bug#64735: 29.0.92; find invocations are ~15x slower because of ignores

[Dmitry Gutov]

On 12/09/2023 19:32, Eli Zaretskii wrote:
> > Date: Tue, 12 Sep 2023 17:23:53 +0300
> > Cc: luangruo@yahoo.com, sbaugh@janestreet.com, yantar92@posteo.net,
> >   64735@debbugs.gnu.org
> > From: Dmitry Gutov <dmitry@gutov.dev>
> >
> > - Dropping most of the setup in read_and_dispose_of_process_output
> > (which creates some consing too) and calling
> > Finternal_default_process_filter directly (call_filter_directly.diff),
> > when it is the filter to be used anyway.
> >
> > - Going around that function entirely, skipping the creation of a Lisp
> > string (CHARS -> TEXT) and inserting into the buffer directly (when the
> > filter is set to the default, of course). Copied and adapted some code
> > from 'call_process' for that (read_and_insert_process_output.diff).
> >
> > Neither are intended as complete proposals, but here are some
> > comparisons. Note that either of these patches could only help the
> > implementations that don't set up process filter (the naive first one,
> > and the new parallel number 5 above).
> >
> > For testing, I used two different repo checkouts that are large enough
> > to not finish too quickly: gecko-dev and torvalds-linux.
> >
> > master
> >
> > | Function                                         | gecko-dev | linux |
> > | find-directory-files-recursively                 |      1.69 |  0.41 |
> > | find-directory-files-recursively-2               |      1.16 |  0.28 |
> > | find-directory-files-recursively-3               |      0.92 |  0.23 |
> > | find-directory-files-recursively-5               |      1.07 |  0.26 |
> > | find-directory-files-recursively (rpom 409600)   |      1.42 |  0.35 |
> > | find-directory-files-recursively-2 (rpom 409600) |      0.90 |  0.25 |
> > | find-directory-files-recursively-5 (rpom 409600) |      0.89 |  0.24 |
> >
> > call_filter_directly.diff (basically, not much difference)
> >
> > | Function                                         | gecko-dev | linux |
> > | find-directory-files-recursively                 |      1.64 |  0.38 |
> > | find-directory-files-recursively-5               |      1.05 |  0.26 |
> > | find-directory-files-recursively (rpom 409600)   |      1.42 |  0.36 |
> > | find-directory-files-recursively-5 (rpom 409600) |      0.91 |  0.25 |
> >
> > read_and_insert_process_output.diff (noticeable differences)
> >
> > | Function                                         | gecko-dev | linux |
> > | find-directory-files-recursively                 |      1.30 |  0.34 |
> > | find-directory-files-recursively-5               |      1.03 |  0.25 |
> > | find-directory-files-recursively (rpom 409600)   |      1.20 |  0.35 |
> > | find-directory-files-recursively-5 (rpom 409600) | (!!) 0.72 |  0.21 |
> >
> > So it seems like we have at least two potential ways to implement an
> > asynchronous file listing routine that is as fast or faster than the
> > synchronous one (if only thanks to starting the parsing in parallel).
> >
> > Combining the last patch together with using the very large value of
> > read-process-output-max seems to yield the most benefit, but I'm not
> > sure if it's appropriate to just raise that value in our code, though.
> >
> > Thoughts?
>
> I'm not sure what exactly is here to think about.  Removing portions
> of read_and_insert_process_output, or bypassing it entirely, is not
> going to fly, because AFAIU it basically means we don't decode text,
> which can only work with plain ASCII file names, and/or don't move the
> markers in the process buffer, which also cannot be avoided.

That one was really a test to see whether the extra handling added any 
meaningful consing to affect GC. Removing it didn't make a difference, 
table number 2, so no.

> If you
> want to conclude that inserting the process's output into a buffer
> without consing Lisp strings is faster (which I'm not sure, see below,
> but it could be true),

That's what my tests seem to show, see table 3 (the last one).

> then we could try extending
> internal-default-process-filter (or writing a new filter function
> similar to it) so that it inserts the stuff into the gap and then uses
> decode_coding_gap,

Can that work at all? By the time internal-default-process-filter is 
called, we have already turned the string from char* into Lisp_Object 
text, which we then pass to it. So consing has already happened, IIUC.

> which converts inserted bytes in-place -- that, at
> least, will be correct and will avoid consing intermediate temporary
> strings from the process output, then decoding them, then inserting
> them.  Other than that, the -2 and -3 variants are very close
> runners-up of -5, so maybe I'm missing something, but I see no reason
> be too excited here?  I mean, 0.89 vs 0.92? really?

The important part is not 0.89 vs 0.92 (that would be meaningless 
indeed), but that we have an _asyncronous_ implementation of the feature 
that works as fast as the existing synchronous one (or faster! if we 
also bind read-process-output-max to a large value, the time is 0.72).

The possible applications for that range from simple (printing progress 
bar while the scan is happening) to more advanced (launching a 
concurrent process where we pipe the received file names concurrently to 
'xargs grep'), including visuals (xref buffer which shows the 
intermediate search results right away, updating them gradually, all 
without blocking the UI).

> About inserting into the buffer: what we do is insert into the gap,
> and when the gap becomes full, we enlarge it.  Enlarging the gap
> involves: (a) enlarging the chunk of memory allocated to buffer text
> (which might mean we ask the OS for more memory), and (b) moving the
> characters after the gap to the right to free space for inserting more
> stuff.  This is pretty fast, but still, with a large pipe buffer and a
> lot of output, we do this many times, so it could add up to something
> pretty tangible.  It's hard to me to tell whether this is
> significantly faster than consing strings and inserting them, only
> measurements can tell.

See the benchmark tables and the POC patch in my previous email. Using a 
better filter function would be ideal, but it seems like that's not 
going to fit the current design. Happy to be proven wrong, though.