imports and python locking

[Neal Becker]

Martin Pool wrote:

> Some benchmarking on bzr's hashcache showed a downside to our current
> policy of doing imports in inner scopes.
> 
> Every import statement results in a builtin_import call, which needs to
> take and release a lock protecting the module table.  This is done by
> semaphores (sem_post, etc) by default.  On Linux, for some reason I
> don't understand yet, this always causes a system call to release the
> futex, even though I would think it's not necessary.  In any case for
> 'bzr st' this can take up to 40% of the time.
> 
> This is compiled out on python built --without-threads, and it would be
> kind of neat to run that for applications that don't use threads, but
> that's not really in our scope here.
> 
> Practically, I think we should change the practice to generally doing
> imports at the top level of the module, except that the top-level code
> should only import modules when they're needed for the selected command.
> 
> Another approach is to use something like the attached hook to defer
> imports.
> 

What distro are you testing on?  Is this a modern glibc with nptl?