[Scons-users] Slowness of SCons with a large number of targets

[Manish Vachharajani]

One thing I noticed a number of years ago In regard to this is having a
node with a large out-degree (a single dependence node with lots of
targets) was much worse than many nodes with few targets.  I wrote a
builder that would take such nodes and break them up into a graph that
divided the sources and targets among a tree of many nodes each with a
small number of targets.  The actual builder would only run once but scons
thought there were many nodes and builders.  This made things much faster.
Unfortunately, the code is locked up at my prior company.

In your tests, are you creating such builders with a very large number of
targets for a single builder?

Manish

On Fri, Apr 14, 2017 at 4:53 AM, Hans Ottevanger <hans.ottevanger at gmail.com>
wrote:

> Hi,
>
> At the company I am working for we are revising our software build
> system and we are currently evaluating build tools. We think that
> SCons is an interesting tool, offering a high degree of consistency
> and features that others are clearly missing.
>
> However, we are facing severe performance issues with an increasing
> number of targets. Experimenting with close to 300000 targets in a
> tree that mimics a large part of our actual tree, we measured times
> for a full build of about 35 hours. We just touched the target files
> and did not invoke real compilers, so those 35 hours are mainly
> overhead from using SCons.
>
> We are aware of the fact that Eric Melski already reported scalability
> issues quite some time ago (see
> https://blog.melski.net/2013/12/11/update-scons-is-still-really-slow/).
> We could almost exactly reproduce Erics results using the tools he
> provides on GitHub (https://github.com/emelski/scons_bench). We are
> using SCons 2.5.1 and Python 2.7 on VDIs with 4 cores and 16GB RAM and
> (perceived) local disk storage, running RHEL6. We need 5000 seconds
> for 50000 targets, but we see the same quadratic behaviour as Eric
> Melski reports.
>
> We understand that the issue was diagnosed as being caused by the way
> Python implements fork() and waitpid() and that relief was expected
> from a wrapper using posix_spawn(). That stubprocess.py wrapper was
> slated for inclusion in SCons 2.5, but apparently did not make it (see
> https://bitbucket.org/scons/scons/wiki/Roadmap).
>
> What are the current plans integrating this stubprocess.py wrapper
> into an SCons release? And is there already an estimate when we can
> expect that?
>
> Best regards,
>
> Hans Ottevanger
> _______________________________________________
> Scons-users mailing list
> Scons-users at scons.org
> https://pairlist4.pair.net/mailman/listinfo/scons-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://pairlist4.pair.net/pipermail/scons-users/attachments/20170414/b26b5aeb/attachment.html>