Distributed Compilation

BagOBones writes "Tired of waiting for your source to compile? Dreaming of having your own cluster and having something useful to do with it? Well Trolltech might have the answer. Trolltech Teambuilder lets you turn your network into a clustered C/C++ compiler."

Comments

I have nothing against the product

[ObviousGuy]

But isn't the point of having subscriptions to avoid advertisements?

That said, I've certainly worked on products that needed at least 3 hours to build an entire tree. These builds were done regularly around 3 in the morning so that the daily drop was available to QA first thing in the morning. It wasn't really necessary to farm out the compilation across machines because it wasn't a big deal to maximize speed.

As for developers' machines, it wasn't like every change was accompanied by a full build. You recompile the files that changed and link the object files together. Any smart build system should be able to handle this type of logic. Such a local build would take about 2 or 3 minutes (if that. This time could be made even shorter by using dlls instead of a single binary). I guess this Trolltech system could reduce this wait even further, but I'm not sure I see the point.

When else are you going to refill your coffee?

Re:I have nothing against the product

[Anonymous Brave Guy]

As for developers' machines, it wasn't like every change was accompanied by a full build. You recompile the files that changed and link the object files together.

Sure. OTOH, the project I work on has well over a million lines of C++ code in it, about five developers full time, and a pretty tidy OO design is used to model much of our stuff. If you're playing with a new feature, this quite often goes into a base class to add it across a whole range of areas of the code, and changing the wrong header file can cause a good 10-15 minutes of rebuild (the whole project on a fast machine takes ~1hr). It's the sort of thing that's not normally a problem, but if you do it 10 times in a day while trying to fix that nagging bug, and you're shipping within the week so time is of the essence, then this sort of functionality is priceless. I guess the kind of team where I work is (aside from developing under Windoze) exactly the sort of place they're targetting. To our clients paying a significant sum of money per hour for our time, the $750 to save several hours at crunch time is nothing.

(And before anyone mentions it, yes, I'm aware of the Cheshire Cat idiom and such, thanks. They aren't in common use on the project, which dates from several years ago, and retrofitting them would be difficult to do without a lot of risk. It's cheaper and safer to just speed up build resources, by using faster PCs or something like the product in question here.)

Why not dmake?

[wickidpisa]

What does this package have that makes it any better than dmake(distributed make)? Convincing anyone that uses linux to go with a $750 product over an open source one seems silly, if they were going to do that they wouldn't be using linux in the first place.

There is a shiney new solution to this

[Tablizer]

use an interpreter. (Okay, it is not new. I lied.)

BTW, is there "lint"-like utilities for any scripting/interpreted languages that detects "suspicious" stuff before running?

This may help one get the best of both worlds.

I personally prefer interpreted languages. They are usually easier to read IMO because there is less formality in the way.

Interpreted languages attempt to reduce problems by making the view clearer, while compiled languages attempt reduce problems by using a fatter air-bag.

It is sort of like chosing between a fighter plane that is highly menueverable by being light, or one that is heavily armored. They are both different strategies to the same goal. The choice often depends on the pilot and training.

ccache

[ryants]

A similar idea that I've been playing around with and seems to work is to use ccache and put the cache on a network drive.

That way, if someone has already built the file you're about to build, it's just a copy. Even better is if you have a continous build script caching results ahead of time for you.

It has worked pretty well for me so far.

Re:ccache

[elmartinos]

ccache is pretty good, but not especially made for distributed compilation, but there is another project at samba.org:
distcc.

/elmartinos

free with PVMGmake

[bohnsack]

You can do this for free with PVMGmake. I have a HOWTO

I don't know about this

[Henry V .009]

Program compilation is not an obvious candidate for parallelization because it is an extremely memory intensive process. So to do this right you'd need very fast communication between your CPUs.

However, what Trolltech is probably doing, is farming out the independant bits of your program that you would link together in the end anyway. That might really help on rebuilding the entire project at once. However, if your program is set up well, you usually only need to recompile one file at a time.

So the times when you could see the biggest speed up in compilation, you will need this least.