this in ~/.bashrc or /etc/profile 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. As a start, take the number of logical processors multiplied with 2, or a larger number if your compile cluster can serve all the compilation jobs. But note that too large numbers may in fact make the build slower (for example if your local machine gets overloaded with preparing more jobs than it can handle at a time).

Here is an example:

 make -j6

WARNING: Never use icecream in untrusted environments. Run the daemons and the scheduler as unprivileged 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" (from a separate repository/package).

make it persistent

If you restart a computer, you still want it to be in the icecream cluster after reboot. Consult your distribution's documentation on this. If you use packages provided by your distribution this should be automatic (or a simple configuration change).

make scheduler persistent:

By adding options --scheduler-host for daemon and --persistent-client-connection for scheduler, the client connections are not disconnected from the scheduler even if there is an availability of better scheduler.

TroubleShooting

Most problems are caused by firewalls and by make using the wrong C compiler (e.g. /usr/bin/gcc instead of /usr/lib/icecc/bin/gcc).

Firewall

For testing purposes, you can stop your firewall like this:

 rcSuSEfirewall2 stop

To open the right ports in your firewall, call

 yast2 firewall

Choose "allowed services" -> Advanced. Enter for TCP 10245 8765 8766 and for UDP 8765.

If you have the scheduler running on another system, you should open broadcasting response:

 yast2 firewall

Choose "Custom Rules" -> Add. Enter Source Network 0/0, Protocol UDP, Source Port 8765.

C compiler

To make sure your compile job uses /usr/lib/icecc/bin/gcc (gcc is used as an example here, depending on your compile job it can also be g++, cc or c++) start your compile using

 make VERBOSE=1

and wait for a typical compile command to appear, like this one:

 cd /root/kdepim/kode/libkode && /usr/lib/icecc/bin/c++  -DTest1Area=5121 -D_BSD_SOURCE
 -D_XOPEN_SOURCE=500 -D_BSD_SOURCE -DQT_NO_STL
 -DQT_NO_CAST_TO_ASCII -D_REENTRANT -DKDE_DEPRECATED_WARNINGS
 -DKDE_DEFAULT_DEBUG_AREA=5295 -DMAKE_KODE_LIB -Wnon-
 virtual-dtor -Wno-long-long -ansi -Wundef -Wcast-align
 -Wchar-subscripts-Wall -W -Wpointer-arith -Wformat-security
 -fno-exceptions -fno-check-new

In this example, the right c compiler is chosen, /usr/lib/icecc/bin/c++. If the wrong one is chosen, delete CMakeCache.txt (if existing) and start the build process again calling ./configure (if existing).

osc build

You can tell osc build to use icecream to build packages by appending --icecream=<n> where n is the number of processes which should be started in parallel. However, for integration with icecream to work properly, you must install icecream on the host where you will run "osc build" and you must start icecream daemon.

some compilation nodes aren't used

If, when using icecream monitor (icemon), you notice some nodes not being used at all for compilation, check you have the same icecream version on all nodes, otherwise, nodes running older icecream versions might be excluded from available nodes.

build with -Werror fails only when using icecream

This problem should not exist with a recent icecream version. If it does, try using ICECC_REMOTE_CPP=1 (see icecc --help).

clang tries to read /proc/cpuinfo and fails

This is a problem of clang 4.0 and newer: https://bugs.llvm.org/show_bug.cgi?id=33008. The most recent Icecream version works around this problem.

Supported platforms

Most of icecream is UNIX specific and can be used on most platforms, but as the scheduler needs to know the load of a machine, there are some tricky parts. Supported are:

  - Linux
  - FreeBSD
  - DragonFlyBSD
  - OS X

Note that all these platforms can be used both as server and as client - meaning you can do full cross compiling between them.

The following platforms are known to work at least as a client, meaning that you can run compilation on them that will compile on remote nodes using cross compilation.

  - Cygwin

Using icecream in heterogeneous environments

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. In the future icecream might gain the ability to know when machines can't accept a different env, 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

to create an archive file containing all the files necessary to setup the compiler environment. The file will have a random unique name like "ddaea39ca1a7c88522b185eca04da2d8.tar.bz2" per default. Rename it to something more expressive for your convenience, e.g. "i386-3.3.1.tar.bz2". Set

  ICECC_VERSION=<filename_of_archive_containing_your_environment>

in the shell environment where you start the compile jobs and the file will be transferred 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 as root.

Cross-Compiling using icecream

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 <native_filename>(,<platform>:<cross_compiler_filename>)*, for example like this:

 /work/9.1-i386.tar.bz2,ia64:/work/9.1-cross-ia64.tar.bz2,Darwin_PowerPCMac:/work/osx-generate-i386.tar.gz

To get this working on openSuse machines there are some packages containing the cross-compiler environments. Here is a sample case showing how to do to get it working. Let's assume that we want to build for x86_64 but use some i386 machines for the build as well. On the x86_64 machine, go to http://software.opensuse.org, search for icecream x86_64 and download and install the version for i586. Then add this to the ICECC_VERSION and build.

 i386:/usr/share/icecream-envs/cross-x86_64-gcc-icecream-backend_i386.tar.gz

Creating cross compiler package

How to package such a cross compiler is pretty straightforward if you look what's inside the tarballs generated by icecc. You basically need a /usr/bin/gcc, a /usr/bin/g++ and a /usr/bin/as. So if you need a cross compiler that uses your OS X running G5 to compile i586-linux for your laptop, you would:

  • go to your OS X and download binutils and gcc (of the versions you use on linux)

  • first compile and install binutils with --prefix /usr/local/cross --target=i586-linux (I have some problems that required setting CC and AR)

  • configure gcc with the same options, go into the gcc directory and make all install-driver install-common - that worked good enough for me.

  • now create a new directory where you copy /usr/local/cross/bin/i586-linux-{gcc,g++,as} into as usr/bin/{gcc,g++,as}

  • now I copy an empty.c (that is empty) into that dir too and call

    chroot . usr/bin/gcc -c empty.c

that will report an error about missing libraries or missing cc1 - copy them until gcc generates an empty.o without error. You can double check with "file empty.o" if it's really a i586-linux object file.

  • now tar that directory and use it on your client as specified above.

My cross compiler for the above case is under http://ktown.kde.org/~coolo/ppc-osx-create-i586.tar.gz

Cross-Compiling for embedded targets using icecream

When building for embedded targets like ARM often you'll 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 symbolic links from where icecc is to the name of your cross compilers (e.g. arm-linux-g++ and arm-linux-gcc), make sure that these symbolic links 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 and ICECC_CXX=arm-linux-g++.

As the next step you need to create a .tar.bz2 of your cross compiler, check the result of icecc --build-native to see what needs to be present.

Finally one needs to set ICECC_VERSION and point it to the tar.bz2 you've created. When you start compiling your toolchain will be used.

NOTE: with ICECC_VERSION you point out on which platforms your toolchain runs, you do not indicate for which target code will be generated.

Cross-Compiling for multiple targets in the same environment using icecream

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 <native_filename>(,<platform>:<cross_compiler_filename>=<target>)*.

Below an example of how to configure icecream to use two toolchains, /work/toolchain1/bin/arm-eabi-[gcc,g++] and /work/toolchain2/bin/arm-linux-androideabi-[gcc,g++], for the same host architecture:

  • 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. icecc-create-env script can be used to create the tarball file for each toolchain, for example:

    • icecc-create-env /work/toolchain1/bin/arm-eabi-gcc
    • icecc-create-env /work/toolchain2/bin/arm-linux-androideabi-gcc
  • Set ICECC_VERSION to point to the native tarball file and for each tarball file created to the toolchains (e.g ICECC_VERSION=/work/i386-native.tar.gz,/work/arm-eabi-toolchain1.tar.gz=arm-eabi,/work/arm-linux-androideabi-toolchain2.tar.gz=arm-linux-androideabi).

With these steps the icecream will use /work/arm-eabi-toolchain1.tar.gz file to cross compilers with the prefix arm-eabi (e.g arm-eabi-gcc and arm-eabi-g++), use /work/arm-linux-androideabi-toolchain2.tar.gz file to cross compilers with the prefix arm-linux-androideabi(e.g. arm-linux-androideabi-gcc and arm-linux-androideabi-g++) and use /work/i386-native.tar.gz file to compilers without prefix, the native compilers.

How to combine icecream with ccache

The easiest way to use ccache with icecream is to set CCACHE_PREFIX to icecc (the actual icecream client wrapper):

 export CCACHE_PREFIX=icecc

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:

 mkdir /opt/ccache/bin
 ln -s /usr/bin/ccache /opt/ccache/bin/gcc
 ln -s /usr/bin/ccache /opt/ccache/bin/g++

And then compile with

 export PATH=/opt/ccache/bin:$PATH

In this case icecc's symlinks in /usr/lib/icecc/bin should not be in your path, as CCACHE_PREFIX is instructing ccache to explicitly delegate to icecc rather than finding it in the path. If both ccache and icecc's symlinks are in the path it is likely the two wrappers will mistake each other for the real compiler and icecc will complain that it has recursively invoked itself.

Note however that ccache isn't really worth the trouble if you're not recompiling your project three times a day from scratch (it adds some overhead in comparing the source files and uses quite some disk space).

Debug output

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 -vvv for full debug).

Some Numbers

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 can't interleave preprocessing with compilation and the file needs to be read/written once more and in between the file is transferred.

But even if the other computer is faster, using g++ on my local machine is faster. If you're (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's 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: max(1.7,1.8)=1.8s
  • (using icecream -j2 on the other machine: max(1.1,1.8)=1.8s)

The math is a bit tricky and depends a lot on the current state of the compilation network, but make sure you're not blindly assuming make -j2 halves your compilation time.

What is the best environment for icecream

In most requirements icecream isn't special, e.g. it doesn't matter what distributed compile system you use, you won't have fun if your nodes are connected through less than or equal to 10MBit. Note that icecream compresses input and output files (using lzo), so you can calculate with ~1MBit per compile job - i.e more than make -j10 won't be possible without delays.

Remember that more machines are only good if you can use massive parallelism, 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 IO 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 doesn't 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 files, you're not safe from a choice where everyone has to wait on the slow machine. Keep that in mind.

Icecream is very sensitive to latency between nodes, and packet loss. While icecream has been successfully used by people who are on opposite sides of the earth, when those users were isolated to their geographic location the speed improved for everyone. In most corporate environments within a single building everything works well, but between two buildings often is troublesome.

If you plan to use Icecream in the cloud or anywhere else you would have more latency than a corporate LAN, you should strongly consider using a dedicated scheduler configured with one of the alternative scheduling algorithms. SeeSome advice on configuration for details on alternative scheduling algorithms. You should also consider using dedicated compile servers as well if at all practical.

Some advice on configuration

Icecream supports many configurations but you need to understand your network to choose what is right for you.

You should ensure that the scheduler is up to the latest version. Many new features require the client and the scheduler to work together to use them. Even though clients should work with old schedulers new features will not work, and may not be disabled correctly.

Version 1.1 gained the ability for multiple schedulers on a single network to decide on the best master. However daemons running earlier versions do not understand this, and it is random if they will find the correct one. In all other ways it is believed that mixing old and new versions of the daemon will work: if you use a new feature only new clients will be used.

Recommended is to start the scheduler and daemon on everybody's machine. The icecream schedulers will choose one to be the master and everyone will connect to it. When the scheduler machine goes down a new master will be selected automatically.

If you need to run mixed icecream versions, then it is best to designate one machine on each subnet to be a scheduler. Icecream nodes will automatically find the scheduler and connect to it. If someone accidentally starts a second scheduler this will cause problems with clients that are older than version 1.1, but they should eventually work. The scheduler should be a reliable machine, but if it fails you use any existing machine as a replacement.

You may also designate a scheduler machine, and then for each client specify the scheduler to use (this is a variation of the previous case). You need to ensure that there is no other schedulers on the same network as this scheduler if you do this. The scheduler machine MUST be reliable, any failure will require reconfiguring all client machines. This setup allows you to specify one scheduler per building which is useful if single developers are scattered around. If you do this check with IT to ensure that icecream traffic won't overload routers.

You might designate a netname. This is useful if your network is using VPN to make it seem like developers who are physically a long distance apart seem like they are on the same sub-net. While the VPNs are useful, they typically do not have enough bandwidth for icecream, so by setting a different netname on each side of the VPN you can save bandwidth. Netnames can be used to work around some limitations above: if a netname is set icecream schedulers and daemons will ignore the existence of other schedulers and daemons.

Finally, you can configure the scheduler to use a different job scheduling algorithm to distribute the load across your compile servers. The default scheduling algorithm is "fastest"; however this may not actually be the fastest algorithm to complete a full build in the real world, so you should try multiple algorithms to find the best for your environment.

Currently, you can choose from the following algorithms:

  • random: distribute jobs randomly across the network.
  • round_robin: schedule jobs on the host that has seen jobs least recently. This scheduler tends to distribute job assignments evenly across the network, but it may distribute actual load unevenly and works best for homogeneous networks of dedicated compile servers.
  • least_busy: schedule jobs to the host with the highest percentage of open job slots. This scheduler tends to distribute load evenly, but tasks may be distributed unevenly and works best for heterogeneous networks of dedicated compile servers.
  • fastest: schedule jobs on the host expected to complete it soonest. This tends to favor using a few hosts in the network, though a portion of the load will be allocated to new hosts and hosts that have not been used in some time to build up statistics. This is the classic algorithm and is suitable for relatively small and heterogeneous networks of compile servers.

Network setup for Icecream (firewalls)

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 (optional)
  • UDP/8765 for broadcast to find the scheduler (optional)

Note that the SuSEfirewall2 on SUSE < 9.1 got some problems configuring broadcast. So you might need the -s option for the daemon in any case there. If the monitor can't find the scheduler, use USE_SCHEDULER=<host> icemon (or send me a patch :)

I use distcc, why should I change?

If you're sitting alone home and use your partner's computer to speed up your compilation and both these machines run the same Linux version, you're fine with distcc (as 95% of the users reading this chapter will be, I'm sure). But there are several situations, where distcc isn't the best choice:

  • you're changing compiler versions often and still want to speed up your compilation (see the ICECC_VERSION support)
  • you got some neat PPC laptop and want to use your wife's computer that only runs intel (see the cross compiler section)
  • you don't know what machines will be on-line at compile time
  • most important: you're sitting in an office with several co-workers that do not like if you overload their workstations when they play doom (distcc doesn't have a scheduler)
  • you like nice compile monitors :)

Icecream on gentoo

  • It is recommended to remove all processor specific optimizations from the CFLAGS line in /etc/portage/make.conf. On the aKademy cluster it proved useful to use only "-O2", otherwise there are often internal compiler errors, if not all computers have the same processor type/version

Be aware that you have to change the CFLAGS during each gcc update too.

  • Create soft link for CHOST gcc/g++ e.g. ln -s /opt/icecream/bin/icecc /opt/icecream/libexec/icecc/bin/x86_64-pc-linux-gnu-gcc; ln -s /opt/icecream/bin/icecc /opt/icecream/libexec/icecc/bin/x86_64-pc-linux-gnu-g++
  • To use icecream with emerge/ebuild use PREROOTPATH="/opt/icecream/libexec/icecc/bin" FEATURES="-network-sandbox" emerge bla
  • Be aware, because your gcc/glibc/binutils are normally compiled with processor-specific flags, there is a high chance that your compiler won't work on other machines. The best would be to build gcc, glibc and binutils without those flags and copying the needed files into your tarball for distribution, e.g. CFLAGS="-mcpu=i686 -O3 -fomit-frame-pointer -pipe" CXXFLAGS="$CFLAGS" ebuild /usr/portage/sys-devel/gcc-yourver.ebuild install ; cp /var/tmp/portage...

Bug tracker

Create a github issue on https://github.com/icecc/icecream

Repository

The git repository lives at https://github.com/icecc/icecream

Mailing list

icecream-users@googlegroups.com