On Mon, May 05, 2014 at 08:41:40PM +0200, Frans de Boer wrote:
>
> @Bruce, so maybe it is timing. Maybe it is timing combined with the type of
> processor. You use a i7 and I use an AMD Phenom II 965. Your SBU's are also
> about 60% compared to my machine.
>
Interesting. My phenom is still using gcc-4.8 (I bought it for
building development versions of the books, but it is running out of
partitions - that is down to details of my backup process : rsync
over nfs to a staging area on my server, then create something akin
to generation data groups for the real backups on a different
filesystem [¹] and isn't going to change in the near future).
Anyway - are you using the ondemand cpufreq driver ? If not,
forget this mail. What I found at some point in the last year,
after the kernel added a new cpufreq driver for recent AMD machines,
was that my phenom (4 real cores, 3.4GHz max) was suddenly slower
than my SandyBridge i5 (2 real cores, but linux sees 4 CPUs because
of the hyperthreading, 3.2GHz max). Looking at the kernel docs,
there was a mention of sampling_down_factor, with an example of
echoing 100 to it. Unfortunately, it has to be set whenever I
change the governor, but at least it is not "per cpu" like the
governors. So, what I now use for the ondemand governor is:
echo 100 >/sys/devices/system/cpu/cpufreq/ondemand/sampling_down_factor
and that again makes it faster than my intel, and it still falls
back to its slowest speed when idle (e.g. during configure scripts
some CPUs will slow down). I have a "speed" alias in my user's
bashrc to report all MHz frequencies from /proc/cpuinfo on a single
line, it differs for each machine.
I've also got a desktop AMD A4 ("trinity APU") which only has two
cores. Mostly, that box is not used for development and at the
moment it is running old LFS's (7.4 or older) and old stable
kernels. But a couple of weeks ago I noticed people were reporting
problems on 3.15-rc kernels with radeons, and decided to try -rc2.
For that box there was a problem, which has now been fixed. Along
the way, I did some kernel bisection and noticed that kernel full
builds were taking longer than I had expected. After checking
things, I discovered that the box was mostly using its slowest
speed. Applying the same workaround for sampling_down_factor seems
to have fixed that.
ĸen
1. GDGs - I'm sure wikipedia knows about them, a form of dataset
organisation on IBM mainframes using VSAM. Arthur Dent would
probably describe my backups as "almost, but not completely, unlike"
GDGs - I only have relative generations, and form them with hard
links.
--
das eine Mal als Tragödie, dieses Mal als Farce
--
http://lists.linuxfromscratch.org/listinfo/lfs-support
FAQ: http://www.linuxfromscratch.org/blfs/faq.html
Unsubscribe: See the above information page
