Re: Optimizing union.scm

[Ludovic Courtès]

Mark H Weaver <address@hidden> skribis:

> address@hidden (Ludovic Courtès) writes:
>
>> Mark H Weaver <address@hidden> skribis:
>>
>>> I want to optimize it anyway, since it takes over 5 minutes to build
>>> my profile, which is a bit painful.
>>
>> Oh, this much?
>>
>> I have 140 packages in my profile and it takes less than 30s to build
>> it; that’s an SSD though, so that probably makes a big difference.
>
> My profile has 155 packages.  I've looked over union.scm and can see
> some extreme wastefulness, most notably in the use of 'others-have-it?'.
> I guess this typically does N 'lstat' calls for every file or directory
> in every package in the resulting profile that cannot be pruned (due to
> being in a directory that's only in one package), where N is the number
> of packages in the profile.

We could use a memoizing version of ‘file-exists?’ in ‘union-build’ and
see what happens.

> I'm fairly sure it is possible to replace those N 'lstat' calls with
> something that requires 0 system calls and at most O(log N) time.  The
> basic idea would be to iterate over all packages in a breadth-first
> manner, as follows:
>
> I think we should readdir the top-level directories of every package,
> and merge them together into a single map structure (vhash?) that maps
> filenames to sets of packages containing that filename.  We don't even
> need to 'lstat' them at this point, all we need are the names.  After
> we've read the directories of every package, we then iterate over the
> map.  For any unique entries, we simply make a symlink in the new
> profile.
>
> For duplicates we'd do conflict resolution, which starts with an
> 'lstat'.  For directories, the default conflict resolution would simply
> create a directory in the new profile and then recurse into that
> subdirectory, considering only the packages that contained that
> subdirectory.
>
> My guess is that this would speed up profile creation by at least an
> order of magnitude, maybe more.
>
> What do you think?

I’m all for it!

It’s not obvious to me that we would gain this much (esp. an order of
magnitude), but if that’s the case, it’s great.  (I’m guessing that the
bottleneck is I/O, not CPU.)

My only concern is that that code should remain readable and tested.

So: patches welcome.  ;-)

Ludo’.