Hello,
In the next kernel patch, you will find an updated AMD K8 PMU description
module. This new version implements the restriction described in the
BKDG for Opteron 64 concerning the NorthBridge (NB) events. In section
10.2, there is a rather confusing paragraph which says that if one
core of a dual-core processor has programmed a counter to measure
Northbridge events (event select >=0xee), then the other core
cannot be measuring Northbridge events.
There are multiple interpretations of the restriction. Yet only one is
fairly easy to implement at this point in perfmon.
In the perfmon implementation, we will have the following restrictions
enforced by the kernel module:
- in system-wide, only one session per socket will be allowed to
measure NB events. The restriction is enforced at pfm_load_context()
or pfm_write_pmcs(). There is no restriction on which core gets
access.
- in per-thread mode, only one session for the entire system can measure
NB events. This severe restriction comes from the fact that we CANNOT
modify/enforce processor affinity to allow two monitored threads on
distinct sockets to co-exist. Modifying affinity can severly impact
program performance which is exactly the opposite of what we are
trying
to measure. This is especially sensitive on Opteron because of the
NUMA
architecture.
--
-Stephane
_______________________________________________
perfmon mailing list
[email protected]
http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/
