parallel loading of debug info in gdb

parallel loading of debug info in gdb

[Yubin via Gdb]

Hi,

Does gdb supports parallel loading of debug symbols? I am debugging some
binary which is 3~4GB itself, and loading and parsing  its debug info takes
minutes.

Thanks,
Yubin

Re: parallel loading of debug info in gdb

[Luis Machado via Gdb]

On 1/7/25 02:14, Yubin via Gdb wrote:
> Hi,
> 
> Does gdb supports parallel loading of debug symbols? I am debugging some
> binary which is 3~4GB itself, and loading and parsing  its debug info takes
> minutes.

Since a few years, gdb has had worker threads, potentially as many as cores
(physical or virtual) on the system.

So it does support trying to read things in parallel. Unfortunately the
nature of the data limits things somewhat, as you have dependencies and have
to look those up before proceeding further.

I've cc-ed Tromey, which worked in this area before.

> 
> Thanks,
> Yubin

Re: parallel loading of debug info in gdb

[Yubin via Gdb]

On Tue, Jan 7, 2025 at 7:08 PM Luis Machado <luis.machado@arm.com> wrote:

> On 1/7/25 02:14, Yubin via Gdb wrote:
> > Hi,
> >
> > Does gdb supports parallel loading of debug symbols? I am debugging some
> > binary which is 3~4GB itself, and loading and parsing  its debug info
> takes
> > minutes.
>
> Since a few years, gdb has had worker threads, potentially as many as cores
> (physical or virtual) on the system.
>
> So it does support trying to read things in parallel. Unfortunately the
> nature of the data limits things somewhat, as you have dependencies and
> have
> to look those up before proceeding further.
>

Yes dependency is a problem, but some dependent libs
are  naturally independent to be read by parallel. If there could be some
way to specify this, it would be great.


> I've cc-ed Tromey, which worked in this area before.
>

Thanks.

Yubin

Re: parallel loading of debug info in gdb

[Tom Tromey]

> Yes dependency is a problem, but some dependent libs are naturally
> independent to be read by parallel. If there could be some way to
> specify this, it would be great.

If you try git gdb, you can enable background reading with
"maint set dwarf synchronous off".

(There's been at least one request to re-enable this by default for GDB
16... I wasn't sure if we should, though, considering that the data race
fixes are pretty new.)

Anyway, background reading is lazier -- the reading is started but gdb
will only pause for it if it is truly needed.

Note that the library debug info isn't really read in parallel.  Or, it
kind of is, depending on how the scheduling shakes out.  That is, the
DWARF reader looks at the number of worker threads and splits its work
up accordingly.  "Overlap" can only really happen if one such worker
finishes and then the next library's job is scheduled.

This might be a bad strategy overall.  Like, testing on a big machine
showed that the reader didn't parallelize past 8 CPUs, so maybe it would
be better to limit the DWARF reader to 8 tasks.  Nobody has worked on
figuring out why it doesn't scale better, though, and I wouldn't want to
change this without knowing more.

Tom

Re: parallel loading of debug info in gdb

[Yubin via Gdb]

On Wed, Jan 8, 2025 at 7:45 AM Tom Tromey <tom@tromey.com> wrote:

> > Yes dependency is a problem, but some dependent libs are naturally
> > independent to be read by parallel. If there could be some way to
> > specify this, it would be great.
>
> If you try git gdb, you can enable background reading with
> "maint set dwarf synchronous off".
>


Thanks, using this command in gdb15.2 indeed make the initial loading
faster.


>
> (There's been at least one request to re-enable this by default for GDB
> 16... I wasn't sure if we should, though, considering that the data race
> fixes are pretty new.)
>
> Anyway, background reading is lazier -- the reading is started but gdb
> will only pause for it if it is truly needed.
>
> Note that the library debug info isn't really read in parallel.  Or, it
> kind of is, depending on how the scheduling shakes out.  That is, the
> DWARF reader looks at the number of worker threads and splits its work
> up accordingly.  "Overlap" can only really happen if one such worker
> finishes and then the next library's job is scheduled.
>
> This might be a bad strategy overall.  Like, testing on a big machine
> showed that the reader didn't parallelize past 8 CPUs, so maybe it would
> be better to limit the DWARF reader to 8 tasks.  Nobody has worked on
> figuring out why it doesn't scale better, though, and I wouldn't want to
> change this without knowing more.
>
> Tom
>

Re: parallel loading of debug info in gdb

[Tom Tromey]

> Thanks, using this command in gdb15.2 indeed make the initial loading faster.
 
Just beware that in 15.2 there are some known data races, which is why
it is off by default.  I don't know whether those races can actually
cause problems.

Tom