On Mon, May 08, 2017 at 04:07:52AM +0100, Ken Moffat wrote:
> On Mon, May 08, 2017 at 12:11:54AM +0100, Ken Moffat wrote:
> 
>       ld.bfd                  ld.gold
> run 1 1413.809s 16.2 SBU      1332.698s 15.3 SBU
> run 2 1318,971s 15.1 SBU      1339.608s 15.3 SBU
> run 3 1390.839s 15.9 SBU      1296.138s 14.8 SBU
> 
> I suspect those are too *few* runs to provide meaningful averages
> (and I would probably have to leave the machine totally idle in case
> using another term really is upsetting things).  At this point, all
> I can say is that ld.gold appears to be marginally faster - if you
> have built it AND have a slower machine, perhaps it does help.
> 
OK, I've done two more runs.

For clarity, I built first with bfd, then with gold, then second run
with bfs, second run with gold, etc.  And I dropped the caches
before each build.  The timings are specific to my machines and the
particular config I happened to use.  Also, I pass -O2 in CFLAGS,
CSSFLAGS: if anything, that should decrease the time spent in
compiling if a package prefers more exotic optimizations - and
therefore the share of hte time taken by ld would rise.  Also, by
setting CFLAGS I'm dropping debug info (ISTR firefox drops that by
default).  Also I made sure that the ld variant I did not want in
this run was renamed to ld.whichever.hidden so that the package
could not try to use it (firefox has been reputed to use gold if it
is available).

First, 5 runs of firefox on my AMD A10, where any reliable speedup
would be welcome.  System is from 20170331.  That box only has 4
cores, at times the keyboard was unresponsive, so I typed the results
in later.  These were all using the performance CPU governor with a
4.11.0 kernel and xscreensaver running after the first couple of
minutes.  These SBUs are rounded to 0.1.

SBU for 5 runs: Avg     Min     Max
ld.bfd          15.0    14.6    15.5
ld.gold         15.0    14.6    15.3

Second, 5 runs of webkitgtk on my intel i7, same system as upthread:

SBU for 5 runs: Avg     Min     Max
ld.bfd          38.6    38.1    39.1
ld.gold         38.6    38.4    38.8

Conclusion: gold was developed by google for use on large C++
applications and no doubt in some test context they can prove a
benefit.  But for ordinary builders any speedup is lost in the
overall noise of varying times.  It is interesting that for my
config of webkitgtk the range of timings with bfd is slightly wider,
but the average is identical.

I'm going to remove all the things I did to use gold in my own
scripts.

ĸen
-- 
I live in a city. I know sparrows from starlings.  After that
everything is a duck as far as I'm concerned.  -- Monstrous Regiment
-- 
http://lists.linuxfromscratch.org/listinfo/blfs-dev
FAQ: http://www.linuxfromscratch.org/blfs/faq.html
Unsubscribe: See the above information page

Reply via email to