Re: [perfmon2] same counter values for several processes onBarcelona

[Rob F. <rj...@re...>]

One detail: As I've mentioned in other messages on the subject, you can
set up a "system-wide" session using pfmon to count the NB events.  This
will continue to count even when the thread that started the session
sleeps.

Drongowski, Paul wrote:
> Hello Sergey --
> 
> Just to add a few comments to Stephane's.
> 
> Northbridge resources are shared across all cores
> in the "node" (the cores that are local to the
> Northbridge and communicate directly with the NB).
> Northbridge events measure memory controller events,
> crossbar events, HyperTransport(tm) events and
> L3 cache events. The events generally do not break
> out by core, i.e., these X events belong to core 0,
> these Y events belong to core 1, etc. They measure
> events due to requests, etc. across all local cores.
> 
> If all cores are measuring the same NB event, the
> events will be counted multiple times -- not good.
> So, perfmon2 follows the recommended practice, which is
> to measure NB events on a single core
> belonging to a given NB.
> 
> There's one potential gotcha. The core measuring
> the NB events cannot go into halt mode (idle) as the
> clock for the performance counter logic may be stopped
> and events are not counted. It may be necessary to
> keep the core awake by running a low-priority
> background task.
> 
> More information concerning Northbridge events can be
> found in Section 3.12 of the BIOS and Kernel Developer's
> Guide (BKDG) for AMD Family 10h processors.
> 
> -- pj
> 
> 
> -----Original Message-----
> From: stephane eranian [mailto:er...@go...] 
> Sent: Friday, October 30, 2009 6:59 PM
> To: Sergey Blagodurov
> Cc: per...@li...
> Subject: Re: [perfmon2] same counter values for several processes
> onBarcelona
> 
> Sergey,
> 
> On Fri, Oct 30, 2009 at 11:23 PM, Sergey Blagodurov
> <bla...@gm...> wrote:
>> Hello,
>>
>> I have the following problem while trying to gather hardware counters
>> on Barcelona: whenever I start several programs with the same counter
>> profiling on AMD Barcelona, all of the program invocations except the
>> first failing with conflicting monitoring session error. The problem
>> can be revealed with smth like that:
>>
>> $ pfmon --events=MEMORY_CONTROLLER_REQUESTS:READ_REQUESTS sleep 20 &
>> [1]
>> $ pfmon --events=MEMORY_CONTROLLER_REQUESTS:READ_REQUESTS sleep 20 &
>> error conflicting monitoring session exists
>>
> 
> This is due to a constraint on all Northbridge events.
> The way the current perfmon code implements the restriction is to
> limit the number of perfmon
> sessions which have access to NB events to 1 per socket in system-wide
> and 1 global in per-thread
> mode.
> 
> 
>> There is no such behaviour on Intel Xeon processors. I wonder what may
>> cause this strange problem and are there any workarounds to it (i.e.
>> is it still possible to gather same counter info for several
>> programs/threads at the same time)?
>>
>> Thank you!
>>
>> --
>> Sincerely,
>>
>> Sergey Blagodurov
>>
> 
> 
> 
> ------------------------------------------------------------------------------
> Come build with us! The BlackBerry(R) Developer Conference in SF, CA
> is the only developer event you need to attend this year. Jumpstart your
> developing skills, take BlackBerry mobile applications to market and stay 
> ahead of the curve. Join us from November 9 - 12, 2009. Register now!
> http://p.sf.net/sfu/devconference
> _______________________________________________
> perfmon2-devel mailing list
> per...@li...
> https://lists.sourceforge.net/lists/listinfo/perfmon2-devel