Thanks for the link. There is some debate on the Linux kernel side as well
recently about the right API for this task.
Perfmon2, the proposal by Stephane Eranian advocated an approach where the
user level library knows about the hardware and schedules them between
{process, system} x {counting, sampling} modes. Therefore, the abstraction
between the user lib and the kernel is register based (read MSRx/write
MSRx).

However, Ingo Molnar recently has been working on an interface that looks
more like freebsd - with some important differences. He argues that the
kernel should be scheduling the available hardware between competing per
process and system wide requests. But unlike FreeBSD, which seems to return
an opaque pmc_id_t, he uses a file descriptor based API, which minimizes the
new system calls needed.

There is continuing controversy about whether this type of an abstraction is
a good thing or  raw hardware should be exposed to user space (and whether
counter scheduling should be managed in userspace). So some of us will be
watching the debate closely.

 -Arun

On Mon, May 18, 2009 at 3:22 AM, Vinayak Hegde <[email protected]> wrote:

> Hi,
>
> This is the first talk that Joseph Koshy gave on FreeBSD / PMCTools as
> a part of the ACM Tech Talk series in Bangalore and he has been a
> FreeBSD committer for more than a decade :-)
> You can read bout the talk and download the presentation from
> http://bit.ly/lT0tB
>
> -- Vinayak
> --
> Twitter @ http://twitter.com/vinayakh
> Blog @ http://thoughts.vinayakhegde.com
> _______________________________________________
> bsd-india mailing list
> [email protected]
> http://www.bsd-india.org/mailman/listinfo/bsd-india
>
>
_______________________________________________
bsd-india mailing list
[email protected]
http://www.bsd-india.org/mailman/listinfo/bsd-india

Reply via email to