Speeding up parallel compilation

Speeding up parallel compilation

[Stefan Monnier]

The recent change I installed which macro-expands Elisp code eagerly
upon load does not speed up sequential compilation very much, but it
seems to help parallel compilation significantly: on a 12-core
24-threads machine, "make -j24 bootstrap" went down from 2:20 to 1:50,
and the CPU utilization went up (from 800% to 1000%).

In other words, this mostly sped up a part of the compilation which does
not parallelize well.  I guess that would be either the construction of
loaddefs, or the compilation of the compiler.

So, lucky those with manycore systems.


        Stefan

Re: Speeding up parallel compilation

[joakim]

Stefan Monnier <monnier@iro.umontreal.ca> writes:

> The recent change I installed which macro-expands Elisp code eagerly
> upon load does not speed up sequential compilation very much, but it
> seems to help parallel compilation significantly: on a 12-core
> 24-threads machine, "make -j24 bootstrap" went down from 2:20 to 1:50,
> and the CPU utilization went up (from 800% to 1000%).
>
> In other words, this mostly sped up a part of the compilation which does
> not parallelize well.  I guess that would be either the construction of
> loaddefs, or the compilation of the compiler.
>
> So, lucky those with manycore systems.

Yay, thus I must get one! BTW was that a dual AMD or Intel btw?

>
>
>         Stefan

-- 
Joakim Verona

Re: Speeding up parallel compilation

[Stefan Monnier]

> Yay, thus I must get one! BTW was that a dual AMD or Intel btw?

I think it's an Intel rig (based on the "12-cores vs 24-threads" and the
fact that I don't think AMD has multithreaded CPUs (they seem to prefer
sharing the FPU part (and sometimes the decoder) between 2 cores
instead)).
[ And the times are with
  "--enable-checking --enable-check-lisp-object-type". ]

Sadly, the machines I use most of the time are pretty far from that,
using low power net(book|top) CPUs, so bootstrapping takes more like
half an hour (so I work hard to avoid bootstrapping).


        Stefan

parallel compilation --with-ns (was: re: Speeding up parallel compilation)

[chad]

On 05 Sep 2012, at 20:13, Stefan Monnier <monnier@iro.umontreal.ca> wrote:

> The recent change [...] seems to help parallel compilation significantly:

Under macosx (i.e. --with-ns), I've tried using 'make -j4' off
and on over the years, and frequently run into problems during
the build. I end up without a complete nextstep/Emacs.app most of
the time, and it seems to be a problem with directories not being
created before one of the workers wants to use it. Three
questions:

1.) Is the ns-app code known to be unable to use parallel
    compilation?

2.) Does it matter that I do full rebuilds each time, starting
    with autogen.sh and ending with make bootstrap install)?

3.) Can anyone suggest a good way to start debugging this?

Thanks,
*Chad

Re: parallel compilation --with-ns

[Paul Eggert]

On 09/06/2012 09:19 AM, chad wrote:
> 3.) Can anyone suggest a good way to start debugging this?

For stuff like this I use "truss" (or its equivalent)
and "make -d".  It's not fun, but sometimes it works.

Re: parallel compilation --with-ns

[Sudish Joseph]

chad <yandros@MIT.EDU> writes:
> Under macosx (i.e. --with-ns), I've tried using 'make -j4' off
> and on over the years, and frequently run into problems during
> the build. I end up without a complete nextstep/Emacs.app most of
> the time, and it seems to be a problem with directories not being
> created before one of the workers wants to use it. Three
> questions:
>
> 1.) Is the ns-app code known to be unable to use parallel
>     compilation?

I was doing regular make -j8 builds of the ns port up until a few months
ago and had no problems with it.

> 2.) Does it matter that I do full rebuilds each time, starting
>     with autogen.sh and ending with make bootstrap install)?

Here's the essence of the function I use to build the ns port.

sj_emacsbuild () {
  local build_dir
  rm -rf sj-build.*
  build_dir=sj-build.$$ 
  sh ./autogen.sh
  mkdir $build_dir || {
    echo "couldn't mkdir $build_dir: $!"
    return 1
  }
  cd $build_dir
  env "$EMACS_CC_FLAGS[@]" ../configure "$EMACS_CONF_FLAGS[@]" && nice -n 20 make -j8 && nice -n 20 make install
}

I then copy over Emacs.app to ~/Applications using a separate func.

You don't need to do a explicit make bootstrap, iirc.
"make maintainer-clean && make -j8" does a full build.

I used to have "git clean -xfd" in the build func above but didn't
adjust for that when switching to the bzr tree.  I.e., full builds
should work in parallel with the ns port.

-Sudish

Re: parallel compilation --with-ns

[chad]

On 10 Sep 2012, at 11:25, Sudish Joseph <sudish@gmail.com> wrote:
>> 1.) Is the ns-app code known to be unable to use parallel
>>    compilation?
> 
> I was doing regular make -j8 builds of the ns port up until a few months
> ago and had no problems with it.

FWIW, the key seems to be that the install target needs to wait for the 
bootstrap or all target to complete (i.e. `make install' in a separate 
process after `make -j8 all/bootstrap').

Thanks for the help, everyone.
*Chad