gcc 10 build times

Tony0945 · Posted: Mon Mar 08, 2021 2:05 am Post subject: gcc 10 build times

34 minutes on a Ryzen 2700X
4 hours 23 minutes on AMD A8-7600 (Bulldozer core)

Phenom II & Athlon II X3, can't bear to think about it. I have a fam10 build environment on the Ryzen 2700X. building gcc10 as a package (--buildpkgonly) for fam10 on the Ryzen right now.
EDIT:
33 minutes 52 seconds on Ryzen building for fam10

mike155 · Posted: Mon Mar 08, 2021 2:16 am Post subject:

Tony0945 · Posted: Mon Mar 08, 2021 3:15 am Post subject:

sdauth · Guru Joined: 19 Sep 2018 Posts: 569 Location: Ásgarðr

On a AMD A10-6800K APU :

Tue Feb 23 19:51:22 2021 >>> sys-devel/gcc-10.2.0-r5
merge time: 4 hours, 31 minutes and 33 seconds.

No overclock, 1600MHz ram, nothing running in parallel. I trashed the log for GCC9 but I remember being more like 3h50min or something.
Both with fortran, graphite, hardened, lto, pgo enabled.

Nothing so terrible to be honest. On my Atom, we would be speaking in days.
That Ryzen build time for GCC makes me jealous though but now is not really the right time to build a new PC.

Especially in the GPU area.

saellaven · l33t Joined: 23 Jul 2006 Posts: 646

Ryzen 3700X:

Machine A:
Sun Mar 7 23:24:02 2021 >>> sys-devel/gcc-10.2.0-r5
merge time: 22 minutes and 45 seconds.

Machine B:
Sun Mar 7 23:24:07 2021 >>> sys-devel/gcc-10.2.0-r5
merge time: 22 minutes and 44 seconds.
_________________
Ryzen 3700X, Asus Prime X570-Pro, 64 GB DDR4 3200, GeForce GTX 1660 Super
openrc-0.17, ~vanilla-sources, ~nvidia-drivers, ~gcc

Anon-E-moose · Posted: Mon Mar 08, 2021 11:08 am Post subject:

I had put off building gcc 10, I just didn't want to mess with it, but lately I've been thinking about upgrading, and was curious as to how much slower that 9.3.
The answer seems to be not much, 9.3 ~ 20 min and from saellaven's posts, I'll spend a couple of minutes more. Good.
_________________
PRIME x570-pro, 3700x, 6.1 zen kernel
gcc 13, profile 17.0 (custom bare multilib), openrc, wayland

Irre · Guru Joined: 09 Nov 2013 Posts: 434 Location: Stockholm

Earlier GCCs seem to be faster.

mike155 · Posted: Mon Mar 08, 2021 1:32 pm Post subject:

mike155 · Posted: Mon Mar 08, 2021 1:41 pm Post subject:

apiaio · Guru Joined: 04 Dec 2008 Posts: 417

Tony0945 · Posted: Mon Mar 08, 2021 3:13 pm Post subject:

Tony0945 · Posted: Mon Mar 08, 2021 3:28 pm Post subject:

figueroa · Posted: Mon Mar 08, 2021 7:54 pm Post subject:

I just got 62 minutes for my Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz w/16GB RAM, amd64, -j9

Also, 126 minutes for the AMD Phenom(tm) 8650 Triple-Core Processor w/8GB RAM, x86, -j4

I've gone ahead and converted both systems to 10.2.0.

On the x86 box, I've got build time stats back to 2004:

eccerr0r · Posted: Wed Mar 10, 2021 7:09 am Post subject:

my results are contaminated with distcc so take with grain of salt, along with tmpfs and running (mostly idle) VM's at the same time. Also machines are distcc'ing each other so that will slow things down...

Also I may have disabled mitigations, with exception of the Atom of course.

"?" means still pending. Atom and P4 are 1C2T, i5 is 2C4T, i7 is 4C8T. I did not load down all 8 threads. All are amd64 except Atom and c2q32/4g. i5/i7 are SSD'd
None of the machines are overclocked currently. i7=~3.5GHz+TB, i5=~1.7GHz+TB, P4=3.4GHz, C2 machines are all around 2.7GHz, Atom is 1.6GHz

dmpogo · Advocate Joined: 02 Sep 2004 Posts: 3267 Location: Canada

eccerr0r · Posted: Wed Mar 10, 2021 8:12 am Post subject: Re: gcc 10 build times

Anon-E-moose · Posted: Wed Mar 10, 2021 10:02 am Post subject: Re: gcc 10 build times

eccerr0r · Posted: Wed Mar 10, 2021 4:20 pm Post subject:

Don't think it's swapping, the numbers seem quite consistent depending on clock and core count...

Then again the -j -l counts technically should be different for gcc/rust compared to other packages; even worse, should be different for stage1/0 than the other stages.

---

For curiosity's sake I disabled distcc and set MAKEOPTS to more reasonable numbers just in case it was swapping.
It looks like my c2d/4g got through stage1 in only 33 minutes. This is corresponding to a 100 minute build time which is a lot faster than with distcc options - so it looks like I will need to make an env for gcc and probably rust as well to optimize for these packages.

If it really can finish in 100 minutes, then I have to fix all my machines... distcc options for gcc is BAD.

---

Nope, not 100 minutes. It took 3h20m this time with -j3 (c2d@2.66/4g).
Stage1 33 min
Stage2 92 min
Stage3 remainder...
Also this will be a multilib gcc, wonder how much this adds to build time...

Oh... gfortran... I'm building this... What USE flags are people using?

(Added for reference for this c2d/4g: )
[ebuild R ] sys-devel/gcc-10.2.0-r5:10::gentoo USE="(cxx) fortran (multilib) nls nptl openmp pch (pie) sanitize ssp (-ada) -d -debug -doc (-fixed-point) -go -graphite (-hardened) -jit (-libssp) -lto -objc -objc++ -objc-gc -pgo -systemtap -test -vanilla -vtv -zstd" 0 KiB
_________________
Intel Core i7 2700K/Radeon R7 250/24GB DDR3/256GB SSD
What am I supposed watching?

saellaven · l33t Joined: 23 Jul 2006 Posts: 646

psycho · Guru Joined: 22 Jun 2007 Posts: 534 Location: New Zealand

64 minutes here on this old i5-3470 (8G RAM, no swap, old spinning hard disks).

eccerr0r · Posted: Thu Mar 11, 2021 12:27 am Post subject:

Same USE flags, but only using local resources (no distcc):

My c2d(2.67GHz)/4g got 3h20m, my i5(1.7GHz+TB)/8g got 3h8m. Seems turboboost and microarchitectural speedups were able to make up for the lower minspeed of the i5. Both are dual core processors though the i5 has 4 threads (and SSD) while the c2d has 2 (rust spinner).

I still would like to know how a 2GHz c2d was able to get 2h49m... If it were omitting fortran, that would explain things among other possibilities.

---

I found a 2.4GHz c2q only needed 1h40m to complete gcc10. This c2q is 32-bit however, this may also be part of the reason for why this machine is finishing faster than the other core2 machines.
_________________
Intel Core i7 2700K/Radeon R7 250/24GB DDR3/256GB SSD
What am I supposed watching?

dmpogo · Advocate Joined: 02 Sep 2004 Posts: 3267 Location: Canada

eccerr0r · Posted: Fri Mar 12, 2021 6:40 am Post subject: Re: gcc 10 build times

dmpogo · Advocate Joined: 02 Sep 2004 Posts: 3267 Location: Canada

eccerr0r · Posted: Fri Mar 12, 2021 8:38 am Post subject:

USE flags match then.
Only other thing that's clearly different is kernel...

The "fast" c2q is running 4.19.86 Scaling to two cores would be in the ballpark with what you're seeing (100 mins). It's using -j6.
the "slow" c2d is running 5.4.97 (mitigations disabled) - So far the best I've seen is 3h20m with -j3 so it's still almost 20% slower.
The "slow" c2qs is running 4.14.65 and 5.4.97, both of these are close together (~130 mins). This doesn't make sense because these c2qs are clocked faster than the "fast" c2q.

Something is very wrong here, but cant put my finger on it... What's bothersome is that the rust numbers are relatively within the ballpark too based on clock frequency. Just gcc is really unexpected.

Doing the math, seems the "fast" c2q is very slow building rust relative to clock speed, so there's some other factor at hand here perhaps...
_________________
Intel Core i7 2700K/Radeon R7 250/24GB DDR3/256GB SSD
What am I supposed watching?