On Tue, 10 Apr 2012, Richard Lowe wrote:
The logical first step in determining why we got SIGILL would be
looking at the instruction we were hoping to execute:
$ mdb core
<rip/i
(or <eip/i if 32bit)
This is a 64-bit binary. This is what I get:
openindiana% mdb
<rip/i
0: addb %al,(%rax)
Note that you're running on Opterons and Hans only very very recently
taught illumos to acknowledge that AVX and its pre-requisites exist on
non-Intel CPUs, which may matter (though I would not expect SIGILL, I
don't think)
I am using a borrowed system remotely which is only available for
perhaps another day. I am sure that this system represents whatever
came on the OpenIndiana install ISO. I could update the system if the
updates are available in the form of a package which can be updated.
Unfortunately, I also have an unrelated day job which mostly limits my
availability to work on this except for in the evenings.
Previously I had noticed that the compiler used had a huge impact on
how well the last 1/2 of the cores could be used on Interlagos.
Using the Open64 compiler, the last 1/2 of the cores behaved like real
cores, producing almost as much speedup as the first 1/2 of the cores
(an almost completely linear ramp). Using GCC with 'barcelona' output
on Linux and OI resulted in unstable performance in the last 1/2 of
the cores, cutting available performance by 1/3rd. My hope was to
test GCC 4.7 with its 'bdver1' output to see if it cures the
performance problem. Since I would be purchasing the system out of my
own pocket, if I can not obtain satisfactory performance then I would
purchase an E5 Xeon system instead. The 64-cores are quite attractive
to me, but not if they can not be effectively used.
Bob
--
Bob Friesenhahn
[email protected], http://www.simplesystems.org/users/bfriesen/
GraphicsMagick Maintainer, http://www.GraphicsMagick.org/
-------------------------------------------
illumos-discuss
Archives: https://www.listbox.com/member/archive/182180/=now
RSS Feed: https://www.listbox.com/member/archive/rss/182180/21175430-2e6923be
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21175430&id_secret=21175430-6a77cda4
Powered by Listbox: http://www.listbox.com
