On 06/26/2014 01:29 PM, White, Joseph wrote:
> Hi,
>
> Attached is a perfevent PMDA for general review/comments/suggestions. This
> linux-specific pmda exposes the hardware performance counters available on
> recent x86 systems. This pmda has been running on our cluster the last 10
> months without problems (I've been meaning to submit this for inclusion for a
> while!).
>
> Notes:
>
> The PMDA uses the libpfm4 library to access the hardware performance counters
> so any counters that are supported in libpfm4 should be available. Also
> included is the ability to read the Intel RAPL counters via direct MSR access.
>
> The PMDA supports automatically loading different counters for each hardware
> architecture. A single configuration file is used to specify the desired
> counters for each chipset family. The configuration file allows different
> counters to be programmed on different CPUs and supports round-robin
> assignment of uncore counters needed for some AMD chips. The default
> configuration is known to work on Intel Sandybridge EP, Intel Westmere DP,
> Intel Nehalem EX and Intel Core2 chipsets.
>
> The PMDA accesses the system-wide hardware counters and runs as root. This
> means that the hardware counters are unavailable to use by normal
> unprivileged user applications when they are in use by the PMDA. The PMDA
> provides a mechanism to temporarily disable the system-wide counters in order
> to allow normal users to be able to use the counters if they wish.
>
> Build requirements:
> - libpfm4
> - flex
>
> Cheers,
>
> Joe
>
Hi Joe,
I am just starting to take a look at the perfevent pmda.
You mentioned that perfevent pmda has been used for the past 10
months. Is perfevent pmda being used to collect data for XDMoD
(https://xdmod.ccr.buffalo.edu/)?
People may want to run this on other architectures such as arm,
aarch64, or power or at least build pcp on those architectures with
same sources. Has there been attempt to compile the code on non-x86
architectures
so this won't break builds on Fedora or distributions with non-x86
architectures?
How portable is the perf event selection to non-x86 architectures?
This might be an error on my part, but do the libraries for math and
pthreads (-lm and -lphtread) need to be stated in the GNUmakefile or
are those passed by $(PCP_PMDALIB)?
-Will
|