Performance regression (binutils and gold)

Performance regression (binutils and gold)

[Richard Purdie]

To allow us to upgrade to the latest binutils we ended up enabling
building of gold. Unfortunately this is clearly visible on our
performance benchmarks, it leads to a 3-4 minute speed loss in overall
build time :(.

Locally, benchmarks for building binutils-cross show:

Just ld:

95.87user 48.36system 1:05.66elapsed 219%CPU (0avgtext+0avgdata
162480maxresident)k
0inputs+1030184outputs (0major+7753895minor)pagefaults 0swaps

with gold:

366.22user 67.25system 1:14.29elapsed 583%CPU (0avgtext+0avgdata
435488maxresident)k
248inputs+1205792outputs (1major+12572236minor)pagefaults 0swaps
richard@dax:/media/build1/poky/build2$ 

So whilst it only takes 10 seconds longer overall for me on what is a
pretty powerful machine, you can see a huge increase in user time and
CPU usage. I believe that on our slower quad core performance benchmark
machine, this equates to a few minutes.

The question is therefore, do we have a better way to fix things? Is
this performance drop acceptable?

Cheers,

Richard

Re: Performance regression (binutils and gold)

[Khem Raj]

On Sun, Apr 12, 2015 at 1:45 AM, Richard Purdie
<richard.purdie@linuxfoundation.org> wrote:
> The question is therefore, do we have a better way to fix things? Is
> this performance drop acceptable?

binutils being part of toolchains is hardly rebuilt so often in
practice. it will be reused from sstate so I am less worried, what
will worry me more
is if ld and gold itself have regressed in linking speed.

We can make it so that we only enable gold for cross build but it will
make it more complex to remember whats enabled where

Re: Performance regression (binutils and gold)

[Martin Jansa]

[-- Attachment #1: Type: text/plain, Size: 1365 bytes --]

On Sun, Apr 12, 2015 at 09:45:31AM +0100, Richard Purdie wrote:
> To allow us to upgrade to the latest binutils we ended up enabling
> building of gold. Unfortunately this is clearly visible on our
> performance benchmarks, it leads to a 3-4 minute speed loss in overall
> build time :(.
> 
> Locally, benchmarks for building binutils-cross show:
> 
> Just ld:
> 
> 95.87user 48.36system 1:05.66elapsed 219%CPU (0avgtext+0avgdata
> 162480maxresident)k
> 0inputs+1030184outputs (0major+7753895minor)pagefaults 0swaps
> 
> with gold:
> 
> 366.22user 67.25system 1:14.29elapsed 583%CPU (0avgtext+0avgdata
> 435488maxresident)k
> 248inputs+1205792outputs (1major+12572236minor)pagefaults 0swaps
> richard@dax:/media/build1/poky/build2$ 
> 
> So whilst it only takes 10 seconds longer overall for me on what is a
> pretty powerful machine, you can see a huge increase in user time and
> CPU usage. I believe that on our slower quad core performance benchmark
> machine, this equates to a few minutes.
> 
> The question is therefore, do we have a better way to fix things? Is
> this performance drop acceptable?

It's acceptable for me, I think most people with slower machines (with
less ram) are already using ls-is-gold, so it won't make any difference
for them.

Regards,
-- 
Martin 'JaMa' Jansa     jabber: Martin.Jansa@gmail.com

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 188 bytes --]