Ah, just as I expected. :)
On the Sun Fire series with multiple system boards, by default memory
is interleaved within board:
> N0.SB0::memory connected configured ok base
> address 0x0, 33554432 KBytes total, 1768104 KBytes permanent
> N0.SB2::memory connected configured ok base
> address 0x2000000000, 16777216 KBytes total
> N0.SB4::memory connected configured ok base
> address 0x4000000000, 8388608 KBytes total
The result of which is that the whole kernel resides in memory off the
board with so-called "permanent memory" -- memory that can't be removed
from the system. Any processor running the kernel will have to access
this range of memory, resulting in the CPUs on this board being busier
than those on the other boards, mostly handling xcalls.
The reason we choose within-board interleaving as the default is to
support Dynamic Reconfiguration (DR) with which you can add or remove
boards without bringing the system down.
To verify my theory, you can do the following:
# cfgadm -c unconfigure N0.SB0 (this will take a short while)
# cfgadm -c configure N0.SB0
Now if you run mpstat again, you will see CPUs off SB2 are busier than
the rest.
There are ways to change this behavior, though not all are recommended
for many other reasons. One of the white papers I wrote many years ago
touched on this subject. I will see if I can dig it up if you are
interested.
Sherry
--
[EMAIL PROTECTED], Solaris Kernel Development, http://blogs.sun.com/sherrym
_______________________________________________
perf-discuss mailing list
perf-discuss@opensolaris.org
