Re: gcc 3.1 branch slowdown?

Re: gcc 3.1 branch slowdown?

[Loren James Rittle]

Responding to HJL, I wrote on the public gcc list:
>> For 0709, total CPU time was: ~62 minutes
>> For 0711, total CPU time was: ~63 minutes
>> For 0717, total CPU time was: ~63 minutes

Someone indirectly asked me in private e-mail why those facts matter
at all.  Perhaps, they don't (which I why I stated the sole
near-absolute conclusion that could be reached from the posted data
before observing the near-constant CPU time ;-).  However, in my
experience, they do matter.

More to the point, I should have added: From another i386 machine
covering those dates with 3.1.1 prerelease, I saw no regression in
total wall update/build/check time (at -j1 on a single CPU with ~512MB
of physical memory which has not used one iota of swap space since
being booted at the end of June):

07/08 posted first final results at 17:54.
07/10 posted first final results at 17:49.
07/12 posted first final results at 17:54.
07/14 posted first final results at 18:04.
07/16 posted first final results at 17:56.
07/18 posted first final results at 18:04.
07/20 posted first final results at 17:49.
07/22 posted first final results at 18:00.

All days above completed the same total work (with the possible
exception of a high-variance wall time to run ``cvs update'' in the
source tree) and started at the same time (13:51).  The jobs ran at
``nice -10'' thus all other jobs on the system tended to get CPU
before these jobs.  At this point, it is unknown whether any other
real jobs ran those days.  It appears to me that the other load was
"light/none".

Regards,
Loren

Re: gcc 3.1 branch slowdown?

[Loren James Rittle]

> 1. "make bootstrap" on 0709 takes [...]
> user    51m12.280s
> sys     10m46.560s

> 2. "make bootstrap" on 0711 takes [...]
> user    52m21.290s
> sys     11m1.410s

> 3. "make bootstrap" on 0717 takes [...]
> user    52m20.570s
> sys     10m53.770s

From the data posted, all we know is that you likely tested on an
[S]MP machine.  Ignoring real time, AKA wall-time (as I think we must
since we don't know your other system load...):

For 0709, total CPU time was: ~62 minutes
For 0711, total CPU time was: ~63 minutes
For 0717, total CPU time was: ~63 minutes

Regards,
Loren

Re: gcc 3.1 branch slowdown?

[David Edelsohn]

>>>>> H J Lu writes:

HJ> I noticed a slowdown on 3.1 branch between 0709 and 0711.
HJ> BTW, it is on Linux/i686. Has anyone else seen it?

	While bootstrap may be slower, it does not necessarily mean that
the compiler is slower.  Possibly the C++ correctness fixes are slowing
down compiling libstdc++?

David

gcc 3.1 branch slowdown?

[H. J. Lu]

I noticed a slowdown on 3.1 branch between 0709 and 0711.

1. "make bootstrap" on 0709 takes

real    34m43.021s
user    51m12.280s
sys     10m46.560s

2. "make bootstrap" on 0711 takes

real    40m26.992s
user    52m21.290s
sys     11m1.410s

3. "make bootstrap" on 0717 takes

real    52m55.189s
user    52m20.570s
sys     10m53.770s

BTW, it is on Linux/i686. Has anyone else seen it?


H.J.