[lkml]   [2011]   [Aug]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [PATCH 6/7] perf, x86: Example code for AMD IBS
    On 02/08/11 22:37, Peter Zijlstra wrote:
    > On Mon, 2011-08-01 at 07:50 +0200, Robert Richter wrote:
    >> On 29.07.11 12:58:49, Peter Zijlstra wrote:
    >>> On Thu, 2011-07-28 at 15:46 +0200, Robert Richter wrote:
    >>>> tools/perf/Documentation/examples/ibs.c | 436 ++++++++++++++++++++++++++++
    >>> That really isn't the place for this..
    >>> Also, how similar is the Alpha PMU to AMD IBS?

    The Alpha PMU (on the EV67 and later CPUs) has two counter modes:
    "Aggregate" which is like counters on other CPUs and implemented in the
    kernel, and "ProfileMe", which is not currently used in the perf. event

    In the "ProfileMe" mode a counter is initialised with max_count-N and
    when the counter overflows (i.e. after execution of ~N instructions) a
    window is opened for profiling. The window closes (roughly) when the
    profiled instruction is retired from the pipeline. The two counters
    count events such as instructions, cycles, Bcache misses, etc. during
    the window. When the window is closed an interrupt to the PMU interrupt
    handler is made. The two counters can be read and there are other
    registers that can be read that provide information on the profiled
    instruction's flight through the pipeline, such as instruction killed
    before being mapped (i.e. it was identified as a nop), instruction
    stalled between fetch and being mapped (usually due to operands not data
    ready), branch direction and branch misprediction (if instruction is a
    branch), instruction was in a new Icache fill stream, instruction
    trapped and trap type, and so on.

    >> Would you prefer
    >> tools/perf/Documentation/examples/x86/ibs.c
    >> instead?
    > Possibly, but having just looked at the example again I don't really see
    > it doing anything perf-record doesn't already do, so why does it deserve
    > to live at all?
    > Initially I thought it was a record+report like example, some code
    > interpreting the 'mess' that comes out of IBS would be most appreciated
    > and I think we can even ship that as perf-ibs-report/perf-ibs-annotate
    > or so (and if its still remotely similar to its Alpha precursor that
    > might make the Alpha folks happy too).

    Sure would be nice if the infrastructure to support ProfileMe mode
    appeared in the perf. events subsystem. I am not going to go full out
    to implement all the support needed for it because there are too few
    users on Alpha to justify the effort. But if we could score an
    implementation of ProfileMe mode with minimal effort on the back of an
    AMD implementation that would make us happy.


     \ /
      Last update: 2011-08-03 11:57    [W:0.025 / U:11.764 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site