Slashdot Mirror
A Distributed Front-end for GCC
format writes "distcc is a distributed front-end for GCC, meaning you can compile that big project across n number of machines and get it done almost n times as fast.
The machines don't have to be identical or be running the exact same GCC version, but having the same OS is helpful." With the advent of faster hardware, I can't complain about kernel compile times anymore, but larger source trees could definitely benefit from this.
The machines don't have to be identical or be running the exact same GCC version
Well, to some extent they probably do. If you're running GCC 3.2 on one, you wouldn't be able to run 3.0 on another because of binary incompatibility.
Yay! My 133 doesn't have to take 25 billion years to compile anymore! Uhm, wait, I don't have any other computers... Shoot.
The sun compiler suite comes with dmake, which does the same on the level of make, rather than cc, but is essentially the same.
Definitelly would make beowulf clusters interesting for compilation as well as hard core numerics (no joke intendend).
You can almost never achevie a speed up of N. You can acheive S(N) = T(1)/(T(1)*alpha+((1-alpha)*T(1))/N+T0) Where T(1) is the time it takes to run the task with 1 computer, alpha is the part of the task that cannot be parallelised (as in startup registers etc.) and T0 is the communications overhead of the task.
:)
Just to clarify.
Sure, kernel compiles are fairly speedy but large projects still take forever to compile. Even on my AthlonXP 1900+. I downloaded the latest CVS snapshots of KDE in an attempt to de-bluecurve my new RH 8.0 machine and it took a couple hours to compile everything. Depressing.
This could really spur the development of OpenOffice.
With 50, 100 machines or so hooked up, OpenOffice's compile time could be reduced to as little as 1 or 2 days!!!
You can almost never achevie a speed up of N. You can acheive S(N) = (1)/ (T(1)*alpha+((1-alpha)*T(1)) / N+T0) Where
T(1) is the time it takes to run the task with 1 computer, alpha is the part of the task that cannot be parallelised (as in startup registers etc.) and T0 is the communications overhead of the task.
This is the text book. Amdahls law, IIRC.
In reality, and also in most text books, there are exceptions where the solution scales with the number of processes.
And it should be easy enough to see: 5 machines compiling one source file each are 5 times as fast as one machine compiling 5 source files.
As long as you start gcc 5 times in a row you have
the same initialization overhead for EACH instance of gcc one after the other.
If you manage to start gcc with a couple of source files as argument to compile you save the laoding time of the binary at least. That would correspondend roughly to the alpha value.
Amdahls law is usefull for a single program/problem: try to paralelize gcc and you find the compiling source can't get speed up very much. So 5 processors running several threads of one gcc instance, those do not scale by 5.
However it says nothing about just solving the same problem multiple times in parallel.
Regards,
angel'o'sphere
Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
I sincerely hope Apple makes this feature into projectbuilder, which compiles insanely slow when compared to codewarrior. If it wasn't for the superior interface and integration with interface builder, I'd swap back to codewarrior right away.
Does anyone here know how good the speed increase is when compiling on dual G4s versus a single proc ?
When will I end this grieving ? When will my future begin ?
Is this better than say, Group Compiler?
Could someone please point out the difference between a parallel and/or distributed make, like pmake?
It sounds not realy reasonable to put the coding work into gcc when you like to have yacc/bison and a bunch of perl scripts and what ever else you have in your makefile also speeded up.
Regards,
angel'o'sphere
Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
From the FAQ:
distcc doesn't care. However, in some circumstances, particularly for C++, gcc object files compiled with one version of gcc are not compatible with those compiled by another. This is true even if they are built on the same machine.
It is usually best to make sure that every compiler name maps to a reasonably similar version on every machine. You can either make sure that gcc is the same everywhere, or use a version-qualified compiler name, such as gcc-3.2 or gcc-3.2-x86-linux.
So in other words, keep them close, especially for gcc versions that break backwards capability.
You think that I'm crazy, you should see this guy!
I thought that the binary incompatibility was only a C++ thing. So for some projects, that's an issue, but not for all. Of course, the idea of being careless about which compiler version you're using for building a large project is rather strange.
I think the biggest plus is that you can have one hella fast machine on your network running distcc that basically does all your compiling for all your other machines. I can see this being a big bonus for server farms like rackspace.com. The customers would be getting compile speeds from a big ass server, rather than just their little dinky Duron.
~LoudMusic
No sig for you. YOU GET NO SIG!
Whether you're looking to install Gentoo on a old pentium to use as a router or sacrificing your first born to compile KDE, it should make things go quite a bit faster.
Well unless every computer you own runs Gentoo you want to emerge world.
The C ABI between *all* GCC versions (and probably other compilers too) are compatible. You can compile libgnome using GCC 2.95.2 and Nautilus using GCC 3.2 and not have any problems at all.
If your system is well designed, compiling the entire thing should be a rare event. In a well designed project, most changes occur in c files or in headers only included in a few c files, so most changes only require compiling a very few files.
Compiling the whole source tree should be the sort of thing you do fairly rarely (for a big project), perhaps once a night, perhaps automated so no one has to watch it.
If compile time is something that is a significant problem for you, you really need to look at your code design.
The cake is a pie
From the ccache homepage, which is also a Samba hosted project
ccache is a compiler cache.
It acts as a caching pre-processor to C/C++ compilers, using the -E compiler switch and a hash to detect when a compilation can be satisfied from cache.
This often results in a 5 to 10 times speedup in common compilations.
Ahem. Amdahl's law still operates, and you even say so yourself. There's a constant part that cannot be removed. Let's say it takes 50 msec to initialize gcc and 500 msec to compile the average source file. Then it takes 5.05 sec to compile ten files with one copy of gcc. Ignoring commiunications, it takes 0.550 seconds to compile them on ten machines. Is 5.05/0.550 == 10? No, it's about 9.2. Therefore, the speedup is LESS THAN N. Note that the faster the actual compile time, the lower the speedup would be!
Cantankerous old coot since 1957.
Running different versions could cause really nasty problems if different versions of gcc support different levels of C (like C99 or older C) or if one version has a compiler bug that another doesn't.
Can you imagine code compiling or failing to compile randomly depending on which machine happens to compile it? Yikes! Debugging nightmare...
The cake is a pie
Oh well, I've been trying to get a distributed kernel compilation system working using Platform's LSF. Guess I can throw that project away. ;)
-- Minds are like parachutes... they work best when open.
i've recently written some dist compile tool using a different approach after i've been using distcc for half an hour ...
the big problem with distcc is that it does all the preprocessing on one machine, which is really an overkill in some situations (and limits the total speed increase one can gain).
what i've done is first running a modified version of make and then distribute all the objects which have to be compiled to the machines.
everything is done on these machines (including preprocessing) which only limits the number of compile machines by the speed of the network (i've been compiling on 60 hosts, with almost linear speed increase).
the only problem involved with this approach is that the same compiler has to be installed on these machines, and that they have to write on some sort of network shared filesystem (for the objects)
the same compiler is an easy thing since i've been using the intel compiler version 6 and common system includes (i've put them into a shared include dir)
any thoughts about this thing ? (or some folks willing to help me create a version basing on gpl stuff so i could release this one ?)
looks like this one is not necessarily a good idea to run on a university workstation cluster...
1.4 Security Considerations
distcc should only be used on networks where all machines and all users are trusted.
The distcc daemon, distccd, allows other machines on the network to run arbitrary commands on the volunteer machine. Anyone that can make a connection to the volunteer machine can run essentially any command as the user running distccd.
distcc is suitable for use on a small to medium network of friendly developers. It's certainly not suitable for use on a machine connected to the Internet or a large (e.g. university campus) network without firewalling in place.
inetd or tcpwrappers can be used to impose access control rules, but this should be done with an eye to the possibility of address spoofing.
In summary, the security level is similar to that of old-style network protocols like X11-over-TCP, NFS or RSH.
time is a funny concept
Once upon a time, Symantec had a C++ compiler, and with version 7.5 (1996), the build process could be spread all over a network. This did speed up compilation times as machines that were running the build service that were more or less idle would be sent files to compile, passing back the objects and binaries as oppropriate.
Oh, by the way, that compiler is now called Digital Mars C++.
That said, all the machines on the network had to be running Windows (and at that time, I think only Windows 95 or NT were the only choices available for that compiler). Further, all had to have the same version of the compiler.
For those of us that are running Linux boxes on a primarily Windows network, this system, whether GCC or something else, would be rather hard to implement without a cross-compiler. Additionally, even if all were Linux workstations (or BSD, or Solaris, etc etc etc) wouldn't binary compatibility be driven by not just the version of the compiler but the target OS as well?
It's a noble undertaking. I hope that the developers are putting thought into all the little things like this that will make it tough to pull off.
Be excellent to each other. And... PARTY ON, DUDES!
This is true. Unfortunately, traditional makefiles tend to encourage compiling each file seperately, so you have to use workarounds like ICC's .il file mechanism to do global optimizations. However, for developers, this distributed processing is a big boost. When you're working on code, you have to recompile a project repeatedly, and distributing the workload pays off in decreased frustration. For those intermediate builds, optimizations don't really matter anyway.
A deep unwavering belief is a sure sign you're missing something...
gcc is not a fast C compiler. It is a portable C compiler and it makes pretty good code, but it is not fast. g++ is a slow C++ compiler.
The Codewarrior compilers were fast compilers that made pretty good code.
Its all in the tradeoffs.
You can't have it all. Apple has been adding engineering hours to improve speed with gcc, but gcc will always value source language selection and target portability above all else.
Emphasis is mine.
I guess it all depends on whether or not you want to work with production quality code or not.
using distcc and ccache
from the above link:
distcc & ccache
Has anybody yet thought of integrating distcc & ccache?
Yes, of course. They work pretty well as separate programs that call each other: you can just write CC='ccache distcc gcc'. (libtool sometimes has trouble with that, but that's a problem that we will fix shortly and it is a separate question.)
This is very nearly as efficient as making them both part of a single program. The preprocessor will only run once and the preprocessed source will be passed from ccache straight to distcc. Having them separate allows them to be tested, debugged and released separately.
Unless there's a really strong argument to do otherwise, having two smaller programs is better practice, or at least more to my taste, than making one monolithic one.
I'd love a speedup, but the time I'd save compiling would be wasted on having to fully install another linux box. Being able to boot a CD with a live linux distro and this software, and then connect to these slave machines to help compile would be immensely helpful. My linux box is a Cyrix 200MHz PC. Being able to stick a CD into my Athlon 1800 to help the compile would be fabulous.
There is an alternative ( http://gecc.sf.net). gecc has a little different approach, it has a central component that distributes the compilation to a number of compile nodes. The set of compile nodes may change (over time). That is: compile nodes may come and go.
gess is work in progress, distcc is much more mature, but maybe you like to take a look at gecc also.
(yes, gecc is my baby)
Amdahl's law still operates, and you even say so yourself. There's a constant part that cannot be removed. Let's say it takes 50 msec to initialize gcc and 500 msec to compile the average source file. Then it takes 5.05 sec to compile ten files with one copy of gcc.
Then you go on to tell how using ten machines provides only a 9.2-fold speedup. But what about a project with 100 files? It would take 50.05 seconds to build everything on one machine, and it takes 5.050 seconds to build ten files on each machine. Now we have a 50.05/5.050 == 9.92 fold speedup. In practice, can you notice the difference between 9-fold and 10-fold speedup?
Does the speedup factor not approach the number of machines asymptotically?
(How can I "Use the Preview Button!" when an accidental Enter keypress in the Subject invokes the Submit button? Scoop gets it right by setting Preview as the default button.)
Will I retire or break 10K?
I followed the 30second install overview, and although it took 20minutes to get working, that 20minutes included a distributed kernel compile across 2 of my systems here: a p3-500 and p3-700, 512meg ram each. very nice. i like it!
Why so negative? They are aware of the gdb issue and are working to fix it. This does not make distcc 'worthless' by any stretch.
Even if you do have a crash gdb will still give you the stack trace of a distcc-compiled -g program complete with function names and line numbers - you just don't get the source code without setting a directory directive within a gdb session. Big deal. It's perfectly usuable.
Sometimes I'll launch production builds mid-day to correct - wait for it - mid-day production problems. This has to be done as quickly as possible, and distcc is very useful.
The gdb thing is by no means a showstopper.
Of course, I'm assuming the developer is smart enough to have a proper build system set up. But when a program is in active development, you'll often have modified files that can take a significant amount of time to recompile. And god forbid you modify a common header. Even if you add one constant to support one source file, almost all build systems will recompile every single file that includes that header.
A deep unwavering belief is a sure sign you're missing something...
This was on Sweetcode months ago.
My most common problem these days is *not* compile time, but *link* time. Would be nice if there was a way to speed that up somehow.
-Bill
SlashSig Karma: Excellent (mostly affected by moderatio
I agree that gcc has different goals, but comeon, the speed difference is really flabbergasting ! codewarrior is twice as fast ! I can understand a speed loss due to portability, but that much ??
Anyways, it doesn't all matter that much if there were to be a speed gain from dual procs or distributed compiling. So to restate the Q : do you have any idea if PB is faster on a dual1GHz than on a signle 1GHz ?
When will I end this grieving ? When will my future begin ?
If you have headers included everywhere that have to be changed often, you've got a bad design.
You'd be amazed at how often it is easy to get rid of those stupid "include everywhere" headers.
The cake is a pie
Later, we upgraded to Motorola 88k. The new system came with a GreenHills compiler - which was so buggy it was unusable. The bugs were acknowleged, but never fixed. It was too buggy to compile gcc - I wasted many days simplifying expressions by hand to no avail.
My solution was to build a cross compiler on the 68k, and run it with a stub on the 88k for cc1 that fed the preprocessed source to the compiler on the 68k, and got back the assembler source. (There was no 88k support yet in gcc, so I had to write my own machine model. It was incompatible with Greenhills in passing and returning structs.) The preprocessor, assembler, and linker would run on the 88k - only the actual compiler pass ran on the 68k. This worked beautifully to build gcc on the 88k! And our 33Mhz 88k was so much faster, that I built a 68k cross compiler for the 88k and speeded up compiles on the 68k a great deal.
Next we upgraded to Power PC. AIX comes with no compiler at all! (IBM's compiler is very good - but expensive.) Fortunately, it comes with an assembler and linker. By dint of copying headers from AIX, and hand preprocessing and compiling to assembler, I got the preprocessor running on AIX. Then it was simple to run the cc1 stub again to compile up a native gcc for AIX. The new PowerPC was so much faster than the old 88k, that a cross compiler to speed up 88k compiles was in order also. (I contributed AIX shared lib support for gcc.)
Recently, I fired up a PowerPC cross compiler on our 600Mhz PII running Linux to speed up compiles on our old 100Mhz PowerPC AIX using the same simple technique. By using the -pipe option to gcc, the compiler on Linux runs in parallel with the preprocessor and assembler on AIX - truly efficient.
In conclusion, I want to thank Richard Stallman and the FSF for making it possible to rise above the stupid problems caused by closed source. Before the 68k, we had SCO Xenix. Basic utilities like tar and cpio were broken, reported, and never fixed - but GNU software was there with solid and reliable replacements. (Yes, I made donations and even ordered a tape.)
I spent a couple of weeks tinkering with the idea a while back and wrote a server script in Perl that can run on any Win32 box (no compilers or anything necessary to be installed), and wrote a client script that plugs into MSVC++'s IDE. The client parses the build commands and contacts all servers through a UDP broadcast, then connects to each server, preprocesses a file at a time, transfers the processed C/C++ file and the compiler and its related binaries to the server, then sends the build commands to the server as well. All output is captured on the server and sent back to the client.
.cpp files together and only compiled that one file every time, so their build was a fixed (short) cost, and never hit any header twice. Dirty code that has module-local statics would choke on that technique, but for good code, it's prolly smarter than distributing it.
It worked pretty well, except I had a lot of problems getting fork to work right in Perl on my Win32 box without crashing, so I could get good parallelism. It even fell back to using the local machine in the event of a failure remotely at any point, so on a multiprocessor development machine and no servers to connect to, it would actually use all the processors to compile--something MSVC doesn't do normally.
The nice thing about the server script is it only has three functions: accept a file, send a requested file, and execute an arbitrary command. So it doesn't really care about what it's doing. At work, we're planning to use it to leech some cycles off the receptionist machines and managers' boxen (we all know they don't do anything all day but email anyway!).
The real limiting factor is that preprocessing is relatively slow due to seeks on the hard drive, where not all the headers fit in the disk cache at once. This is a bigger bottleneck than you'd believe.
For what it's worth, I've learned of another company that concatenated all their
Any connection between your reality and mine is purely coincidental.
Unfortuantely GCC is really poor at optimizing in any case, so it's pretty much moot.
Just compare GCC at max optimization and SUN cc at minimum or first level. SUN cc beats GCC even there.
So while it's true that more data -> better optimization is correct I don't think it's a very pressing issue for GCC front ends.
gmake has parallel makes. See the -j option.