Stephane,
On Mon, May 5, 2008 at 6:02 PM, stephane eranian <[EMAIL PROTECTED]>
wrote:
> Vivek,
>
>
> On Mon, May 5, 2008 at 11:42 PM, Vivek Thakkar <[EMAIL PROTECTED]> wrote:
> > Hi Stephane,
> >
> > Do you have some examples of configuring load miss event using IBS ? The
> > solitary example for IBS shipped with libpfm-3.4 (smpl_amd64_ibs.c)
> doesn't
> > show how this can be done. I need values of PMD_IBSDCLINAD (given in
> > pfmlib_amd64.h) from
> > the overflow signal handler, much like PEBS records.
>
> AFAIK, this is exactly what smpl_amd64_ibs is doing. It does print the
> load and data
> addresses. OF course, it prints more, you need to adapt the program to
> your needs.
>
>From what I understood, in the switch case statement, it does print all the
control information regarding load/store (this is given in IBSODATA2 and
IBSOPDATA3 MSRs) . However, I don't find a case for the linear/physical
registers causing load miss. May be it goes in the default case. Perhaps
pmd for linear address is PMD 12 , for RIP , it is PMD 8 and for physical
register PMD 13. I am guessing these from the order that the MSR registers
are defined in Software Optimization manual. Is there some way to verify
this by looking in libpfm ?
>
> >
> > I am also reading the documents mentioned by Anoop. It looks like IBS
> > supports time based sampling and not event based sampling ala PEBS. I am
> not
> > sure how much lossy the information would be. Anyways, a related
> question is
> > that using IBS, can I get *a set of * linear addresses for load miss
> > instructions after kernel sample overflow ?
> >
> It is true that the sampling period is expressed in cycles.
> What do you mean by 'after kernel sample overflow'?
I meant when the kernel sampling buffer overflows. But I guess that is
happening here.
>
>
> Another key point, with IBS you cannot filter by priv level, it
> captures instructions at all levels.
>
Sure, I got that after reading your and Anoop's replies .
--
Vivek Thakkar
Graduate Student and Research Assistant
CSC Department, NC State University
-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference
Don't miss this year's exciting event. There's still time to save $100.
Use priority code J8TL2D2.
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone
_______________________________________________
perfmon2-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
